log.md

log.md

Append-only log of every /ctx-compile, /ctx-query promotion, and /ctx-lint --apply operation. Newest at top. (Entries below 2026-08-08 use the old /wiki-* command names — they are history and are left as written.)

2026-08-25 — The attribution advantage widens with corpus size

citation_doc_f1 — the only metric whose gap exceeded its own noise — was re-measured at 48× corpus (828 files, ~172,000 tokens, 800 deterministic distractors, blocklist re-verified before the run).

Gap against compiled: −0.056 at 24 sources, −0.091 at 828. Long-context fell from 0.491 to 0.455, a drop of 0.036 against a within-arm swing of 0.011, while its answers stayed perfect at 1.00 content recall. It answers as well as ever and attributes worse.

The mechanism is legible. Gold lists 1.5 evidence documents per case. The compiled arm offers 6.3 citations across 4.2 distinct documents — several precise anchors inside a tight set. The agent offers roughly one citation per document, 4.8 distinct at 24 sources and 5.1 at 828: it widens as the haystack grows. The advantage is not finding the right document, since both do that; it is not dragging in extra ones, and tightening rather than loosening with scale.

This reframes the defensible claim. Not "better answers than an agent" — measured false at both sizes, twice. Rather: when a conclusion is challenged, the compiled package points at a tighter set of sources, and stays tight as the corpus grows. Narrower, and the one the evidence carries.

Read as an upper bound, not a result: the compiled package holds only the 24 needle sources, so this compares a perfectly curated index against an agent facing noise rather than two systems on one corpus. Decisive in one direction only — had compiled failed to win under conditions this favourable it would not win at all. Compiling all 828 and repeating is the experiment that would settle it, now justified by a signal rather than a hope. Single run at scale, n=12, one model.

citation_resolvable stayed 1.00 at 828 as at 24. Nothing fabricated a citation at either size. Any pitch resting on "agents invent sources" has now failed measurement twice.

Verification: 66/0, doc-paths clean.

2026-08-25 — The benchmark's broken metrics were repaired, and auditability was measured

Four defects fixed, the comparison repeated twice under the corrected contract.

status never stated its allowed values, so the scorer compared ANSWERED against answered and reported a formatting mismatch as a capability score. The prompt now states them and the scorer normalises. Re-scoring the existing predictions with no new model calls moved it 0.0 → 0.83 / 0.75 / 0.67.

assertion_keys asked for a vocabulary the model was never given. Every task now carries the same 20-key label space with allowed and forbidden keys mixed, so choosing between assumes-brl-usd-5.50-as-of-june and ...-5.70-as-of-june stays a real discrimination while the metric becomes answerable.

citations was worse than a # prefix. Both arms cite the correct source document and differ only in addressing scheme: the compiled arm uses line ranges (L23-L24), gold uses section slugs, and exact matching gave the more precise scheme a score of exactly 0.00 on both precision and recall. Replaced by citation_doc_f1 (right source, scheme-independent) and citation_resolvable (can a reader follow the pointer to real text — heading slug, frontmatter anchor, or in-range line), the latter negative-controlled against fabricated sources, invented slugs and out-of-range lines. Anchor-exact numbers stay in the output for continuity and are not provenance quality.

Claim resolution 0.75 → 1.00. The earlier reading of that figure was wrong and is corrected above: no CLM-* id was ever unresolvable — 65 catalog ids, zero fabrications. Three answerable cases supplied no ids at all, because the contract never asked for them. Stated explicitly, all twelve now supply resolvable ids, stable across both runs. A contract defect, not a compiler defect.

Measured result, two runs at 24 sources. citation_doc_f1 compiled 0.552 / 0.541 against long-context 0.492 / 0.490 — a gap of +0.060 / +0.051 with a within-arm swing of 0.011, so roughly five times its own noise, and the ordering held in both runs. The compiled package cites the right source documents better than an agent reading the folder. That is the first measured advantage for the compiled arm recorded in this repository. assertion_score moves the same way (+0.071 / +0.038) but its gap is the size of its swing (0.051) and is not established. content_recall shows no difference.

Weight it accordingly: one axis, about 0.05, at a corpus size where the compiled arm is still slower per question and costs ~25 minutes of compilation first. citation_resolvable was 1.00 for every arm in every run — nothing fabricated a citation, so the premise that an agent invents provenance did not hold here. Two runs of twelve cases on one corpus with one model cannot size a 0.05 effect; they can only show the ordering did not collapse. bm25's third run aborted on a transient CLI failure and is reported as two arms rather than three.

Verification: 66/0 under both locales; verify-northstar-benchmark.sh green after the scorer's output shape changed; doc-paths and doc-claims clean.

2026-08-25 — The decision benchmark was run for the first time, and the docs were aligned to it

benchmarks/northstar/runs/ was empty. The benchmark had never been executed, so every claim about compiled context was UNMEASURED by the project's own definition — while 61 CI checks and 20 gates guarded whether the machinery ran. Results are recorded in benchmarks/northstar/RESULTS.md and transcribed there because runs/ is git-ignored.

At the shipped corpus size (24 sources, 28 files, ~3,537 tokens): compiled 1.00 content recall / 0.92 refusal / 0.75 claim resolution, 52.3s per question, after ~25 minutes of compilation. long-context 1.00 / 0.92, 39.0s. bm25 0.73 / 0.83, 25.0s. The compiled arm is slower than an agent reading raw/ and no more accurate. (Correction, same day: the 0.75 was first read as a quarter of cited IDs failing to resolve. Every cited ID resolved — 65 catalog IDs, zero fabrications. Three answerable cases supplied no claim IDs at all: current-01, pit-01, super-01, the temporal and supersession cases where an auditable identifier matters most.)

At 48× corpus (828 files, ~172,000 tokens): long-context fell only 1.00 → 0.97, and the point-in-time, supersession and contradiction cases — the rungs the product thesis rests on — held at 1.00. Cost grew 1.4× for 48× the data, because the agent greps and opens a handful of files rather than reading everything. That is what "do the work once, then queries are cheap reading" was meant to buy, arriving with no build step. The compiled arm was not run at that size: the decision rule was fixed beforehand — if long-context held, compiling at scale would pay hours to tie. It held.

Three of seven metrics are defective and return 0.0 for every arm at every size: status (model returns ANSWERED, scorer wants answered, the prompt never states the allowed values), citation_* (model returns #anchor, gold stores anchor), and assertion_score, which asks the model for a label vocabulary it is never given and which cannot be supplied without leaking forbidden_assertion_keys. Consequence: the axis where a compiled package should win produced no signal in either direction, and the tie on content_recall is a ceiling effect rather than evidence of equivalence.

A structural observation: compiling 24 sources (~3,537 tokens) produced 83 files (~39,058 tokens) — 11× the material it was built from — and the benchmark then feeds 26 of them back per question against long-context's 28. There is no context reduction at this size; per-source overhead dominates.

New: tests/eval/northstar-scale/gen-northstar-filler.sh, a deterministic distractor generator reusing the design of gen-scale-filler.sh but emitting raw sources, which is what the competing arms read. Byte-identical across runs, blocklist-enforced against every discriminative needle term, negative-control tested. Two defects were found in the verification itself before it was trusted: the blocklist first matched the workspace directory name scale-ws and flagged every real source as contaminated, and an exit code was read through a pipe to head, turning a correct failure into an apparent pass.

Documents aligned to what is demonstrated. docs/WHAT-IS-THIS.md was written earlier the same day, before the measurement, and its capability ladder implied the compiler was required for the temporal and supersession rungs; it now separates what a compiled package is built to answer from what has been shown, and its comparison gained a fourth option — an agent with the folder — which is the strong alternative rather than a strawman. docs/CONTEXT-COMPILER.md keeps the "Not RAG" argument, which survives against BM25, and adds a dated note that it does not survive against an agent; "beats a fleet of specialized agents" became "should beat", never having been tested. site/index.html gained a section separating machinery numbers from answer quality. docs/SELLING.md separates the provenance property, which is checkable, from a quality advantage, which is not established. README.md now cites evidence where it previously relied on modesty. AGENTS.md needed no change: its line 526 already scoped its guarantee and excluded semantic entailment.

RESULTS.md and the scale generator ship into generated compilers, so a user receives the evidence and can re-run the comparison rather than taking it on trust.

Verification: 66/0 under both LC_ALL=C and C.UTF-8; ratchet clean at 15 gates; installer verifier 10/10; doc-paths, doc-claims, site-claims and definition-consistent all green.

2026-08-25 — The reachability gate was failing on fork pressure, not logic

CI went red on affef57 with R33 (gate-reachable), R36 (gate-ratchet, purely a consequence) and R47 (gate-definition-consistent). Both were legible immediately because gate_diag now replays a failed gate's own output — the change that landed the day before precisely so this would not require guessing.

R47 fired on its own test fixture. tests/gates/definition-consistent/dirty/docs/ SELLING.md deliberately contains "second brain"; the scan prefers git ls-files, the fixtures were untracked when the gate was first run, so it passed locally and broke the moment they were committed. The identical trap templates/QUICKSTART-fresh.md set for gate-doc-paths.sh a day earlier. tests/gates/* is now exempt, with the reason recorded in the gate header.

R33 was never the locale bug it was assumed to be. It failed on both legs this time. The sixteen oracles it called unreachable are all wired correctly. The cause was the fixed-point loop: one grep per (candidate x reached-file) pair, ~n^2 per pass over ~150 scripts — 86 seconds locally and tens of thousands of forks. A grep that fails transiently under CI load is indistinguishable from "basename not present", so it marks a correctly wired oracle unreachable. That explains every observation that made no sense before: green locally and in a clean clone, intermittent on CI, and a different set of oracles each run (three on the C leg, two on C.UTF-8, none reproducible). Adding one script in affef57 made it worse by widening the loop.

Collapsed to a single pattern-file grep per pass: build the comment-stripped corpus of everything reached, collect every pending basename into one pattern file, and match once. Same verdict — 57 oracles, 0 standalone — verified by running the old and new implementations back to back. 86s to 15s. Fixtures still discriminate, the comment-strip still holds (a script named only in a comment stays an orphan), --count unchanged.

Honest about confidence: fork pressure fits all the evidence but was not reproduced on Linux, and no Docker daemon was available to try. What is certain is that the fork storm is gone and the verdict is identical, so the failure mode cannot recur whichever theory was right.

Also corrected: a claim in this session that smoke-all.sh contained NUL bytes. That was a shell-quoting artifact — $'\0' collapses to an empty pattern that matches every line. The file has zero NULs and is valid UTF-8; the binary-file theory was wrong.

Verification: 66/0 under both LC_ALL=C and C.UTF-8, identical; ratchet clean at 15 gates; fixtures discriminate at 20; shellcheck -S error clean.

2026-08-25 — One definition, one linear explanation, and a gate for both

An audit of every explanation surface found six competing framings of what this project is — context compiler, build system, factory, AI workspace, second brain, LLM-wiki knowledge base — two of them contradictory. docs/SELLING.md was titled "the productized second brain" while docs/CONTEXT-COMPILER.md carries a section headed "Not a second brain". AGENTS.md:27 said compiler and AGENTS.md:29 said knowledge base, two sentences apart. Nothing was broken: prose that disagrees with prose produces no red.

Worse, the three surfaces a newcomer meets first — README.md, START-HERE.md, site/index.html — contained no why at all. README reached git clone by line 29 without stating a problem. The genuine why-sentences existed but each appeared exactly once, buried: docs/CONTEXT-COMPILER.md:368 ("sources arrive far less often than questions"), :365 (the RAG contrast, at line 365 of 478), docs/EXPLAIN.md:22-25.

• docs/WHAT-IS-THIS.md — one canonical linear explanation for a non-specialist, 146 lines, each step depending on the last: the problem, why filing and search fall short, the move, the unit, what it buys you, why "compiler", what it is not. It ships, so a generated compiler carries it. docs/CONTEXT-COMPILER.md ships alongside it now, because the new doc links it and that link must resolve in a generated compiler.
• Four representations, each grounded in a tested artifact rather than invented for a pitch: the five questions every statement carries (from the real claim schema — what, how we know, when, how sure, what it replaced); the capability ladder drawn from the twelve independently-scored cases in benchmarks/northstar/gold/cases.json; one question answered three ways (folder, search, compiler); and the cost line that justifies the architecture.
• Six framings reduced to one, carried verbatim by all five primary surfaces. docs/SELLING.md retitled, killing the contradiction. AGENTS.md no longer defines itself twice. README.md:188 "How the factory works underneath" was deliberately kept — the site CTA anchors to it, and there "factory" names a mechanism, not the product's identity.

New gate R47 — DEFINITION-CONSISTENT. The canonical sentence must appear verbatim on every primary surface, and no tracked file may call the product a "second brain" except the two that explicitly deny the label. Negative controls both ways: dropping the sentence from README fires one violation; reintroducing the label fires one. log.md and LEARNINGS.md are exempt as dated history, the same carve-out gate-doc-paths.sh makes and for the same reason — rewriting an accurate record of what the project used to call itself would be falsifying it.

On whether this made the system easier to use: it did not, and the measurement says so. eval-onboarding.sh scored 3/5 before and 3/5 after, failing the same pair (C3, C4). The honest reading is that this work targeted a check with no headroom — C5 ("at most two new concepts to first value") was already green in all five runs of this session. C4 fires when the agent opens AGENTS.md, which all four shipped shims instruct it to do and which /ctx-compile genuinely needs; it is structurally unpassable without degrading compilation. Tool actions also rose (32 → 31 → 45 → 35 → 76) — one data point, direction noted, cause unseparated between variance and longer docs. The docs are more correct and now mutually consistent; measured ease is unchanged. The next useful work is fixing the instrument, not writing more prose: C4 measures whether an agent read the schema, not whether a human could start.

Verification: 66/0 under both LC_ALL=C and C.UTF-8, identical, "All 61 checks green" in both logs; ratchet clean; fixtures discriminate at 20; reachability 57 oracles; installer verifier 10/10; shellcheck -S error clean; verify-ai-workspace, verify-self-service-profile and verify-site-claims all green after the surface edits.

2026-08-24 — CI tells the truth, and the suite is green

Two defects, one of which had been hiding the other.

• ci.yml:50 masked every smoke failure. bash scripts/smoke-all.sh --no-build | tee … takes its exit status from tee, which is always 0, so a red suite reported the job green. Only the downstream LOCALE-INVARIANT job noticed anything was wrong, and only indirectly — it greps for the All N checks green line, which a failing run never prints. Fixed with shell: bash (GitHub's -eo pipefail). Added if: always() to the tally upload, because with the status now propagating a failing suite would skip the artifact and leave LOCALE-INVARIANT with nothing to compare — the tally is most needed exactly when the suite failed.
• The copy drift that kept the suite red. extensions/context-workspace/extension.js labelled the action "Teach it another kind of work"; verify-self-service-profile.sh asserted "Teach this workspace another kind of work". Three phrasings existed across the repo. Settled on the verifier's: "it" is ambiguous in a sidebar, and the longer form matches START-HERE.md. Fixing it advanced the verifier to a second assertion my own QUICKSTART rewrite had broken — the doc no longer named the Create Specialization action, which all six user-facing surfaces are required to mention. Restored, and audited all six rather than only the one that failed.

scripts/verify-self-service-profile.sh passes for the first time in this session, and with it the whole suite: 65/0 under both LC_ALL=C and C.UTF-8, exit 0, identical. The All N checks green line is present in both logs, so LOCALE-INVARIANT has what it needs. The CI step name still said R1–R44; corrected to R1–R46.

2026-08-24 — Complete operator reference, and a gate that keeps it complete

An inventory found seven of twenty-two commands documented nowhere at all — not in README.md, ADVANCED.md, START-HERE.md, docs/QUICKSTART.md, or the generated workspace's own README: /ctx-start, /ctx-gate, /ctx-discover, /ctx-client-assumptions, /ctx-client-decisions, /ctx-client-lint, /ctx-client-review. /ctx-start is the entry point — the one command a newcomer is meant to run first appeared in no document a newcomer would read. Four maintainer scripts (gate-fixtures.sh, new-gate.sh, new-corpus.sh, auto-ingest.sh) were undocumented too.

Nothing was broken, which is why it survived: an undocumented command fails silently and permanently. The operator never learns the capability exists, and no test can notice one nobody invoked. gate-command-aliases.sh already proved every alias RESOLVES; that says nothing about whether a human can find what it resolves to.

• ADVANCED.md is now the complete operator reference (30 → 166 lines), split by audience because it ships into every generated compiler: §1–5 operate a compiler (bootstrap, specializations, demo, evidence in, asking for work, render/export), §6–8 maintain the bootstrap repo (suite, evals, gates, ratchet), explicitly marked as not shipped. Descriptions are taken from each command's own description: field rather than written fresh, so the reference cannot drift from the contract on day one.
• README.md points at it and lists the eight commands worth reaching for, now including /ctx-start.
• site/index.html had three defects: the 5 slash commands stat read as a total (it is pinned by verify-site-claims.sh S3 to the five CORE lifecycle commands by design — reworded to say so, keeping the pin's literal intact), the example query had drifted from the one every entry doc now shares, and the CTA anchor #how-the-factory-works did not match the heading #how-the-factory-works-underneath. Added the operator-reference link the landing page never had.

New gate R46 — COMMAND-DOCUMENTED (scripts/gate-command-documented.sh). Every shipped command must be named in ADVANCED.md, and every command the reference names must exist so a typo cannot invent one. The reverse check is anchored to the backtick form because commands are written `/ctx-foo` while scripts are written scripts/wiki-metrics.sh — without the anchor the slash inside a path matches and the gate reports a command nobody claimed. Honest limits recorded in the header: it proves a command is NAMED, not that the description is accurate; alias files are skipped deliberately; and the reference path is fixed to ADVANCED.md, because one canonical reference is the point.

Verification: 64/1 under both LC_ALL=C and C.UTF-8, identical; ratchet clean at 14 gates; fixtures discriminate at 19; installer verifier 10/10; shellcheck -S error clean repo-wide; and a freshly generated compiler ships the reference with 0 of its 22 commands missing. Negative control on the new gate: removing /ctx-start's row makes it fire on exactly that command. The one remaining suite failure is the pre-existing verify-self-service-profile.sh / extension.js copy drift, untouched.

2026-08-24 — Bootstrap path made survivable for a stranger

The machinery was production-grade; the on-ramp was not. Five dead ends on the documented path, each verified against a freshly generated compiler.

• preflight.sh called python3 optional. False: /ctx-start, /ctx-add, /ctx-compile, use-profile.sh and every /client-* command invoke it by that exact name. A user without it got "Ready." then failed on their first action. Now a hard requirement, checked by name (a python that is not also python3 fails).
• README.md advertised two demo scripts that never existed and linked two docs never written. Repointed at stage-northstar.sh / run-northstar-benchmark.sh, profiles/client-decision/COMPILATION.md and benchmarks/northstar/README.md.
• docs/QUICKSTART.md SHIPS and named scripts that do not ship — including its own "Verify installation" step. Split: templates/QUICKSTART-fresh.md is the installed copy, sourced the same way README.md ← templates/README-fresh.md already was.
• verify-guided-workspace.sh SHIPS and ran four tests/guided-workspace/ tests; two were missing from the manifest. The documented verification step and the VS Code task both died in every generated compiler while passing here.
• "Start here" was not the first step — it opened with a VS Code extension build, impossible before cloning, while the clone appeared 100 lines down under "How the factory works underneath". The extension path also needs Node, npm, network and the code CLI, declared only in a doc that does not ship.

The installer now ships itself (create-context-compiler.sh, its verifier, the manifest and the six FRESH template sources). That removed seven dangling references at once — cheaper than rewording seven documents — and generated compilers are now self-replicating: gen1 → gen2 produce byte-identical 345-file trees, and a generated compiler passes its own installer verifier.

New gate R45 — DOC-PATH-RESOLVES (scripts/gate-doc-paths.sh). gate-doc-claims.sh already stopped a NUMBER in prose from rotting; nothing did the same for a PATH, and every defect above is that one uncovered rule. Two sub-rules: a path named in a tracked doc must resolve, and a scripts/-or-tests/ asset that a SHIPPED doc or SHIPPED script depends on must itself ship. The second is what a dev-repo-only check structurally cannot see. Scoped deliberately: prose mentions are exempt (rewriting accurate prose about this repo's own CI would make the docs worse — the three shipped docs carrying such mentions got an honest header note instead), log.md and LEARNINGS.md are exempt as dated history, and a guarded invocation ([ -f x ] && bash x) is recognised as correct rather than suppressed — proven by a mutation in the dirty fixture where a guard naming a different file does not launder the command.

Onboarding eval. C2 ("one unambiguous start-here") went red → red → red → green across four runs, each red a genuinely different cause: no canonical path, then colliding literal dates between demo and own-work examples, then README's first query differing from the other three. All five entry surfaces now carry a byte-identical first query and follow-up block; own-work examples use <your date> so no two literal dates compete. Score stayed 3/5. C4 is not an onboarding-doc defect and was deliberately not chased: all four shipped agent shims instruct "read AGENTS.md before any work" and /ctx-compile genuinely needs the schema — passing C4 would mean degrading compilation to move a metric. C3 flipped red once after three clean runs; treated as the variance already recorded, not as a regression.

Also: verify-synthesize.sh ships and used shasum (Perl, absent on Alpine/debian-slim) → openssl dgst -sha256, matching body-hash.sh. Two tests/eval/grain-corpus/runs/ scripts hardcoded an absolute home-directory path, unrunnable for anyone else → $HOME default. 18 occurrences of a local username redacted to /Users/<user>/… across eight committed files, log-line shape preserved. AGENTS.md, CLAUDE.md, GEMINI.md, .cursor/rules and copilot-instructions.md all still said "five slash commands"; there are 22. AGENTS.md gained a "you do not need this to start" banner above its five-version changelog. verify-ai-workspace.sh pinned the literal date the C2 fix removed — loosened to the assertion's intent (a delta-shaped request exists), not reverted.

Verification: 63/1 under both LC_ALL=C and C.UTF-8, identical (locale-invariant); ratchet clean at 13 gates; fixtures discriminate at 18; installer verifier 10/10. The single remaining failure is pre-existing and untouched — verify-self-service-profile.sh asserts extension.js says "Teach this workspace another kind of work"; it says "Teach it another kind of work". That is a product-copy decision, not a gate to loosen.

2026-08-21 — Guided professional workspace

• Added START-HERE.md, user-owned EVIDENCE-INBOX/, generated BRIEFS/ and REVIEWS/, and VS Code extension/task recommendations to fresh compilers.
• Added /ctx-start, /ctx-inbox, and /ctx-create-profile with natural-language routing, explicit clarification rules, and a bounded profile-generation contract.
• Added deterministic scripts/inbox.py state tracking and generic scripts/profile-check.py portable-asset validation, including path/symlink safety.
• Client workflows now support deterministic --save views under BRIEFS/ and REVIEWS/ without mutating evidence or compiled claims.
• Added guided-workspace acceptance tests, installer/package integration, regression wiring, and documentation for non-developer VS Code operators.

2026-08-21 — Bootstrap factory positioning

• Reframed README.md and the GitHub Pages landing page around two distinct products: the bootstrap that creates/configures specialized compilers, and the specialized compiler that continuously turns evidence into a usable context package.
• Added concrete profile-creation, daily operation, consumed-output, and current-limit sections using the shipped client-decision profile as the reference implementation.
• Preserved measured site claims and explicitly states that new profile creation is currently LLM-assisted development, not a no-code workflow.

2026-08-21 — Decision-context profile integration

• Added opt-in compile/lint hooks, claim-aware packaging, clean-install assets, deterministic integration verification, and Northstar task preparation.
• Generic compilation remains default; schema stays v6 because profile version 1 is additive and independently versioned.
• Task preparation is unmeasured. Only saved model predictions scored against gold are reported as measured.

2026-08-11 — segmented-source anchors never resolved, and nothing checked

Rationale. AGENTS.md tells /ctx-compile to cite a segmented section with its positional range dropped — ## Power Envelope (lines 13-19) → #power-envelope. citation-audit.py only ever slugified the whole heading, giving power-envelope-lines-13-19. The schema and the resolver could never agree, and the compiler was following the schema exactly.

Measured downstream on a 2,414-source corpus: 15,440 unresolvable citations, 82% of all failures, none of them the compiler's fault. Eleven gates in that repository ran green throughout, because no gate ran C1/C2 resolution over the compiled root. A broken anchor fails silently — the page renders, names a real file, and reads as sourced.

The resolver was wrong, not the convention. A line range belongs to one segmentation run: an anchor carrying it breaks the moment the source is re-segmented with a different word budget, which is precisely the fragility the schema's form avoids. Both forms now resolve, so packages already citing the full slug keep working. Ambiguity stays fatal — if dropping ranges makes two headings collide, that is a coin flip, not a match.

Regression test, and it fails without the fix. verify-citation-audit.sh gained a case running the real engine against the schema's own example, asserting the range-dropped form resolves, the full form still resolves, and a nonexistent heading still does not — the third clause being what stops the "fix" from being a resolver that says yes. Reverting the one-line change turns it red.

CITATION-FLOOR (R44), the gate whose absence caused this to run unnoticed. It runs the audit over the compiled root and fails when an anchor does not resolve. Detection delegated wholesale to citation-audit.py — a second implementation of "does this anchor resolve" would eventually disagree with the first, and then a corpus is sound by one and broken by the other.

Found by the gate it exposed. CTX-ROOT-ADOPTION went 80 → 81 on the new gate's own error message, which named the legacy root as a literal. Reworded rather than baselined up — declared failure mode 2 of that gate, met for the second time.

Suite 46 → 47. verify-site-claims and gate-doc-claims both billed for the change, as designed.

2026-08-10 — the adapter blueprint moves upstream, where compilers are generated

Rationale. new-corpus.sh and gate-adapter-contract.sh were built in one deployment while compiling a 2,413-document corpus. They are the machinery that makes pointing a compiler at a NEW source a bounded job — and they were trapped in the single deployment they were meant to generalise beyond. A freshly generated compiler got none of it. They now ship in the installer manifest.

What a new corpus costs, after this. A corpus-<name>.json declaration and two functions — _catalog() and _fetch_one(). Everything else is generic or scaffolded, and ADAPTER-CONTRACT settles with an exit code whether the result is something this compiler can actually compile.

A name collision, resolved honestly. scripts/stage-corpus.py already existed here and was YouTube-transcript-specific despite the general name — down to the padded/unpadded timestamp headings it emits. It is now stage-transcripts.py, which is what it always was, and the corpus-agnostic stager takes the general name. Its oracle (verify-corpus-eval.sh, R27) moved with it and stays green.

TWO BUGS THE PORT ITSELF EXPOSED, both invisible in the source repo:

1. new-corpus.sh scaffolded a 0-byte daemon and reported success. It sed'd from gain-watchd.sh, which exists only in the deployment it came from. A scaffolder whose output is empty is worse than one that fails, because the file looks written. The daemon is now a template (templates/corpus/watchd.sh.tmpl), a missing template is an error, and an empty result is refused. R43 checks the scaffold produces a non-empty, fully substituted, parseable adapter.

2. gate-adapter-contract.sh exited 2 in a freshly generated compiler, failing the installer's own I6 check, because there was no corpus.json yet. That conflated DECLARED-AND-BROKEN with NOT-DECLARED. A compiler may legitimately have no harvest adapter and feed raw/ by hand through /ctx-extract — the same reasoning ctx-lint-rules.sh already applies to an absent rules/ directory. An absent declaration is now clean and says so.

New checks. R42 (an adapter's output is compilable) and R43 (the scaffold produces real files). The suite goes 44 -> 46; the bound smoke-guard-range claim, the published page's count and README's all moved with it, because three separate gates refused to let them drift.

Files. scripts/{stage-corpus.py,new-corpus.sh,gate-adapter-contract.sh} (new), scripts/stage-transcripts.py (renamed), templates/corpus/*.tmpl (new), tests/gates/adapter-contract/** (new), scripts/installer-skeleton-manifest.txt, scripts/smoke-all.sh, gates/baseline.tsv, scripts/gate-fixtures.tsv, README.md, site/index.html, .github/workflows/ci.yml.

2026-08-10 — schema v6: the schema can now describe a document corpus honestly

Rationale. v5 could compile a corpus of reports; it could not label one truthfully. Both gaps were found by building a compiler over 2,413 USDA FAS GAIN attaché reports, and both bite on day one of any document corpus.

(a) source: document. The page enum was video | analysis | external | mixed — no value for a document. Compiling GAIN, the agent reasoned correctly that a USDA report is neither a video nor its own analysis and wrote external on every page: the value reserved for material fetched off the open web, and in a closed-corpus deployment the value that means contaminated. 43 pages, each citing its source perfectly, were indistinguishable from 43 pages pulled off a search engine. The mislabel is invisible — the page is right, the citation resolves, only the frontmatter lies.

(b) rule_domain: artifact | world, and R-01 is now scoped to artifact. rule_class asks whether a rule has a checkable SHAPE. It does not ask whether the rule has a checkable SUBJECT here. Compiling 279 reports harvested 540 deterministic rules, every one of the second kind:

"A soybean meal import into Indonesia must hold a permit issued under the
 MOT 11/2026 and MOA 11/2026 licensing procedures."

Perfectly deterministic. Completely uncheckable in that package, which contains no consignments — a gate for it would have nothing to open. All 540 defaulted to scope_include: ["wiki/**.md"]; the same placeholder on every single rule is the tell that the field was meaningless for them.

The cost was not theoretical. R-01's count then grows with every source ingested, so the ratchet reddens on a schedule and gets bumped without being read — three times in one session before the category error underneath was spotted. §1 of docs/deterministic-gates.md is about rules the MAINTAINERS must obey; it quietly assumed the output's rules are about the output. True for a repository, false for a document corpus.

World rules are not excused: /ctx-lint counts and prints them every run, they stay catalogued and citable, and one that declares a gate anyway still faces R-02..R-07 in full.

A dead fixture, found by making this change. With world as the fallback the dirty fixture's prose-only rule stopped firing R-01 — and the pair still went red, so it looked fine. Two different checks were both labelled R-01: the real one and the discarded-rule/discard_reason check. The label collision hid the loss. The discarded check is now R-00, the prose-only fixture declares rule_domain: artifact (which is what it tests), and the clean tree gained a NEGATIVE CONTROL — an ungated world rule that must stay green, so the exemption is proven to work rather than merely to exist.

Migration for an older client: none required. rule_domain defaults to world, so an older /ctx-compile that never writes the field produces exactly the new behaviour. source: document is opt-in and nothing validates the enum mechanically. The one visible change is R-01 no longer firing on harvested rules — deliberately.

Files. AGENTS.md (enum, rule_domain section, version), scripts/ctx-lint-rules.sh, .claude/commands/ctx-compile.md, tests/gates/ctx-lint-rules/{clean,dirty}/**, scripts/smoke-all.sh (R4's version pin).

2026-08-08 — three enforcement gaps closed: the standing rule applied to this repo

Rationale. docs/deterministic-gates.md §1 says no deterministic rule stays prose-only, and /ctx-lint R-01 enforces exactly that — on the compiler's output. Three places in the compiler's own repo did not meet the standard. Two were live bugs.

1. A generated compiler could not package itself. scripts/package-wiki.sh ships in the installer skeleton and hard-sources scripts/lib/ctx-root.sh (line 72), which was absent from scripts/installer-skeleton-manifest.txt. Every repo built by scripts/create-context-compiler.sh exited 2 on its first packaging attempt — the context package, the fifth property in docs/CONTEXT-COMPILER.md, was unreachable from a generated compiler. Fixed by adding the manifest row. An audit of every manifest-shipped script for unlisted lib/ dependencies found no others.

2. The ratchet's registration check missed the directory /ctx-gate writes to. scripts/gate-ratchet.sh enforced "every gate owns a baseline row" over scripts/gate-*.sh and context/gates/*.sh, but /ctx-gate writes to gates/<RULE-ID>.sh (scripts/new-gate.sh:38). The first gate the compiler ever generated for itself would have been exempt from the requirement gates/baseline.tsv declares non-negotiable in its own header. Latent only because gates/ holds no .sh files yet. Two other call sites already had the glob right (gate-ratchet.sh:80, gate-fixtures.sh:141), which is what marks it a slip rather than a design choice.

It survived because the gate that catches unproven gates was itself unproven: tests/gates/ratchet/{clean,dirty} existed and --repo worked, but scripts/gate-fixtures.tsv had no row and gate-fixtures.sh excluded gate-ratchet.sh by name as "recurses". That exclusion was over-broad — the recursion only exists for the default invocation, which reads the real baseline; in --repo mode it reads the fixture's, which lists stubs only.

The pre-existing dirty fixture could not have caught it either: it fails for three other reasons, so it exits 1 whether or not gates/ is scanned. Pinning the fix needed a tree whose only defect is an unregistered gates/*.sh — tests/gates/ratchet/gates-arm-dirty/. Verified by reverting the glob: the old code reports "clean — none unregistered" on that tree while an unregistered gate sits in it.

gate-ratchet.sh remains the one gate with no baseline row, now stated in gates/baseline.tsv rather than left implicit: it has no --count mode, because "count your own violations" is undefined for the script that reads the baseline, and a row naming it would make it invoke itself once per run forever. Its coverage is two fixture pairs instead.

3. CTX-ROOT-ADOPTION (R41) — a new gate. AGENTS.md tells new code to resolve the compiled root with ctx_root(). Measured adoption on the day the gate was written: 4 files called it, 41 files hardcoded wiki/ on 80 non-comment lines. gate-context-root.sh (R40) proves ctx_root() resolves; it states in its own header that it proves nothing about the scripts that bypass it.

scripts/gate-ctx-root.sh counts hardcoded roots under scripts/ and is ratcheted at 80, not migrated to 0 — AGENTS.md argues directly against a bulk rewrite of the oracles that constitute this repo's safety net, and this gate deliberately creates no pressure toward one. Five declared failure modes in the header; a five-way mutation proof confirmed the fixture pair catches each break. The negative controls (a commented path, and meta-wiki/-style names that must not match) live in the clean tree on purpose: gate-fixtures.sh compares exit codes only, so a control planted in a tree that already exits 1 proves nothing.

Wired as R41 via --count rather than a bare run: this is the first gate with a nonzero baseline, so a bare run exits 1 by design and forever. R41 proves the gate is operable, R36's ratchet enforces the threshold, R31 proves it discriminates.

Not done, and why. The plan called for adding a RATCHET-NO-INCREASE row to gates/baseline.tsv. That would recurse infinitely — see the exception recorded above.

Three claim gates fired on this change, and each was answered by correcting the claim rather than the gate. gate-doc-claims.sh caught README's bound smoke-guard-range still reading R1–R40; verify-site-claims.sh caught both site/index.html and README still advertising 43 deterministic checks against a suite that now runs 44. Adding a check to a repo that publishes its own numbers is supposed to cost exactly this, and it did.

Files. scripts/gate-ctx-root.sh (new), scripts/gate-ratchet.sh, scripts/gate-fixtures.sh, scripts/gate-fixtures.tsv, templates/gate-fixtures-fresh.tsv, gates/baseline.tsv, scripts/installer-skeleton-manifest.txt, scripts/smoke-all.sh, tests/gates/ctx-root/** (new), tests/gates/ratchet/gates-arm-dirty/** (new), tests/gates/ratchet/clean/gates/RULE-9999.sh (new), AGENTS.md (the ctx_root() paragraph now names the gate that enforces it), README.md (bound claim smoke-guard-range → R1–R41; check count → 44), site/index.html (check count → 44), .github/workflows/ci.yml.

2026-08-08 — schema v5: the commands are /ctx-*, the project is context-compiler-bootstrap

Rationale. The names had drifted from the thing. /wiki-ingest described the weakest part of what it does: the output carries provenance, typed relations, valid time, a portable bundle and its own verifier, and "wiki" named none of that. The canonical commands are now /ctx-init, /ctx-extract, /ctx-compile, /ctx-query, /ctx-lint, /ctx-visualize, /ctx-flashcards, /ctx-diagram, /ctx-discover. The project is renamed llm-wiki-bootstrap → context-compiler-bootstrap, and scripts/create-llm-wiki.sh → scripts/create-context-compiler.sh.

Migration note: none required. All nine /wiki-* names survive as forwarders that print one deprecation line and delegate to the canonical file; they are not scheduled for removal. /ingest still works alongside the new /compile. scripts/create-llm-wiki.sh remains as an exec forwarder. An older client that knows only the old names is fully functional.

What was deliberately NOT renamed. raw/ (immutable evidence — the old name appears inside sources and stays there), log.md history below this entry, committed test fixtures whose bytes are asserted, .scratch/ design records, and the wiki page okf-vs-llm-wiki-bootstrap.md, whose slug names the project as it was called when that comparison was made. Tooling scripts keep their wiki-* names (wiki-to-kg.py, wiki-metrics.sh, wiki-lint-*.sh) — they operate on the wiki layer, which is still called wiki/.

Also in this change. docs/CONTEXT-COMPILER.md's Vocabulary section previously promised the opposite ("No identifier in this repo is named after the right-hand column ... they stay that way"); that paragraph is rewritten to say what changed and why, rather than quietly contradicted. Two new gates: ALIAS-RESOLVES (R37) proves every one of the 19 alias files resolves to an existing canonical command, and EXEC-BIT-PRESERVED (R38) catches the redirect-and-move rewrite that dropped +x from 55 scripts mid-rename.

2026-08-08 — valid-time contract satisfied: seven honest unknowns

The last red row on the package-quality scorecard. Every one of the seven raw sources now records an asserted_at, and every one of them records unknown — with a specific reason and a specific way to resolve it.

That is the honest answer, not an evasion. Each body was read looking for a passage that states the document's own date, and none has one. The near-miss is worth naming: the Devsplainers transcript says the OKF spec was published "On June 12th" — but that dates an event the video reports, not the video. It establishes only that the video postdates 2026-06-12, which is a lower bound, and writing a lower bound into a date field would be the same class of error as hand-writing a hash: a receipt nobody derived. The note records the bound instead. Two sources (the Google Cloud blog, the OKF spec) were fetched with extraction_status: degraded, so any dateline is in the part that never arrived — the note says so and names the fix. Two are synthetic fixtures with no real date and never will have one; those say "unknown by design". Copying fetched_at was never on the table — /wiki-extract forbids exactly that conflation.

The append-only gate caught a first attempt. The notes were originally written as YAML block scalars (asserted_at_note: | plus indented lines). gate-raw-append-only.sh authorises the asserted_at family as keys, but its diff parser only recognises one continuation shape — - wiki/… under ingested_pages: — so the indented note lines read as an unauthorised body edit and all six files were rejected. Correct behaviour from the gate given what it can express. Rewritten as single-line double-quoted scalars, so every added line is a recognised key: value; no gate change was needed, and the contract stayed as documented.

Frontmatter-only throughout: every body hash is unchanged, verified against body-hash.sh per file. No citation moved, no drift fired, nothing needed re-ingesting.

asserted-at-audit.py is clean in both default and --all mode (the latter also covers canary-scanned.pdf.md, which nothing cites yet and which default mode skips — annotated too, so the stricter mode passes as well).

Scorecard: 10/11 → 11/11. All eleven structural measures pass. Which is the floor, not the ceiling: it says the package is sound, not that it is useful. Tier 2 — the behavioural half — still has no measured numbers for this package.

• Updated: all 7 files in raw/, frontmatter only (asserted_at, asserted_at_note)
• Contradictions flagged: none

2026-08-08 — the commitment lint was quote-blind (one reader, four copies)

The new package-quality scorecard reported "4 of 6 cited sources lack an ingest commitment", and the obvious next move was to re-run /wiki-ingest on each. That would have been wrong, and expensive: all four hashes were present, correct, and current. Recomputing each with scripts/body-hash.sh matched the stored value exactly.

The four store the hash unquoted (ingested_hash: d1d2986…). wiki-lint-commitment.sh tested for it with grep -q 'ingested_hash: "[0-9a-f]' — a literal opening double-quote. Unquoted is valid YAML, and scripts/commit-source.py (the canonical writer) reads quote-optionally at line 101 while always writing the quoted form at line 116. Tolerant reading was already the intended contract; two readers just did not implement it.

The diagnosis was wrong in a second, worse way. The error text also claimed "hash-drift detection is disabled for this body". It is not — wiki-lint-hash-drift.sh reads the field through an awk helper that strips quotes. Verified empirically rather than by reading: tampering with two bodies in a temp copy, one quoted and one unquoted, produced 2 raw source(s) no longer match their ingest commitment. Drift was being caught the whole time on exactly the sources the lint called uncommitted. Message corrected.

The same inline grep had been written four times — wiki-lint-commitment.sh, wiki-metrics.sh, eval-retrieval.sh, and verify-retrieval-eval.sh. The last one is the oracle for the third, and it reimplemented the predicate instead of calling it, which is precisely why the bug outlived its own test. Replaced with scripts/lib/commitment.sh — one has_commitment(), quote-agnostic, sidecar-aware, hex-validated (a non-hex value is a hand-written receipt and must not read as verified). Same reasoning as body-hash.sh being the one hasher: a second implementation of a shared predicate does not stay identical to the first.

verify-metrics.sh's P1 fixture was all-quoted, which is how the bug survived. It now carries committed-unquoted.md with a bare hash, and exp_ok/exp_total moves 2/3 → 3/4. Mutation-tested: reverting the reader to quote-only makes the oracle report committed=2/4, i.e. it now fails.

Net effect on the scorecard: 9/11 → 10/11, with no model calls, no writes to raw/, and no wiki pages rewritten. The remaining red row is the asserted_at gap on 6 sources, which is genuine debt.

• Created: scripts/lib/commitment.sh (added to installer-skeleton-manifest.txt — both callers ship in fresh installs)
• Updated: scripts/wiki-lint-commitment.sh (reader + corrected message), scripts/wiki-metrics.sh, scripts/eval-retrieval.sh, scripts/verify-retrieval-eval.sh, scripts/verify-metrics.sh (fixture)
• Contradictions flagged: none

2026-08-08 — context-compiler framing (positioning, no mechanical change)

"Wiki" names the shape of the output; it does not name the function of the system. Adopted context compiler as the category: a context compiler transforms unstructured source material into a structured, provenance-aware, machine-navigable context package that LLMs can navigate, retrieve from, and reason over. Every clause of that definition was already implemented and already gated — extract/segment/ingest for the transform, the page template and typed-relation grammar for structure, the citation + hash + entailment stack for provenance, [[link]]/KG/MCP for navigability, and package-wiki.sh + MANIFEST + the in-bundle verifier for the package. So this is a naming change, not a build change.

The last clause was rewritten before shipping. The draft read "…that an LLM can reliably use." That word promises what this layer cannot deliver: a compiler controls the artifact, not the model reading it, and a flawless package can still be misread. Replaced with three verbs — navigate, retrieve from, reason over — because each is a property of the artifact rather than of the reader, and each is already measured: R1 needle retrieval, R4 citation locus, M1/M2/M5 multi-hop, R5 stale evidence, M6 citation integrity. "Reliable" is a word to earn from a scorecard, not to assert in a definition.

So the scorecard now exists. scripts/package-quality.sh aggregates the structural half — eleven measures, no LLM, no network, recomputed on demand — from the tools that already owned each one (citation-audit.py, corpus-health.py --json, wiki-lint-commitment.sh, wiki-lint-hash-drift.sh, asserted-at-audit.py). Nothing is recomputed locally, for the same reason body-hash.sh is the one hasher: a second implementation of "count the citations" drifts from the gate that enforces it. The two tiers are kept apart on purpose — structural measures are properties and recompute for free; the eval-* family measures events, costs real model runs, and its numbers must carry a date (the rule verify-site-claims.sh already applies to every score on the public page). Printing them as one number would launder an event as a property.

It reports rather than gates by default (exit 0; --strict exits 1), because a repo with known debt must still be able to read its own numbers — and this repo has debt: 9 of 11 pass, with 4 of 6 cited sources missing an ingest commitment and 6 sources missing asserted_at. That is the scorecard doing its job on its first run.

scripts/verify-package-quality.sh (Q1–Q7, wired as smoke R35) is the oracle, because a scorecard that only ever prints green launders debt as quality. It proves the thing still fails an unsound package, agrees with its own --json, never edits what it measures, and does not report a perfect score on an empty package. Q2 pins a real bug found in the first draft: corpus-health.py prints thin (<2 related): 0, whose label contains a 2, and a first-number grab reported the THRESHOLD as the count — a fully-linked package scored "2 thin". Fixed by reading --json fields instead of prose. Both mutations (reintroduce the bug; force the citation row green) were confirmed to make the oracle fail before it was wired in.

Adding R35 moved the suite 37 → 38, which verify-site-claims.sh caught immediately on site/index.html and README.md — the guard behaving exactly as designed. Counts updated in the same commit, along with the smoke-guard-range claim marker (R1–R34 → R1–R35).

New docs/CONTEXT-COMPILER.md defines the category independently of this repo's name: the definition, the five conformance properties with their verify scripts, the vocabulary mapping, and — stated up front rather than buried — the limit. The transform stage is LLM judgment and is not reproducible; determinism is scoped to the body-hash cache key, the synthesis artifacts, the OKF export, and the manifest. The guarantee on offer is verifiable output, not reproducible output.

Two features that look like exceptions are named instead of hidden. /wiki-query promotion writes into its own source tree mid-run — framed as dependency resolution (npm install fetching a missing dep during a build), since the fetched source goes through the same no-bare-URL, frontmatter, and entailment rules as any /wiki-extract acquisition. Flashcards, slides, diagrams, /wiki-discover, and wiki/journal/ are a viewer tier over a built package, which is what AGENTS.md already said by calling them "not lifecycle steps".

AGENTS.md gained a ## Output format — the context package section: packaging existed only in shell scripts and docs/SELLING.md, so the schema layer had no notion of its own distributable — the weakest-surfaced clause of the definition. No /wiki-export was added; wiki/commands.md records that a unified exporter was deliberately rejected in favour of a split output tier, and that stands. Schema version deliberately not bumped: no command, frontmatter, or layer rule changed, and smoke-all.sh pins the literal **Schema version:** 4.

• Created: docs/CONTEXT-COMPILER.md, wiki/context-compiler.md
• Updated: README.md (lead, packaging framing, vision), AGENTS.md (what-this-is + output format), docs/EXPLAIN.md (source-maps and npm-pack rows), site/index.html, CLAUDE.md, GEMINI.md, .clinerules, .github/copilot-instructions.md, .cursor/rules/llm-wiki.mdc, wiki/index.md, wiki/core-idea.md
• Unchanged on purpose: repo name, git remote, /wiki-* commands, raw/ and wiki/ paths, every script name, the bundle format
• Contradictions flagged: none

2026-07-24 — cross-modality retrieval eval (the measuring instrument)

"Can a bootstrapped wiki retrieve accurate, point-in-time info about anything?" was asserted, never measured — the existing evals cover multi-hop traversal and citation faithfulness, neither of which touches tabular or thread sources, and none of which has a time axis. Added scripts/eval-retrieval.sh: five binary checks (R1 needle retrieval per modality, R2 point-in-time as-of + current in one run, R3 refusal on absence, R4 citation locus, R5 stale evidence blocks the answer) run against a wiki built by the REAL installer and loaded through the REAL /wiki-extract → /wiki-ingest path — no hand-authored wiki fixture, because the gaps being hunted (tabular truncation past the 20-row preview, thread flattening) live in extract/ingest and a fixture would paper over exactly them.

Corpus is generated, not committed (tests/eval/retrieval-corpus/gen-corpus.sh, deterministic — the numbers are only comparable across runs if the input is fixed): a 1200-row CSV with the needle at row 947, a 14-message thread with the needle in message 11, a 24-section 11k-word report with the needle in section 21, and two vintages of one throughput figure (412 Q1 → 389 Q3) with the vintage stated in the body text, so the eval runs against today's schema with no asserted_at field. Every needle is a synthetic NEEDLE-<MODALITY>-<hex> planted past the boundary its extractor truncates at — unguessable from a preview and absent from pretraining, so it can only be produced by actually reaching it.

scripts/cite-span.py resolves a citation to its passage by importing citation-audit.py's anchor grammar rather than reimplementing it (same reason body-hash.sh is the one hasher — a second copy would make the eval measure something the audit doesn't enforce). R4 is the Goodhart-hardened check: containment alone passes a whole-file cite, tight-span alone passes a cite of the wrong lines, so it demands one passage that does both.

Graders live in scripts/lib/eval-common.sh so scripts/verify-retrieval-eval.sh (E1–E6, wired as smoke R23) can exercise them with no LLM and no spend: corpus determinism, needles past the first-40-lines boundary, and — the point — that whole-file, wrong-line, and oversized-but-containing citations all FAIL. An eval whose graders are unverified reports a perfect score on a broken system. Holdout: tests/eval/retrieval-questions-holdout.md plus gen-corpus.sh --holdout (plain-text meeting notes, own needle, supersession twist), never run by default, never in CI, never to be "fixed" by editing. Smoke now 28 checks, all green. The eval itself has NOT been run yet — that costs claude -p invocations and is the next step.

2026-07-26 — first eval run came back VOID; hardened the instrument

Ran scripts/eval-retrieval.sh for real. It printed retrieval score: 6/12 (R1 3/3, R2 2/2, R3 1/1, R4 0/5, R5 inconclusive) and that number is not a measurement — recorded here because a plausible score from a broken harness is the most expensive artifact this project can produce.

Three faults, in ascending order of importance. (1) /wiki-ingest died on a transient API Error: 529 Overloaded, leaving wiki/ at one empty stock index.md; claude_p() now retries once on 5xx/overload/rate-limit. (2) The question loop ends done < "$tmp_q", so every nested claude -p inherited stdin pointed at questions.tsv and swallowed the later eval questions into the first one's prompt — the model said so in its own answer. All three invocation sites now pin </dev/null. (3) The one that matters: with an empty wiki, /wiki-query answered all six questions correctly by reading raw/ directly, and disclosed it. So R1 and R2 scored full marks while measuring "an agent can grep a CSV" — a false pass straight through the middle of the loss function, on the checks that exist to prove wiki retrieval.

Fix for (3) is two-part, because detection and visibility are different problems. A VOID gate refuses to score when wiki/ has <= 1 page or no raw source carries a real ingested_hash — it prints why and exits 3 instead of spending on queries that cannot measure the thing. And a via column tags every answer wiki / raw-only / unknown, so a raw bypass stays visible once the wiki is populated, which the gate alone would not catch. verify-retrieval-eval.sh grows E7 for both, negative-tested: deleting the gate fails it, unpinning stdin at one of three sites fails it.

Also --work=DIR: state persists outside a temp dir and install/extract/ingest/each-question are skipped when already complete, so a kill costs the current stage instead of the whole run, and answers on disk re-grade for free when a grader changes. Two prior runs were lost to kills before this existed.

Worth recording against the prediction made before the run — R1-csv fail, R2 fail both legs, R1-email unknown: every needle came back correct from raw text alone, across CSV row 947, email message 11, report section 21, and both report vintages. That is evidence about the agent's reading, and none about the wiki's retrieval. The run designed to measure the wiki passed without the wiki existing.

2026-07-26 — second run: R1 3/3 via wiki, and three more grader faults

Re-ran with ingest healthy: 7 raw files, 14 wiki pages, every answer tagged via: wiki. First genuine signal — R1 3/3, needles retrieved through the wiki across CSV row 947, email message 11 and report section 21; R4 2/5, with the email needle earning a tight correct locus (raw/thread-q3-planning.eml.md#L120-L127) and the CSV and report needles not. Those two numbers stand.

Three more faults found, all in the instrument. (1) --work resumability poisoned the corpus: R5 corrupts capacity-report-q1.md (412 -> 999) and the marker files let a resumed run re-ingest the mutated body, committing ingested_hash over it — so R5 is structurally unscoreable and the as-of question silently changes answer. R5 now snapshots and restores the file, and a .mutated marker surviving a kill forces a clean re-extract on the next run. (2) R2-asof PASSED with a headline of 999 GB/day — the wrong figure — because expects was a document-wide substring search and 412 appeared eight paragraphs down in a caveat about the drift. (3) R3-absent FAILED on a textbook refusal: "No Q2 2026 sustained throughput figure exists in this wiki" matched no marker, because every marker required "for"/"in" straight after the noun.

Fix for (2) is per-question, not grader-wide: both vintage questions now carry a headline-anchored forbids-pattern (^[^a-zA-Z0-9]{0,4}(389|999)). Scoping expects to the first N lines was tried first and is strictly worse — it fails answers that open with a preamble while still passing the buried case it was written for. Reverted, and the reasoning is in the grader so it is not retried. E3b/E3c/E5b cover all three; oracle is E1-E7 across 12 checks.

Re-graded verdicts (cached answers, no spend — the point of --work): R1 3/3, R2 1/2 (asof now correctly FAILS), R3 1/1, R4 2/5, R5 inconclusive. R2-asof leading with the raw's current value rather than the figure of record on the asked-for date is the missing asserted_at axis showing up as a measurement rather than an assertion.

One process note worth more than any of the above: the first re-grade run reported R1 0/3, which was wrong — the ad-hoc loop ran under zsh, which does not word-split unquoted $expects on IFS=',', so multi-token expectations collapsed into one literal. The eval itself is #!/usr/bin/env bash and unaffected. Grader changes get verified by re-running the oracle or bash -c, never by a zsh one-liner.

2026-07-27 — R4: 0/7 -> 7/7. The spec gap was the whole story.

Re-ran after the /wiki-query citation-contract fix. 9 raw files, 14 wiki pages, every answer via: wiki.

R4 went from 0/7 to 7/7. Three runs of zero had pointed at the model; the defect was an instruction that never asked for the citation. With the form stated, every answer emitted it, and the loci are tight and correct — raw/sales-2026.csv#L948 for the CSV needle at row 947, raw/thread-q3-planning.eml#L120-L127 for the email needle in message 11, raw/field-report.md#instrumentation-debt-lines-1043-1098 for the report needle in section 21. Nothing about the model's retrieval changed between the 0/7 run and this one.

Everything measured passed: R1 3/3, R2 4/4 (both capacity legs and both silent-memo legs), R4 7/7 — 7/7 questions, 14/14 including citations.

R3 and R5 are INCONCLUSIVE, not failures. Both answer files contained You've hit your session limit · resets 10:40am. The run hit the account cap on its last two questions. The E8 machinery added earlier the same day caught the API-error family but not this phrasing, so both were scored FAIL — again indistinguishable in the report from a refusal defect and a drift-detection defect. RETR_BROKEN_MARKERS now covers session/usage/quota/credit-balance/login forms, and E8 asserts the exact string that bit. Re-graded from cached answers at no cost: 7/7 measured, 2 inconclusive.

That is the third distinct way a non-answer has been scored as a capability failure (529 overload, ENOTFOUND, session cap). The pattern is worth stating plainly: any transport failure that returns text will be graded as text unless something explicitly recognises it, and every such mis-grade looks exactly like the system being bad at its job.

Entities remain unmeasured. eval-entities.sh reports NO ENTITIES CAPTURED against this wiki, and the cause is timing, not capability: the run started 06:49:32 and Step 3.5 was committed 06:56:46. The run's own installed copy of wiki-ingest.md contains no Step 3.5, so entity capture was never requested. N6 earned itself here — reporting "the section was never written" instead of "E1 0% recall" is the difference between a scheduling note and a false bug report. Needs one fresh extract+ingest once the session cap resets.

2026-07-27 — R4 root cause was a spec gap; entity extraction (phase 2) instrumented

R4's 0/5, 0/5, 0/7 was never the model. /wiki-query's output template asked for - Wiki: and - Web: and never asked for an inline (source: raw/<file>#<anchor>) at all. Answers that emitted one were improvising, so the shape varied per run: a backticked path, the path merged into a wikilink parenthesis, or provenance listed only in the Sources block. Every one is unverifiable, because citation-audit.py locates evidence by grepping the literal (source: form. The provenance was real and tight every time — the last run cited raw/retry-budget-memo-feb.md#L20-L22, exactly right — it just was not in the contracted shape. Three runs of evidence pointed at the model; the defect was in the instruction.

/wiki-query now states the form literally, adds a - Raw: line, tabulates the near-miss shapes that actually occurred with why each fails, and requires the narrowest anchor (a citation resolving to 60 lines gestures at a document; one resolving to 2 proves a claim). Guarded by verify-query-citation-contract.sh (Q1-Q5, smoke R25). Q5 is the load-bearing check: it round-trips the taught form through the grader's real extractor, so contract and audit cannot drift apart silently and send R4 back to zero with no visible cause.

Phase 2 (entities) is instrumented but not yet measured. Design decision: entities feed the EXISTING KG rather than a parallel store. type: entity is already a page type and wiki-to-kg.py already builds edges from ## Related links, so a second entity store would be two sources of truth that disagree — the failure the drift lint exists to prevent. Step 4 already handled entities but gated page creation on "2+ raws OR structurally important", which is why a person who made one decision in one email was never captured at all. New Step 3.5 captures entities as cited bullets in an ## Entities section on the summary page; promotion to a full page still respects Step 4's threshold. That gets contextualization without a page per person.

The spec names what is NOT an entity, in a table, because listing every capitalized phrase is the failure mode: it looks thorough, doubles the noise, and makes entity retrieval useless — a list where everything is an entity identifies nothing.

scripts/eval-entities.sh scores it deterministically against a gold set already present in the corpus (the five From: authors in the email thread) and planted decoys (Title Case section headings from the field report). E1 recall >= 80%, E2 decoy capture <= 20%, E3 provenance >= 80% resolvable to a containing passage <= 10 lines. E1 and E2 are each other's mitigation and are never to be read separately: recall alone is won by dumping every capitalized token, precision alone by emitting the two obvious names. verify-entity-eval.sh (N1-N7, smoke R26) builds a wiki that plays each strategy and asserts it LOSES.

Two portability traps hit while building, both worth recording because they produce silent wrong numbers rather than errors: BSD sed has no \| alternation (GNU extension), so name extraction returned the whole bullet and E3 read 0% on correct data; and grep -q under pipefail closes the pipe, the producer takes SIGPIPE, and the pipeline reports failure on correct output. Both now covered by N7 and by N6's variable-capture pattern.

Smoke 29 -> 31 checks (R25, R26), all green. The entity numbers are NOT measured yet — that needs an ingest run against the new Step 3.5.

2026-07-27 — T2 run: as-of survives without prose, and the eval mis-scored a correct answer

Ran the eval against the T2 pair. 9 raw files, 16 wiki pages, every answer via: wiki.

The extract contract holds. /wiki-extract populated asserted_at on all 7 sources unprompted and reasoned correctly about each: end-of-period for the two capacity reports (2026-03-31, 2026-09-30), the bare Published: line for both memos, the thread's last message date for the .eml, and unknown + a note for the two genuinely undated sources (field-report.md, sales-2026.csv.md). fetched_at was 2026-07-27 on every file and appears in zero asserted_at fields — the A5 false pass was not taken once. All 7 anchors resolve to a passage containing their stated date.

T2-silent-asof PASSED. "As of 2026-03-01, the ingest retry budget was 3 attempts per message", citing raw/retry-budget-memo-feb.md#L20-L22. As-of reasoning survives with no relational prose to lean on — which is what T2 existed to find out, and the answer is yes.

T2-silent-current "FAILED" and the answer was right. The run happened 2026-07-27; the second memo was dated 2026-09-22, two months in the FUTURE. The system answered 3 attempts, explained that only the February memo was in effect, flagged the future date as an anomaly worth chasing with the author, and recorded the two memos as a genuine contradiction. My question demanded 7. A corpus with fixed dates cannot ask about "now" — the question has to be answerable from the documents alone, with no reference to when the eval runs. Both dates moved into the past and both "now" questions rewritten to "the most recently published …". E9 now fails the oracle if a now question reappears, and it immediately caught a second instance I had not noticed: R2-current had the same defect and had been passing only because the Q3 report's prose says "supersedes … as the current number" — the relational prose was covering for a future-dated document. Exactly the dependency T2 was built to expose.

R5 "FAILED" on an API error. The answer file contained API Error: Unable to connect to API (ENOTFOUND); the drift logic was never exercised. Graded as FAIL, which in the report is indistinguishable from the system serving a stale claim — the same class of defect as scoring an empty wiki. retr_answer_broken now detects API/network/empty answers, the eval marks them INCONCLUSIVE and excludes them from the denominator rather than counting them as losses, and the report says how many were excluded. E8 covers it. The .r5-pristine restore did work: raw/ was back to 412 and the drift lint clean afterwards, so the poisoning fix from yesterday holds.

R4 0/7. Third consecutive run at zero, and the cause is confirmed as format, not provenance: the T2 answer cited raw/retry-budget-memo-feb.md#L20-L22 — exactly the right locus — but wrapped as ([[page]], source: `raw/…`) rather than the bare (source: raw/…) that AGENTS.md hard rule 4 mandates and retr_citations greps for. The provenance is real and tight every time; the emitted format is not the contracted one. Still out of scope here, still not to be fixed by loosening the grader, and now with three runs of evidence behind it.

Oracle: E1–E9, 12+ checks. Smoke 29 green.

2026-07-26 — valid time: asserted_at enforced (timestamps phase 1 of 2)

The clean run's R2 pass was load-bearing on luck: the corpus states its vintage in prose ("reporting period", "figure of record for Q1", "supersedes ... as the current number"), so an as-of answer could be assembled from phrases without any vintage reasoning. Delete those clauses and the capability goes with them, because nothing structured carried the document's own date. fetched_at is transaction time only.

Added the valid-time axis. asserted_at is the date the DOCUMENT claims for its content, and it is never silently absent: either an ISO date plus asserted_at_source — an anchor into that file's own body where the date appears, resolved through citation-audit.py's grammar, with the date required to be inside the passage — or the literal unknown plus asserted_at_note saying why. unknown is a first-class answer on purpose: plenty of real sources have no discoverable date, and a lint that blocked them would push authors to fabricate one. What is forbidden is silence.

The check that earns the design is A5. Copying fetched_at into asserted_at is the cheapest possible false pass — 100% field coverage, zero information, and every as-of query silently defeated. The mandatory anchor is what prices it out: a date that must resolve to a passage stating it costs more to fake than to read off the page. A5 asserts both the rejection and that the message names the pattern, so the next author to try it is told why.

scripts/asserted-at-audit.py (reuses wikitext.parse_frontmatter and citation-audit.py's resolve_anchor — a second copy of either would let the lint accept anchors the citation audit rejects), scripts/wiki-lint-asserted-at.sh, oracle scripts/verify-asserted-at.sh A1–A11, smoke R24, /wiki-lint check 9, /wiki-extract step 4 populates all three fields, installer manifest. Smoke now 29 checks, all green. Prose date forms are accepted ("March 31, 2026" satisfies 2026-03-31) — an ISO-only lint would force authors to edit raw/, which hard rule 1 forbids. Frontmatter-only writes leave the body hash untouched, so this does not trip the drift lint.

T2 corpus pair added: retry-budget-memo-feb/sep.md, identical bodies apart from one figure, a bare Published: line, and no "supersedes"/"current"/quarter names anywhere — verified zero relational cues. Two questions (T2-silent-asof, T2-silent-current) with headline-anchored forbids. Honest limit, recorded so it is not overclaimed later: the date still lives in the body, so T2 does not prove the asserted_at FIELD is load-bearing — a source with no date anywhere would be unsatisfiable, since /wiki-extract would correctly record unknown. T2 proves the narrower thing: as-of survives with no relational scaffolding to lean on.

Not yet run against T2 — that costs a full extract+ingest and is the next step. Phase 2 (entity extraction, checks E1–E3, feeding the existing KG rather than a parallel entity store) is not started. The AGENTS.md schema addition is a patch at scratchpad/agents-md-asserted-at.patch; context-gate blocks agent writes to that file.

2026-07-26 — clean run: 7/12, and point-in-time actually works

First run where all five checks were measured on one honestly-ingested corpus — fresh install, no cached stage, capacity-report-q1.md ingested at 412 so R5's mutation had a real commitment to violate. 7 raw files, 11 wiki pages, every answer via: wiki.

R1 3/3 (second independent sample), R2 2/2, R3 1/1, R5 1/1, R4 0/5. Score 7/12 — numerically identical to the contaminated run and completely different in meaning.

R2 is the correction. The prior run's failure was called evidence that point-in-time does not work; that was an artifact of the poisoned corpus. On clean data the as-of leg leads with "Sustained throughput as of 2026-04-15 was 412 GB/day" and the current leg with 389, both right, in one run. The honest scope of that capability: the corpus states its vintage in body text ("period through 2026-03-31", "supersedes Q1"), so the wiki is reasoning about vintage from content, not from a valid-time axis. It works because the documents self-describe. A silently-revised document with no stated vintage would still defeat it, and that is what an asserted_at field would fix — so the gap is narrower than claimed but real.

R5 passes convincingly, not incidentally. The answer names both hash prefixes (21dca3d0… vs recorded 2a394578…), identifies the edit as uncommitted and post-ingest, refuses to serve the 999, prescribes /wiki-ingest if intentional, and explicitly declines to git checkout away the user's working-tree change. The drift commitment added earlier today is enforced at query time, not just by the lint.

R4 0/5 is a format-compliance defect, not a missing-provenance defect, and R4 currently conflates the two. The provenance is present and tight — the email needle is cited as raw/thread-q3-planning.eml#L118-L125 — but rendered inside a Sources: list in backticks instead of the inline (source: raw/<file>#<anchor>) form that AGENTS.md hard rule 4 mandates. Only 1 of 7 answers used the required form here; the previous run used it more often. So the finding is that the inline citation convention is applied inconsistently across runs, which citation-audit.py cannot catch either since it only inspects committed wiki pages, not query output. Fixing this belongs in /wiki-query's output contract, and R4 should keep failing until it lands — a grader taught to also accept the backticked Sources form would launder the violation exactly the way restamping ingested_hash launders drift.

2026-07-24 — hash-drift lint (ingest commitments enforced)

ingested_hash was written but never re-checked: re-extract a source (or hand-edit a sidecar) and every (source: raw/<file>#<anchor>) citation keeps resolving to text that no longer supports the claim. /wiki-lint check 4 is page age, check 7 is frontmatter fields, and citation-audit.py only proves the target exists — nothing compared the body against its commitment. Added scripts/wiki-lint-hash-drift.sh (recomputes via body-hash.sh, the canonical hasher; names the drifted file, both hash prefixes, and the ingested_pages now at risk), its oracle scripts/verify-hash-drift.sh (H1–H5), and wired both in: /wiki-lint check 8 (fix = re-run /wiki-ingest, never restamp the hash — that launders the drift), smoke R22 (H5 gates the repo's real raw/), installer manifest. Not-yet-ingested sources (ingested_hash: "") are skipped — that's /wiki-ingest's job, and noise there would train users to ignore the lint. Smoke now 27 checks, all green.

2026-07-07 06:20 — /wiki-ingest

• Processed: raw/okf-spec-v0-1.md (hash 7d4121ee)
• Processed: raw/google-cloud-okf-blog.md (hash 8d5556f6)
• Processed: raw/devsplainers-okf-llm-wiki-video-transcript.md (hash d1d2986d)
• Created: wiki/okf-spec-v0-1-summary.md, wiki/google-cloud-okf-blog-summary.md, wiki/devsplainers-okf-llm-wiki-video-transcript-summary.md, wiki/open-knowledge-format.md, wiki/okf-vs-llm-wiki-bootstrap.md
• Updated: wiki/core-idea.md, wiki/division-of-labor.md, wiki/index.md
• Contradictions flagged: none
• Faithfulness gate: 42/43 SUPPORTED; 1 UNSUPPORTED marker on okf-vs-llm-wiki-bootstrap.md:43 (moat→three-layer paraphrase, anchor #6:15 — left for /wiki-lint per contract)
• Note: both OKF web sources are extraction_status: degraded (blog partial, spec condensed) — full-page WebFetch blocked by context-gate egress hook this session. Re-extract from a fresh session to upgrade to verbatim.

2026-06-22 — vision hardening wave 7 (path-traversal confinement)

Closing wave 6's allowlist exposed a path-traversal hole: (source: raw/../secret.txt#L1) passed both --no-bare-urls (prefix raw/) and C1/C2 (read a file outside raw/, earned coverage) — defeating the "snapshotted in raw/" invariant. Fixed at both points in citation-audit.py (stdlib os.path): _is_allowed_target normalizes (normpath) and requires the target to stay under raw/ (so raw/../x→reject, raw//x→ok); C1 confines realpath(rawpath) to be inside realpath(raw_dir) (escape → c1=False, doesn't resolve). Also reject NUL/control-char citation targets in both _is_allowed_target and audit() before realpath (an embedded NUL raised ValueError, crashing the gate) — now flagged + non-resolving, no crash. Follow-up noted: apply the same confine-to-dir everywhere a wiki string becomes a path (bundle/kg/anki).

2026-06-22 — vision hardening wave 6 (bare-url guard → allowlist)

Re-audit found Wave 5's --no-bare-urls closed the web-cite gap by name, not by class: it matched only ://-scheme and literal www., so protocol-less / bare-host / uppercase cites ((source: cnn.com/x), (source: example.com), (source: WWW.x/y)) bypassed it. Inverted to an allowlist: an inline (source: X) is legal only if X is raw/<file>[#anchor] or exactly analysis; everything else is flagged by construction (no URL parsing). The check now also runs in the faithfulness gate's ingest mode (was promote-only). Removed the dead regex. This closes the last HIGH gap.

2026-06-22 — vision hardening wave 5 (web sources must snapshot to raw/)

Closed V2 gap 3 (web claims invisible to the gate). Promoted web findings can no longer be cited with a bare (source: <url>) — the receipts rule now extends to the web: a web source must be snapshotted into raw/ (via /wiki-extract) before citing, so the claim is raw-backed, coverage-counted, and entailment-checkable. New deterministic guard citation-audit.py --no-bare-urls (flags (source: …://…) / www., leaves (source: analysis) and raw/ alone), wired as smoke R20 (25 checks) and as a promote-mode floor in the faithfulness gate (bare-url page → exit 3). /wiki-query promote prose + AGENTS.md updated; verify-no-bare-urls.sh added. The actual fetch stays LLM-driven (CI has no network); the guard is the keyless mechanical enforcement.

2026-06-22 — vision hardening waves 2–4 (gate, features, discover, edges)

Re-audit raised the vision from 72%→88%→(targeting ≥99%). Wave 2: faithfulness gate fails CLOSED (exit 3) with no judge (--allow-unjudged opt-out + loud warning); /wiki-query promote path now invokes the gate before synthesis; flashcards carry a Source column; KG stamps causal edges sourced:true/false; diagram contract requires raw-anchored footers. Wave 3: CI gains R18 (bundle round-trip) + R19 (real-wiki coverage gate), R3 scoped to *.md + block-HTML patterns (22→24 checks). Wave 4 (closing re-audit residuals): wiki-discover.py now reads the sourced flag and marks causal chains resting on uncited edges ⚠ [unsourced] (was a real HIGH leak — KG stamped the flag but discover ignored it); flashcard citation is now card-local (closed the sticky-citation laundering bypass); citation-audit.py skips fenced code blocks and no longer false-flags empty/frontmatter-only pages; verify-bundle success line + README + SELLING + ci.yml label all corrected to not overclaim.

2026-06-22 — vision hardening wave 1 (floor + coverage + bundle)

Adversarial audit of the 6 vision checks (72% achieved) drove fixes. Floor/coverage (citation-audit.py): skip frontmatter when matching citations; page_frontmatter returns {} on an unclosed fence (no self-exempt); timestamp anchors match on a token boundary (not substring); reject line-ranges inside frontmatter; fail C2 on duplicate-slug ambiguity; anchorless whole-file cites no longer count toward --coverage (a bare file cite could exempt a page of fabrications); journal added to EXEMPT_TYPES. Bundle: package-wiki gains G4 coverage gate, verify-bundle gains B5; both refuse symlinks; verify-bundle hard-fails on missing python3 and runs faithfulness checks over the full tree even under --post-use; optional WIKI_SIGN_KEY detached signature; buyer-facing wording downgraded from "faithfulness" to "citation integrity" (C3 entailment is a write-time seller attestation, not reproducible offline). New verify-bundle-roundtrip.sh.

• Wiki content: karpathy-video-slide-ingest-pipeline-summary.md re-cited with #body-verbatim-numbered-0107 (was an anchorless whole-file cite the stricter coverage gate correctly flagged).

2026-06-22 — citation-coverage gate (vision check #5)

Added the inverse of the citation-audit floor: instead of "do citations resolve?", "does every claim-bearing page carry one?". scripts/citation-audit.py --coverage lists pages with no resolving (source: raw/...) citation and exits 1 if any exist (R17 in smoke-all.sh, test: scripts/verify-citation-coverage.sh). Two exemptions: type: navigation (structural pages point inward) and a new optional provenance: none frontmatter knob for meta pages that make no external claims.

• Marked provenance: none: commands.md, glossary.md, source-attribution.md, synthesis-artifacts.md — each already self-declares as source: analysis design/interpretation, not external claims. (The other 7 source: analysis pages cite raw and stay gated.)
• Schema: additive/opt-in field documented in AGENTS.md; no version bump (per the bump policy).
• Why: codifies the project vision — "a second brain that ships with receipts" — as a deterministic gate, so no feature can quietly add unsourced claims. See README "Vision".

2026-06-09 — schema v2 → v3: synthesis layer

Added a mechanical synthesis layer. /wiki-ingest (Step 8), /wiki-query (on promote), and /wiki-lint --apply now regenerate four derived artifacts via scripts/synthesize/all.sh: wiki/open-questions-dashboard.md, wiki/tensions.md, wiki/decision-timeline.md, and wiki/knowledge-graph.json. Generation is deterministic (byte-identical on no change) and does no LLM work — it only aggregates markers the LLM already wrote (## Open questions sections, > CONTRADICTION FLAGGED flags, log.md headers, [[links]]). The graph JSON reuses scripts/visualize/graph-html.py --json (the same parser /wiki-visualize uses), so JSON and rendered graph can't diverge.

This partially and deliberately re-introduces aggregate-view machinery cut on 2026-06-08 (a76196f, "infrastructure ahead of single-user demand") — but only the narrow, deterministic, zero-LLM-cost core, and pushed onto the read/write-completion path rather than as bespoke per-source LLM synthesis. Did not restore the causal-KG, factory, or self-updating-brain layers.

• Added: scripts/synthesize/{build.py,all.sh}, scripts/verify-synthesize.sh, --json mode on graph-html.py, manifest entries, AGENTS.md "Synthesis artifacts" section.
• Migration note for older clients: a wiki scaffolded before v3 has no scripts/synthesize/all.sh; commands skip Step 8 silently (the four artifacts simply won't exist/update). Re-scaffold with scripts/create-llm-wiki.sh to adopt it. No frontmatter or layer-ownership rules changed; the four artifacts are additive type: navigation pages + one JSON.

2026-06-08 — fix stale verification note in operation-ingest.md

Corrected a stale "honesty note" that claimed the 7-step ingest pipeline was specified, not demonstrated and that /wiki-ingest had never been invoked. Both are false: scripts/smoke-build.sh drives claude -p "/wiki-ingest raw/smoke-source.md" in a real session and scripts/smoke-all.sh runs it in CI on every push (C1 confirms the follow-up query reaches the right answer); scripts/eval-onboarding.sh exercises extract→ingest→query independently. Flipped wiki/operation-ingest.md lines 40–52 to "demonstrated end-to-end and CI-gated," preserving the honest nuance that the original karpathy-derived meta-wiki pages were hand-bootstrapped (only the smoke-fixture pages were machine-ingested) and the residual per-step granularity unknowns. The diagrams/ingest-pipeline.* poster (gitignored, regenerable) carried the same stale line and was corrected locally; it will regenerate correctly from the fixed source.

2026-06-08 — cut to core (factory + brain + causal removed)

Trimmed the system to the core that earns its keep over "a folder + Claude": init / extract / ingest / query / lint + the output commands (visualize / flashcards / diagram) + the blank-wiki installer (create-llm-wiki.sh). Rationale: value-vs-cost review concluded the scale machinery was infrastructure ahead of single-user demand. Full prior system recoverable at git tag archive/full-system-v1.

• Removed — factory: /wiki-new /wiki-skill /wiki-registry (+ new/wikis aliases), scripts/{new-wiki,registry,verify-multi-wiki,verify-skill-install}.sh, templates/skill/. Smoke R5/R10 dropped.
• Removed — self-updating brain: /wiki-learn (+ learn), and this branch's brain/scale guards: scripts/privacy-scan.sh, scripts/hooks/pre-commit, scripts/wiki-near-duplicates.py, their verifiers + fixtures. Smoke R12/R13 dropped; core.hooksPath wiring removed from the installer.
• Removed — causal layer: scripts/{wiki-lint-causal.sh,wiki-to-kg.py,wiki-graph-walk.py,verify-causal.sh,eval-causal.sh}, templates/causal-vocab.txt, causal fixtures, raw/causal-smoke-source.md, causal step 1.5 in wiki-query.md + authoring prose in wiki-ingest.md + the "Causal relations" section in AGENTS.md. Smoke R11 dropped.
• Smoke is now 13 checks (C1–C5 + R1–R8, renumbered contiguous); jscpd ceiling tightened 3.25% → 2.5% (post-cut duplication ~1.9%). AGENTS.md/CLAUDE.md/GEMINI.md + README/QUICKSTART/EXPLAIN + wiki/commands.md/glossary.md updated to core-only.
• Kept (degrade gracefully): the typed-relations evals (eval-multi-hop*) still reference the removed wiki-to-kg.py behind an [ -f ] guard, so they run sidecar-less. presentations/lseg-sales/ (a deck for the removed skill feature) left untouched as a separate artifact.

2026-06-01 — conversion pass: prove it runs, show it, one start-here, CI

First-visitor → adoption polish. The decisive change: the README no longer says "Runtime behaviour is untested" — because it now is tested. The Claude Code happy path is proven end-to-end (scripts/eval-onboarding.sh drives claude -p as a fresh newcomer through extract→ingest→query and confirms they reach the right answer). Durable wins: C1 (pipeline reaches the correct answer) reliably green, and C2 (one unambiguous start-here) structurally fixed by the README-FRESH move. The post-change run scored 5/5; C3 (agent write-before-read) and C4 (AGENTS.md probe) have run-to-run variance, so treat 5/5 as a ceiling, not a guarantee.

• README hero rewrite — one 10-second pitch, a real demo GIF (assets/demo.gif, a genuine /wiki-query against the shipped wiki, recorded via asciinema→agg), CI + License badges, and a single copy-paste "first answer in one block." Disclaimers and the meta-wiki keep/wipe decision moved below the fold.
• Removed the "untested" disclaimer, replaced with the truth: the happy path is verified by smoke-all.sh (14 deterministic checks, in CI) + eval-onboarding.sh.
• One start-here — moved README-FRESH.md → templates/README-fresh.md. There is now exactly one README at the repo root; the fresh-wiki template no longer reads as a competing entry point (this is what flipped onboarding check C2). Updated consumers: create-llm-wiki.sh, verify-create-llm-wiki.sh (R8 still green), eval-onboarding.sh judge path, README tree.
• CI — new .github/workflows/ci.yml runs scripts/smoke-all.sh --no-build (new flag: skip the LLM build phase, verify committed artifacts + R1–R9; no API key / no claude needed). Badge on README line 1.
• No schema change. Deferred (not done): launch kit, per-tool smoke expansion, /wiki-verify, CONTRIBUTING/issue templates, release tag.

2026-05-29 — visual answers: /wiki-query --visual (html/pdf/png)

/wiki-query can now return a diagram of its answer alongside the text, in html / pdf / png, reusing the vendored Infographic-extractor archetype system (no new design system). This ships into generated wikis (it's an operating command), so any produced wiki can answer visually.

• New scripts/visualize/render.sh <html> --pdf|--png — HTML poster → PDF/PNG. Detection: system headless browser (Chrome/Chromium/Edge/Brave, PATH or macOS app bundle) → Node+puppeteer (temp project, fetches its own Chromium) → fail-soft (keep HTML, print install hint, nonzero). PDF single-page (height fit to content); PNG full-page @2×. Functionally verified end-to-end via the puppeteer path (real 152 KB PDF, 1600×2048 PNG). A RENDER_DISABLE=1 seam lets the oracle test fail-soft deterministically.
• .claude/commands/wiki-query.md — added --visual [html|pdf|png] + --archetype <name> and Step 5.5: score all 8 archetypes against the synthesized answer, auto-pick the top (override with --archetype), generate a self-contained HTML poster to diagrams/query-<slug>.html per the generator-contract, then render to pdf/png via render.sh. Text answer is always produced; visual is additive; footer cites wiki pages + any web URLs used.
• .claude/commands/wiki-diagram.md — consistency: --pdf/--png now also run render.sh.
• scripts/installer-skeleton-manifest.txt — ships scripts/visualize/render.sh into generated wikis (installer oracle still green; templates/infographic/* + wiki-diagram already shipped).
• scripts/visualize/verify-visualizers.sh — render.sh existence + deterministic fail-soft check always; functional PDF smoke when a system browser is present (puppeteer not auto-probed — too heavy).
• Docs: AGENTS.md, README.md, README-FRESH.md (ships to generated wikis), docs/VISUALIZATION.md (new render.sh §), docs/QUICKSTART.md, and the 5 cross-tool shims now document --visual / render.sh.

Schema version: unchanged (v2). Additive, opt-in — --visual is a new optional flag; no per-wiki schema change.

2026-05-29 — multi-wiki factory (/wiki-new, /wiki-registry)

Added a factory tier: the repo can now generate other domain-shaped wikis and track them in a local catalog. Built additively on the proven single-wiki installer — scripts/create-llm-wiki.sh, scripts/installer-skeleton-manifest.txt, and scripts/verify-create-llm-wiki.sh were left byte-for-byte unchanged (the installer oracle still passes).

• New scripts (factory-only; not added to the manifest, so generated wikis remain leaves): scripts/new-wiki.sh (composes create-llm-wiki.sh with workspace placement + registry), scripts/registry.sh (owner of registry.jsonl — add/mark-seeded/list/prune over JSONL, no jq/python3), scripts/verify-multi-wiki.sh (oracle M1–M5).
• New commands (factory-only): .claude/commands/wiki-new.md (+ alias new.md), .claude/commands/wiki-registry.md (+ alias wikis.md).
• Layout: workspace default ${LLM_WIKI_WORKSPACE:-~/llm-wikis}, each wiki its own git repo, registry.jsonl at the workspace root; --target <path> escape hatch registers an out-of-workspace wiki by absolute path (in_workspace:false).
• Generate-on-demand domains: no hand-authored domain content ships. /wiki-new --domain "<desc>" scaffolds deterministically, then the LLM authors a ## Domain conventions section + navigation index + 3–5 seed pages. Seed pages are source: analysis with an interpretive disclaimer (no fabricated (source: raw/...) citations) — enforced by the M4 oracle.
• Verification: verify-multi-wiki.sh green on M1–M3 (deterministic scaffold+register, enumerate+drift, escape-hatch+refuse-clobber) plus a committed edge battery E1–E9 (default-workspace resolution, missing --domain, relative --target absolutization, adversarial-domain JSON-escaping/injection resistance, empty-registry listing, mark-seeded/has error paths, non-slug rejection, duplicate-name fail-fast). M4–M5 (seeded schema-validity + domain relevance) green on a real /wiki-new run, and the oracle correctly reds intentional defects (broken link, unlinked seed, provenance gap). A 4-domain generator battery (medieval guild economics, kubernetes operators, competitive bouldering, grief counseling) all passed M4–M5, and the verify-and-fix loop was demonstrated end to end (inject broken link → red → fix → green). Both /wiki-new and /wiki-registry command bodies were exercised as real slash-command invocations.

Schema version: unchanged (v2). The per-wiki schema is untouched; this is a new tier of factory commands that sits outside the generated wikis. Additive, opt-in — no bump per the AGENTS.md policy.

Deferred (phase 2): snapshot (freeze a built wiki into a reusable static template), a remote/published registry, and recursive factories (generated wikis that can themselves generate).

2026-05-28 — /wiki-diagram (semantic output command)

Added /wiki-diagram (+ alias /diagram), the semantic member of the output tier: it takes a natural-language intent, retrieves relevant wiki pages (reusing /wiki-query discipline), scores all 8 diagram archetypes, presents a candidate menu, and on the user's pick generates a self-contained HTML poster to diagrams/. Distinct from /wiki-visualize (mechanical render of existing structure). Wiki-read-only; no web search/promotion by default.

• New: .claude/commands/wiki-diagram.md + alias diagram.md.
• Vendored (self-containment, no external skill dep): templates/infographic/{archetypes.md, scoring-rubric.md, generator-contract.md, example-poster.html} — from FrancyJGLisboa/Infographic-extractor. License pending upstream (add MIT/Apache to that repo).
• .gitignore: added diagrams/, wiki-graph.html, anki.csv (generated artifacts; also closes the prior litter gap from the visualize/flashcards tier).
• Manifest: 6 new files added to scripts/installer-skeleton-manifest.txt (verifier I4a).
• Docs: output-tier tables in AGENTS, README, README-FRESH, CLAUDE, GEMINI, copilot, clinerules, cursor, wiki/commands, QUICKSTART now list three commands, with the mechanical-vs-semantic boundary stated.

Schema version: unchanged (v2) — additive, opt-in. Untested: the command's prompt-body (retrieval + 8-lens scoring + HTML generation) is unrun; first real /wiki-diagram invocation is its smoke test.

2026-05-28 — output-command tier (/wiki-visualize, /wiki-flashcards)

Added two output commands that render/export an already-built wiki, sitting outside the five-command lifecycle loop. Both are read-only on raw/ and wiki/ (they write only new output artifacts) and are thin LLM dispatchers over existing scripts — no parsing logic reimplemented.

• New: .claude/commands/wiki-visualize.md + alias visualize.md — dispatch to scripts/visualize/{graph,mermaid,slides,serve}.sh.
• New: .claude/commands/wiki-flashcards.md + alias flashcards.md — wrap scripts/wiki-to-anki.sh.
• Manifest: added the 4 command files to scripts/installer-skeleton-manifest.txt so fresh installs ship them (verifier I4a).
• Docs: AGENTS.md, README.md, README-FRESH.md, CLAUDE.md, wiki/commands.md (+ cross-tool shims) now document the output tier alongside the five; resolved the dormant /wiki-export open question in wiki/commands.md.

Schema version: unchanged (v2). Per the AGENTS.md bump policy, new commands are a strictly opt-in addition (older clients simply lack them) — additive, not breaking, so no bump. Recorded here per the schema-versioning "record the change" guidance.

2026-05-27 17:45 — phase-2 kg-traversal

verdict: phase-2b-abandoned-on-gate. baseline 7/7 (100%, >> 40% threshold). C5 gate FAILED.

Implemented .scratch/phase-2-kg-traversal/GOAL.md Phase-2a only. New artifacts:

• tests/eval/sparse-fixture/ — 7 hand-crafted pages of fictional supply-chain entities (zerlon, quirpal, bryntex, mordax, velnar, thalox, glivex). Each page body ≤ 100 words, no body sentence states direction or position. 8 typed ## Related lines using 7 distinct verbs (feeds, powers, produces, ships-via, terminates-at, succeeds, replaces).
• tests/eval/sparse-multi-hop-questions.md — 7 multi-hop questions, all tagged baseline-absent: true, all hops: 2 or hops: 3. Expects-tokens (upstream/downstream/transitive/parallel/unreachable) are NOT verb literals and are absent from the unstripped fixture prose (C4(d) + C4(e) both green).
• scripts/eval-multi-hop-sparse.sh — thin variant of eval-multi-hop.sh pointed at the sparse fixture. Stays sidecar-less in Phase 2a; would conditionally invoke scripts/wiki-to-kg.py in the typed work dir (only) if Phase 2b were active.

Gate result (./scripts/eval-multi-hop-sparse.sh):

• baseline: 7/7
• typed: 7/7
• delta: 0
• verdict: null-result on the eval itself; gate verdict: FAILED (baseline >> 40%)

Reading. The LLM extracts enough signal to answer the direction questions from sources the verb-strip does not touch:

1. Connectivity survives stripping. Stripped ## Related lines keep slug-pair links — the baseline LLM still sees that zerlon connects to mordax, mordax connects to velnar, etc. Direction has to come from somewhere else, but only some-where else is needed.
2. Page semantics leak through tags. tags: [sparse-fixture, alloy] on zerlon, tags: [..., vessel] on mordax, tags: [..., port] on velnar are intact in both variants. The LLM uses tag semantics + world knowledge ("ore feeds refinery feeds alloy ships to port") to recover direction even without verbs.
3. The question text is a teacher. Every question names the verbs explicitly ("Following only the forward supply edges (feeds, powers, produces, ships-via, terminates-at)"). Even when those verbs are stripped from the markdown, the LLM reads the verb list in the prompt and applies it to the connections it sees.

Conclusion: typed relations do not provide measurable retrieval signal even on a fixture engineered to need them. The signal the typed graph carries is already encoded — redundantly — in connectivity + tags + question framing. A parallel KG layer would record what the LLM already infers. Not justified.

Decision per §7: Phase 2b abandoned. No scripts/wiki-to-kg.py written, no modifications to .claude/commands/wiki-ingest.md or .claude/commands/wiki-query.md, no wiki/_kg.jsonl artifact. C10 absence guards intentionally green.

Future avenues (not pursued here, recorded for phase-3 design):

• Strip tags AND verbs in the baseline (not just verbs) — would test whether tag semantics alone explain the result.
• Make question text neutral (do not name the verbs in the prompt) — would test whether the prompt is the teacher.
• Use questions whose answers require numerically combining attrs across hops (attrs get stripped; the LLM cannot reconstruct them from tags or world knowledge).
• Revisit the parallel-KG idea only after one of the above produces a measurably stricter discriminator.

The user-research signal recorded today — "users expect /wiki-ingest to create a KG" — is not addressed by this work and should not motivate building a KG on a null result. It is a phase-3 design input.

C1 regression-oracle: ./scripts/smoke-all.sh exit 0; ./scripts/eval-multi-hop.sh on the rich-prose Brazilian-ag fixture verdict null-result (no regression vs. phase-1's 5/5 = 5/5).

2026-05-27 14:00 — phase-1 typed-relations eval

verdict: null-result (baseline 5/5, typed 5/5, delta 0).

Implemented .scratch/typed-wikilinks-semantic-viz/GOAL.md (annotated ## Related lines with optional verb + attr; pure CommonMark; backward-compat preserved by treating untyped and multi-link lines as implicit related-to). New scripts/wiki-lint-typed-relations.sh validates the verb regex ([a-z][a-z0-9-]*). scripts/visualize/graph-html.py extended: per-edge verb field in the JSON, a <select id="verb-filter"> UI + filterByVerb() JS handler + per-verb edge colouring via d3.schemeCategory10.

Fixture: 6-page Brazilian-agriculture wiki built during the 2026-05-27 new-user simulation, frozen at tests/eval/wiki-fixture/. Typed verbs applied to 14/19 single-target Related lines (74%); 6 distinct verbs (researches-for, defined-by, credit-for, complement-of, enables, summarized-in). Verb tokens were chosen to be absent from the baseline prose (0 hits each) so stripping them removes a real signal.

Eval (scripts/eval-multi-hop.sh, 5 multi-hop questions, 3 tagged baseline-absent: true):

• baseline: 5/5
• typed: 5/5
• delta: 0

Reading: typed verbs alone did not improve /wiki-query accuracy on this fixture. The LLM is smart enough to infer the typed relationship from surrounding prose context even when the verb token is stripped — on Q1-Q3 (the baseline-absent set), the baseline LLM still produced the exact kebab-case verb (enables, credit-for, complement-of) via paraphrase + format-mimicry of the question. On a Wikipedia-derived wiki with rich prose, the marginal information added by typed verbs is dominated by what the prose already encodes implicitly.

Implications for phase-2 (parallel knowledge graph):

1. Typed-verb-in-markdown is not enough on its own; the eval signal is masked by LLM inference from prose.
2. Either the eval needs sparser-prose fixtures (so typed verbs are the only path to the answer), or phase 2 must add explicit graph traversal to /wiki-query (the LLM gets the typed-relation graph as a separate retrieved structure, not just inline markdown).
3. The null-result is itself the answer to the original "do we need a parallel KG?" question: on rich-prose wikis, typed verbs are redundant; on sparse-prose wikis or for multi-hop queries beyond the LLM's inference horizon, they may help — but only when paired with traversal logic, not just syntax.

All 9 success checks (C1–C9) green; installer regression (verify-create-llm-wiki.sh) and core smoke (smoke-all.sh) both still exit 0. scripts/installer-skeleton-manifest.txt extended by 13 lines so the new lint, eval harness, fixture, and canary tests ship in fresh installs.

2026-05-26 10:56 — visualization toolchain landed

Three-iteration build (per .scratch/visualization-tools/GOAL.md) shipping scripts/visualize/ — a bespoke Python+D3 graph generator (stdlib only, no npm), plus npx-wrapped slides (marp-cli), mermaid (mermaid-cli), and a python3 -m http.server wrapper. Includes tests/canary/graph-fixture/ (4 nodes/4 edges, flat) and tests/canary/graph-fixture-nested/ (2 nodes/1 edge with sub/leaf.md — anti-gaming against non-recursive parsers). Oracle at scripts/visualize/verify-visualizers.sh runs 5 sub-checks; skip-when-absent semantics for the npx tools. Documented in docs/VISUALIZATION.md; recommended heavier alternatives (Quartz, mdBook, SilverBullet) noted but not bundled. scripts/installer-skeleton-manifest.txt extended (44 lines) so visualizers ship in fresh installs. First end-to-end run: all 5 smokes green on a machine with both marp-cli and mermaid-cli reachable via npx. No schema change.

2026-05-26 09:17 — installer toolchain landed

Per .scratch/installer-fresh-skeleton/GOAL.md. New: scripts/create-llm-wiki.sh (manifest-driven scaffolder generating a fresh repo at <target-dir>), scripts/verify-create-llm-wiki.sh (oracle: I3 manifest iteration + I4 tree-shape + content tripwire + frontmatter parse + I5 target preflight + I4(d) template substitution), scripts/installer-skeleton-manifest.txt (single source of truth, 31 lines initially, later extended), README-FRESH.md and wiki/index-FRESH.md (fresh-skeleton templates), tests/installer-output/.gitignore. Removes the wipe-meta-wiki.sh step from the new-user flow. Post-adversarial-pass revisions to the spec: I4 reformulated from negative-spot-list (gameable by wholesale cp -R) to positive-shape + content-tripwire + frontmatter-parse triple; I4(d) added to byte-match installed README.md/wiki/index.md against FRESH templates. No schema change.

2026-05-26 08:00 — smoke infrastructure landed

Per .scratch/plug-and-play-curator-smoke/GOAL.md. New: scripts/smoke-build.sh (LLM-driven, idempotent via body-hash.sh), scripts/smoke-check.sh (pure-shell asserts C1–C5), scripts/smoke-all.sh (umbrella: build + check + 4 regression guards), scripts/r3-obsidian-patterns.txt (patterns file for the no-Obsidian-syntax guard, avoids backtick quoting hazards in shell), tests/smoke/smoke-source.md (fictitious Phase Coherence Engineering fixture with Quortex protocol / Dr. Alma Voss / 47 phase rotations anchors), tests/smoke/expected-query.md, tests/smoke/.gitignore. Adversarial-pass introduced C5 (ingested_hash populated cryptographic proof) — without it, an agent could hand-author log.md + last-answer.md to satisfy 6 of 8 checks without running claude -p. First smoke run produced the 06:52 /wiki-ingest entry below — the empirical demonstration that the 7-step pipeline executes correctly. No schema change.

2026-05-26 06:52 — /wiki-ingest

• Processed: raw/smoke-source.md (hash ba2159c8)
• Created: wiki/smoke-source-summary.md, wiki/quortex-protocol.md, wiki/dr-alma-voss.md, wiki/phase-coherence-engineering.md
• Updated: wiki/index.md
• Contradictions flagged: none (smoke fixture's domain is disjoint from the existing meta-wiki about the LLM-wiki pattern; nothing to disagree with)

wiki/quortex-protocol.md carries the literal "Quortex" and "47 phase rotations" anchors required by smoke check C2.

2026-05-26 05:30 — schema bump 1 → 2

Three additions to the schema landed together. Each is opt-in on its own, but the journal exception introduces a new rule on /wiki-ingest behavior — by the bump policy in AGENTS.md, that's a behavior-changing edit and requires a version bump even though no v1 client would currently violate it in default repos.

• Journal exception (wiki/journal/): new user-owned directory. Files there are not rewritten by /wiki-ingest. New type: journal value added to the page-type enum. Template at templates/journal-entry.md. The directory is reserved by an empty .gitkeep; entries are created by the user, not the LLM.
• ## Flashcards content convention: any wiki page may declare Q/A pairs in a ## Flashcards section. Exporter at scripts/wiki-to-anki.sh emits an Anki-importable CSV with the page slug as the card tag. Canary fixture at tests/canary/canary-flashcards.md; smoke test via scripts/verify-wiki-to-anki.sh.
• Optional MCP read surface: scripts/mcp-server.sh launches @bitbonsai/mcpvault pointed at wiki/, exposing read/search/write tools to any MCP-aware client. Setup details in docs/MCP.md. Pure addition; does not change the three-layer model or the slash commands. scripts/preflight.sh now reports whether npx is available.

Migration note for v1 clients. A v1 client that scans wiki/**/*.md without journal awareness will see entries under wiki/journal/ as ordinary wiki/ files. If such a client runs /wiki-ingest autonomously and decides a journal entry needs rewriting (e.g., as part of step 4 "update existing pages"), it could clobber user-authored content. Mitigations until clients upgrade: (a) ingest sources from raw/, not from journal entries, (b) if running v1 ingest on a repo that has journal entries, snapshot wiki/journal/ first, (c) /wiki-lint is unaffected — link-checking still works against journal entries.

2026-05-25 17:15 — backfill extraction_method on legacy raws

Gap surfaced during PR #3's smoke-test batch: the two raw files shipped before PR #1 (karpathy-llm-wiki-video-transcript.md and karpathy-video-slide-ingest-pipeline.png.md) predate the extraction_method frontmatter field, so running verify-extract.sh on them fails the "extraction_method required" check. This is a one-time human-authorized migration to bring the legacy raws up to schema_version 1.

• raw/karpathy-llm-wiki-video-transcript.md: added extraction_method: passthrough. The transcript was pasted into the conversation that bootstrapped the project (per its own notes: field) — passthrough is the closest match in the enum (raw text imported as-is, no parser).
• raw/karpathy-video-slide-ingest-pipeline.png.md: added extraction_method: llm-vision. The slide is a binary PNG; its sidecar was produced by vision extraction (per the 2026-05-25 08:30 entry below).
• Both edits are in the YAML frontmatter only; body content unchanged. Body hashes were recomputed via scripts/body-hash.sh and verified identical to the recorded ingested_hash — idempotence preserved, /wiki-ingest will still skip both files.

Hard-rule note: modifying files in raw/ normally violates the LLM's read-only rule on that layer. This migration is treated as a one-time maintainer-authorized edit (not an autonomous LLM act). Future cases where the LLM's /wiki-extract writes extraction_method on a new raw file are fine and unrelated.

Follow-up gap surfaced during this commit: the verifier's ingested_hash check warns "/wiki-extract should leave this empty" — accurate for fresh extract output, but noisy on files already processed by /wiki-ingest (the normal post-ingest state). Not fixed here to keep this PR's scope minimal. Tracked for next iteration.

2026-05-25 17:00 — verify-extract.sh: surface extraction_status visually

Gap surfaced during the prior session's smoke-test batch: the DOCX-degraded run produced a sidecar with extraction_status: failed, but the verifier reported "Passed. Shape checks all green." with no visual indication that extraction actually failed. Shape was fine — but the user reading the verifier output wouldn't know they need to install pandoc.

• scripts/verify-extract.sh: new check block after ingested_hash. When extraction_status is present in the frontmatter, emit ✓ for ok, ⚠ for degraded / failed, ⚠ (with "unknown value" message) for anything else. Exit code unchanged on degraded/failed (shape is still fine — the warn is the signal).
• Docstring updated to document the new behavior in the "Scope" section.

Tested on 4 synthetic frontmatter values: failed → ⚠, ok → ✓, degraded → ⚠, bogus → ⚠ (unknown). Exit code 0 in all cases.

2026-05-25 16:50 — tests/canary/canary-csv.csv: tracked CSV fixture

Smoke-tested the CSV path of /wiki-extract end-to-end this session. The fixture used (5-row world-cities CSV) was a one-shot file in working tree; promoting it to a tracked fixture so the smoke test is reproducible.

• tests/canary/canary-csv.csv — 5 data rows + header. Plain ASCII. Designed to exercise the ≤100-row markdown-table render branch.
• Sibling to tests/canary/canary-smoke-test.md (plain-text fixture from PR #3). Same purpose: known-good input for the verifier.

The smoke test flow (per docs/QUICKSTART.md):

/wiki-extract tests/canary/canary-csv.csv   # in AI tool
./scripts/verify-extract.sh canary-csv      # in shell

Expected: raw/canary-csv.csv (copy) + raw/canary-csv.csv.md (sidecar with extraction_method: csv-passthrough and a markdown table of the 5 cities). Verifier green.

2026-05-25 15:30 — canary + verify-extract.sh: smoke test for first /wiki-extract

"Specified, not demonstrated" was the honest status of /wiki-extract after PR #1. This commit ships the smallest possible shape test: a known-good canary source + a shell verifier. The user runs /wiki-extract on the canary in their AI tool, then runs the verifier in shell to get a green/red signal on whether the produced output has the expected frontmatter and body. Shape only — not semantics.

• New tests/canary/canary-smoke-test.md: tiny plain-markdown source (~30 lines) used as the known-good input. Self-describing — explains what should happen and what shouldn't.
• New scripts/verify-extract.sh <slug>: pure bash verifier. Locates the produced raw/<slug>.<ext> or sidecar <slug>.<ext>.md, parses the frontmatter, checks that required fields (source_url, source_type, fetched_at, extraction_method) are non-empty, that ingested_hash is present-but-empty, and that body content exists. TTY-aware coloring, exit 0 on pass / 1 on fail / 2 on usage error.
• Honest scope: shape only. Verifier can catch: missing output file, malformed frontmatter, absent/empty required field, empty body. Verifier CANNOT catch: wrong source_type value, hallucinated source_title, extraction_method recorded incorrectly. Semantics need a human eye on raw/<slug>.*.
• docs/QUICKSTART.md: new "Smoke test (recommended)" section between "Before you start" and "The 5 operations" walking the user through the canary flow.
• README.md: project-layout tree updated with the new script and the tests/canary/ directory.

Also in this commit: drift cleanup. The QUICKSTART operations table (separate from the per-tool sequences below it) still used the old verb names fetch and ask — leftover from PR #1's sed which only touched wiki-fetch/wiki-ask literals. Updated to extract/query to match the rest of the doc.

Verified on the working tree (without touching tracked files): 4 cases — no raw file → exit 1 with "Not ready" message; happy-path raw → exit 0 with all green; missing extraction_method → exit 1 with one fail line; malformed frontmatter (no closing ---) → exit 1 with multiple failures. Cleanup confirmed raw/ returned to its 3 tracked files.

2026-05-25 15:15 — scripts/wipe-meta-wiki.sh: clean-slate helper

Today's "start fresh" flow was three commands in QUICKSTART (rm -rf wiki/*.md raw/* && touch wiki/index.md) — easy to typo and easy to skip the index touch, leaving the next /wiki-extract confused. Added a single-command helper.

• New scripts/wipe-meta-wiki.sh: pure bash, mirrors body-hash.sh/preflight.sh style. Wipes wiki/*.md and raw/*, recreates wiki/index.md as a minimal stub with valid frontmatter, resets log.md to its header line.
• Safety: interactive [y/N] confirmation by default; --yes flag to skip. Inventories file counts before wiping so the user sees what's about to disappear.
• Idempotent: re-running on an empty wiki shows "nothing to do" and exits 0. (Actually re-wipes the stub index and rewrites it; functionally idempotent.)
• Preserves: AGENTS.md, README.md, all shims, .claude/commands/, scripts/, docs/, LICENSE — everything that isn't generated content.
• README.md + AGENTS.md + docs/QUICKSTART.md: replaced the three hand-rolled rm -rf instructions with the new script. Project-layout tree in README updated.

Verified on a /tmp/ copy: before (23 wiki files, 3 raw files) → after wipe (only index.md stub remains; raw/ empty; log.md reset). Idempotence checked — second run with --yes produced same end state.

2026-05-25 15:00 — AGENTS.md: schema_version = 1 declared

Started declaring a schema version on AGENTS.md so future schema changes have a coordination marker. Today's AGENTS.md (as merged in PR #1, before this change) is retroactively designated as the version-1 baseline. Future bumps are reserved for breaking/behavior-changing edits.

• AGENTS.md: added **Schema version:** 1 (introduced 2026-05-25) line near the top. Added a "Schema versioning" section before "When in doubt" describing the bump policy (breaking-change bumps only; additive opt-in changes don't bump; no runtime enforcement V1).
• No code changes. Slash commands today don't read this field. It's a marker for humans reviewing diffs and for future tooling.

Migration impact: none. Existing slash commands continue to work unchanged.

2026-05-25 14:30 — scripts/preflight.sh: fail-fast tool & permissions check

User-side install today is git clone && cd — no health check. Failures of /wiki-extract only surface at first invocation, sometimes silently (e.g., missing pandoc falls back to python-docx, which the user may not have either). Added a preflight script that probes the environment before the user runs any slash command.

• New scripts/preflight.sh: pure bash (mirrors the style of body-hash.sh), TTY-aware coloring, platform-aware install hints (brew on Darwin, apt/dnf/pacman on Linux).
• Three check tiers:
  • Hard requirements (bash/awk/openssl/git + write permissions on raw/ and wiki/) — fail fast with exit 1 if missing.
  • Recommended optional tools (pdftotext/pandoc/xlsx2csv + python-docx/openpyxl fallbacks) — warn only.
  • AI runtimes (claude/cursor/code/copilot/gemini) — informational.
• Summary line classifies extraction coverage: "Ready. Full first-try: ... Partial: ... Degraded: ...". Maps directly to the extraction_method matrix in AGENTS.md.
• Manual-run only. No auto-execution, no post-install hook, no agent-side invocation. The user runs it when they want a snapshot.
• README.md: new optional-but-recommended line under Install pointing at the script; project-layout tree updated.
• AGENTS.md: one-line mention near the body-hash.sh reference so the agent knows the script exists and can suggest it when users hit extraction_status: failed.
• Verification: ran on this host (macOS, partial tool coverage). All hard reqs green, summary correctly classifies PDF as full / XLSX as partial / DOCX as degraded. Piped output drops ANSI codes correctly. Exit code 0 when hard reqs met.

This PR stacks on PR #1 (feat/extract-rename) because the preflight references the renamed /wiki-extract and the new extraction_method matrix in AGENTS.md. Merge PR #1 first.

2026-05-25 10:45 — /wiki-extract gains DOCX / XLSX / CSV handlers + PDF LLM-vision fallback

Real new behavior in the (just-renamed) /wiki-extract command. Previous coverage was URL, plain text, image, and PDF-via-pdftotext. New coverage adds DOCX (via pandoc), XLSX (via xlsx2csv), CSV (passthrough + markdown-table preview), and an LLM-vision fallback path for PDF when pdftotext is missing or returns near-empty output.

• .claude/commands/wiki-extract.md: step 1 (format detection) and step 3 (acquisition) extended with the new formats. Step 4 (frontmatter spec) gains extraction_method and extraction_status fields. Tool-availability check (command -v) is now mandatory before invoking any optional binary.
• AGENTS.md: new "Supported source formats and extraction" subsection with the format → handler matrix. File-naming rule updated to cover DOCX/XLSX (binaries → sidecar) and CSV (tabular text → sidecar). Frontmatter YAML block updated with the two new fields and the new source_type values.
• wiki/commands.md: /wiki-extract spec rewritten to mirror the new behavior table; tool-policy paragraph added.
• wiki/open-questions.md: new top entry under "Operational questions" — "Are the new extraction handlers actually correct?" — listing concrete unknowns per format. Same posture as the existing "do the 7 steps actually happen?" entry.
• README.md, docs/EXPLAIN.md: one-line updates to the /wiki-extract row to mention multi-format support.
• docs/QUICKSTART.md: no content update needed (the rename sed already swept it; existing examples still work).

Verification status: all four new handlers (DOCX/XLSX/CSV/PDF-LLM-vision) are specified, not demonstrated. First real /wiki-extract on each format is the smoke test. Same as the 7-step pipeline.

Principle preservation: every shell binary is optional with a declared fallback (pandoc→python-docx, xlsx2csv→openpyxl, pdftotext→llm-vision). A user with zero shell tools installed gets degraded but functional extraction. BYO-AI guarantee intact.

2026-05-25 10:30 — verb rename: fetch→extract, ask→query

User-facing mental model survey: new users naturally describe the workflow as "extract content from this file" and "query the wiki," not "fetch" and "ask." Renamed the two slash commands and updated all references across the repo. Pure mechanical rename — no behavior change.

• Renamed: .claude/commands/wiki-fetch.md → wiki-extract.md
• Renamed: .claude/commands/wiki-ask.md → wiki-query.md
• References updated in 22 other files via sed: AGENTS.md, CLAUDE.md, GEMINI.md, .clinerules, .cursor/rules/llm-wiki.mdc, .github/copilot-instructions.md, README.md, docs/QUICKSTART.md, docs/EXPLAIN.md, docs/pitch-vscode.html, .claude/commands/wiki-init.md, .claude/commands/wiki-lint.md, wiki/commands.md, wiki/division-of-labor.md, wiki/glossary.md, wiki/index.md, wiki/ingest-pipeline.md, wiki/karpathy-llm-wiki-video-transcript-summary.md, wiki/layer-raw-sources.md, wiki/operation-query.md, wiki/query-as-write-loop.md, wiki/source-attribution.md, wiki/three-layer-architecture.md.
• Strategy: hard rename. Repo is V2 and runtime-untested with no external users — aliases would be permanent bloat.
• Gate verified: grep -rEn 'wiki-(fetch|ask)\b' . --exclude-dir=.git returns only historical references inside this log.md (the 2026-05-25 06:45 bootstrap entry, which records the original file names at creation time).
• What did NOT change: the 7-step ingest pipeline, the body-hash algorithm, the three-layer model, frontmatter conventions, or the command bodies' behavior. Two prompts now carry their new names internally (/wiki-extract $ARGUMENTS and /wiki-query $ARGUMENTS); everything else is identical.

Follow-up commits on this branch (feat/extract-rename) will (a) expand /wiki-extract to handle DOCX, XLSX, CSV with graceful tool-chain fallback, and (b) refresh docs prose where the rename diff isn't sufficient.

2026-05-25 09:35 — docs/QUICKSTART.md: per-tool first-use sequences

User asked: "will users get a fluid experience?" Honest answer was no, for most. Highest-leverage fix identified: a per-tool first-use guide. Built it.

• New docs/QUICKSTART.md, English, ~260 lines.
• Structure: prereqs (incl. what to do with the shipped meta-wiki — keep / wipe / archive) → 5-operation mental model → per-tool sequences for Claude Code, Copilot CLI, VSCode + Copilot Agent Mode, Cline, Cursor, "other tools" → expected output (with symptom→cause→fix table for partial pipeline execution) → recovery via git → cost expectations → honest caveat about untested runtime → quick reference card.
• Each per-tool section gives the exact natural-language phrasing to invoke each workflow, since non-Claude-Code tools don't have slash commands. This was friction #1 from the earlier honest assessment.
• "What success looks like" section addresses friction #3 (no expected-output gabarito) by listing concrete observable outcomes (1 summary page, 3-10 wiki pages, log.md entry, index updated).
• README.md gains a callout to QUICKSTART right after the install command, plus the file is listed in the project layout.

Still NOT addressed by this commit:

• The /wiki-init --replace-content affordance for wiping the meta-wiki cleanly (friction #2). For now QUICKSTART tells the user the manual command.
• Recovery beyond git checkout (no --dry-run, no rollback command). Documented as a caveat, not solved.
• The actual runtime untested-ness (friction #4 + #5). Cannot be solved without someone running the system.

2026-05-25 09:15 — fix: pitch HTML had outdated Copilot claim

User correction: I claimed "Copilot é conversacional, não agentic" in the AI-extension tier table. Wrong — (a) GitHub Copilot CLI is a standalone agentic CLI competing with Claude Code, (b) VSCode + Copilot in Agent Mode (shipped 2025) does multi-step autonomous work.

• Rewrote the AI-extension tier table in docs/pitch-vscode.html. New top tier (★★★★★) groups Copilot Agent Mode, Copilot CLI, and Cline as peers. Continue/Roo/Cody one step below. Copilot Chat classic (without Agent Mode) demoted to ★★★. "No AI extension" stays bottom.
• Added project memory ai-tool-capability-claims to enforce: don't make blanket capability claims about AI tools from training data; verify first. Indexed in MEMORY.md.
• README.md needed no change (it already listed Copilot CLI as a peer tool — I just contradicted myself in the pitch).

2026-05-25 09:00 — pitch artifact: docs/pitch-vscode.html

User asked how I'd present the project to a VSCode-only audience. Wrote the pitch in chat, then user asked if it was persisted as HTML. It wasn't. Created docs/pitch-vscode.html — self-contained single-file pitch in PT (audience = self / internal reference, per follow-up clarification).

• Single file, zero external dependencies (no CDN, no JS, no remote fonts).
• Dark/light theme via prefers-color-scheme.
• Print-friendly (Cmd/Ctrl+P → PDF).
• Content: hook, 30s premise, 5-step flow, AI-extension tier matrix (Cline > Continue/Roo/Cody > Copilot Chat > nothing), 4 concrete problems, honest caveats, 3 implicit principles, closing slide.
• README.md updated to list docs/ in the project layout.

This is documentation, not part of the wiki layer — it lives in docs/ because it's a one-shot artifact for external presentation, not a knowledge page that grows via ingest/lint. No frontmatter, no link convention. Pure HTML by user choice.

2026-05-25 08:45 — codified verification-gap honesty

User asked the sharp epistemic question: do the 7 steps actually happen, or are they just specified? Direct answer: the latter — no LLM has invoked /wiki-ingest in this project; the wiki was hand-written in the design conversation. To make this surface in the wiki itself rather than only in chat:

• Added "Verification status (as of 2026-05-25)" subsection to wiki/operation-ingest.md, naming what is and isn't tested.
• Added a new section at the top of wiki/open-questions.md under Operational questions, titled "The most important open question: do the 7 steps actually happen?".

No body content was changed in any other page; the existing 7-step descriptions stand as specifications. The wiki is now honest about which of its claims are demonstrated and which are aspirational.

2026-05-25 08:30 — second raw source (slide) + project-wide rename CHANGELOG.md → log.md

User shared the "What happens when you ingest a source" slide from the same video as a screenshot. The slide is more prescriptive than the spoken transcript in two places: it names the index file as index.md and the log file as log.md, and it adds that step 6's index entry is "a catalog entry with link and one-line summary."

• New raw source: raw/karpathy-video-slide-ingest-pipeline.png (binary) + raw/karpathy-video-slide-ingest-pipeline.png.md (sidecar with vision-extracted text + visual description). Hash cfe8e91a. Dogfoods the image-acquisition convention (binary + sidecar) for the first time.
• Project-wide rename: CHANGELOG.md → log.md to honor the source video. Touched 18 files: AGENTS.md, README.md, CLAUDE.md, GEMINI.md, .clinerules, .cursor/rules/llm-wiki.mdc, .github/copilot-instructions.md, all 5 .claude/commands/wiki-.md, and 6 wiki/.md pages. The rename is recorded in wiki/source-attribution.md as a video-aligned decision (previously a project deviation).
• Refined wiki/ingest-pipeline.md steps 6 and 7 to match the slide's exact wording: step 6 now says "catalog entry with link and one-line summary"; step 7 explicitly names log.md. Both steps now carry a second source citation pointing at the slide.
• Created wiki/karpathy-video-slide-ingest-pipeline-summary.md (type: summary, source: video) — per-source recap of the slide per the ingest-pipeline's step 3 convention.
• Created wiki/karpathy-llm-wiki-video-transcript-summary.md (type: summary, source: video) — the missing summary page for the original transcript. Backfills a convention violation from the initial bootstrap (no summary page existed for the first source).
• Updated wiki/index.md with a new "Source summaries" section listing both summary pages.
• Updated wiki/source-attribution.md to add the slide as a second source and to record the CHANGELOG.md → log.md rename as video-aligned.
• Updated raw frontmatter on both sources: transcript's ingested_pages adds the new summary; slide sidecar's ingested_hash/ingested_at/ingested_pages populated.

Net: 2 raw sources, 23 wiki pages. Body hash of transcript unchanged (3054546f) — still skippable on next /wiki-ingest. Slide sidecar now has its ingested_hash populated.

2026-05-25 08:05 — follow-up: extracted "knowledge compounds" as a first-class page

User observation: the compounding-of-knowledge point (LLM doesn't restart from scratch per question; cross-refs pre-built) is one of the most important arguments in the video, but it was scattered across core-idea, problem-with-naive-rag, operation-ingest, and query-as-write-loop rather than living in a dedicated page.

• Created wiki/knowledge-compounds.md (source: mixed) — synthesizes the compounding thread from three angles (negative-RAG case, positive-wiki case, the per-source compounding effect) and names the two engines (ingest + query-as-write-loop). Citations to transcript timestamps 1:12-1:30, 1:36-1:55, 5:30-5:40.
• Updated Related sections in: core-idea.md, problem-with-naive-rag.md, operation-ingest.md, query-as-write-loop.md, ingest-pipeline.md, division-of-labor.md, four-principles.md — all now link to knowledge-compounds.
• Updated wiki/index.md under Foundations + the literal-reading-order list.
• Updated raw/karpathy-llm-wiki-video-transcript.md ingested_pages to include the new page. Body hash unchanged (3054546f), so future /wiki-ingest runs still skip the source.

This is the kind of follow-up /wiki-lint would surface as a "gap" — a heavily-referenced concept without its own page. The next time /wiki-lint runs, it should report zero such gaps for "compounding."

2026-05-25 07:10 — V2 multi-tool portability shims added

• Created .cursor/rules/llm-wiki.mdc — Cursor rules pointing at AGENTS.md + workflows.
• Created .clinerules — Cline shim.
• Created .github/copilot-instructions.md — GitHub Copilot shim.
• Created CLAUDE.md — shim for older Claude Code versions that load CLAUDE.md instead of AGENTS.md.
• Created GEMINI.md — shim for Gemini CLI.
• All shims reference AGENTS.md as canonical; they exist to give every tool's auto-loader something to discover. The Claude Code slash commands in .claude/commands/wiki-*.md remain the only first-class implementations; other tools invoke the same workflows via natural language.
• Updated README.md with the per-tool support matrix and revised project layout.

2026-05-25 06:45 — bootstrap (manual ingest)

Initial bootstrap of llm-wiki-bootstrap. The repository's first wiki was produced by an interactive design session rather than by /wiki-ingest, but the result honors the same conventions and the raw source is now ready for re-ingest at any time.

• Processed: raw/karpathy-llm-wiki-video-transcript.md (hash 3054546f)
• Created:
  • wiki/index.md (navigation)
  • wiki/core-idea.md, wiki/problem-with-naive-rag.md, wiki/three-layer-architecture.md (foundations)
  • wiki/layer-raw-sources.md, wiki/layer-wiki.md, wiki/layer-schema.md (architecture)
  • wiki/operation-ingest.md, wiki/operation-query.md, wiki/operation-lint.md, wiki/ingest-pipeline.md (operations)
  • wiki/division-of-labor.md, wiki/four-principles.md, wiki/query-as-write-loop.md, wiki/use-cases.md (foundations)
  • wiki/commands.md, wiki/implicit-constraints.md, wiki/open-questions.md, wiki/source-attribution.md, wiki/glossary.md (analysis)
• Created (system files):
  • AGENTS.md (canonical schema)
  • README.md (install + quickstart)
  • .claude/commands/wiki-init.md, wiki-fetch.md, wiki-ingest.md, wiki-ask.md, wiki-lint.md
• Contradictions flagged: none
• Notes:
  • raw/karpathy-llm-wiki-video-transcript.md is a third-party YouTuber's walkthrough, not Karpathy's tweet verbatim. See wiki/source-attribution.md.
  • 6 pages marked source: analysis: 5 analytical content pages (commands, implicit-constraints, open-questions, source-attribution, glossary) plus 1 navigation page (index).
  • 3 pages marked source: mixed (layer-raw-sources, layer-wiki, layer-schema — video extraction + project-specific convention details).
  • 11 pages marked source: video (all literal extractions from the transcript).
  • The raw file's frontmatter has been updated with ingested_hash, ingested_at, and ingested_pages so future /wiki-ingest runs skip it unless the body changes.
  • scripts/body-hash.sh ships as the canonical hash algorithm. The recorded ingested_hash was computed via this script; future /wiki-ingest runs must use the same script (per AGENTS.md) for idempotence.
  • Slash command runtime is NOT yet validated. All five .claude/commands/wiki-*.md files exist and are well-formed, but no command has actually been invoked. The first real-session invocation will be the smoke test. The bootstrap of the wiki itself was done by direct file writes during the planning conversation, not by /wiki-ingest.

2026-08-21 — AI workspace and single evidence intake

• Added AI-WORKSPACE.code-workspace, which foregrounds START-HERE.md, EVIDENCE-INBOX/, BRIEFS/, and REVIEWS/ while hiding compiler internals from the daily Explorer view.
• Added /ctx-add (/add) as the sole first-run evidence action for files, folders, URLs, pasted text, and the drop folder. Existing /ctx-inbox, /ctx-extract, and /ctx-compile contracts remain available as advanced controls.
• Added a deterministic scripts/add-evidence.py boundary for safe, atomic local staging with recursive folders, collision handling, no-op detection, and symlink rejection.
• Rewrote generated-workspace onboarding around ordinary work requests and automatic saved outputs. ADVANCED.md keeps the full compiler visible when an operator needs it.
• Added installer, package, smoke, and regression coverage for the AI workspace assets and intake contract.

2026-08-22 — Local VS Code Context Workspace extension

• Added a local-first VS Code extension that embeds a clean compiler template and exposes creation, one evidence intake, briefs, deltas, evidence chains, historical state, exception review, health checks, and exact citation navigation.
• Pasted evidence now supports --text-stdin, keeping evidence bodies out of process argument lists while preserving the original /ctx-add contract.
• Added a versioned product/security specification, enterprise VSIX deployment guide, locked packaging dependency, VSIX content verifier, CI artifact build, and extension regression suite.
• The extension stores no credentials, emits no telemetry, requires workspace trust, asks before URL acquisition, and never replaces deterministic compiler validation with model judgment.

2026-08-22 — Universal evidence intake and compiler-owned checkpoints

• Made Add Evidence open the file/folder chooser directly and added sidebar drop plus clipboard text or URL-batch intake.
• Any regular local file is preserved before extraction; unfamiliar binaries remain available and are reported as degraded instead of silently discarded.
• Added scripts/checkpoint-context.sh, which commits only compiler-owned state after successful validation, leaves unrelated staged work untouched, never pushes, and degrades safely when Git is unavailable.
• Reframed README, quickstart, generated guidance, extension documentation, and the GitHub page around one daily intake action and outcome-oriented views. /ctx-extract, /ctx-compile, and manual Git remain advanced controls.

2026-08-22 — Self-service specialization builder

• Added a conversational /ctx-create-profile lifecycle that interviews for domain meaning, scaffolds safely, customizes from synthetic examples, validates, previews behavior, records explicit approval, activates, checkpoints, and ends at Add Evidence.
• Added deterministic profile scaffolding and readiness tools. Readiness exposes technical validity, behavioral coverage, and domain-owner approval separately; self-service profiles cannot activate until all three pass.
• Added Context Workspace actions for Create Specialization and Check Specialization, keeping schemas, tests, commands, and Git behind the AI-chat surface.
• Added fresh-compiler end-to-end coverage for scaffold → behavioral examples → approval → activation → first evidence, plus distribution and documentation assertions.