Integrate contributor PRs #62-#65: four doctrine rules, covenant-completed, 5-round three-family gate - #67
Merged
Merged
Conversation
…d-pass signature A two-sided good/bad validation proves a grader for one invocation shape at one time. Reusing it later without re-validating let a directory-vs-file argument mismatch make every candidate throw MODULE_NOT_FOUND before its code ran, while the harness's own error-tolerant edge-case branches (accepted-throw) silently absorbed the load failure and got credited — producing an inverted signature (edge cases pass, happy path fails) and a known-bad score of 2/6 against a recorded 0/6. The recorded expected score was the only tripwire. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
…, not evidence Extends the completion-claim auditing bullet: a claimed FAILURE deserves the same reproduction as a claimed success. A subordinate CLI's own sandbox can fabricate a RED gate exactly as easily as a model fabricates a green one — acting on it unreproduced either reverts working code or teaches the next session to distrust RED gates in general. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
…d endpoint vs auth/gateway A single run can't classify empty/dead-looking model output. Two ladders: (1) a single endpoint returning no bytes needs own-parse -> auth/gateway -> same-key alternate-model differential before it's called dead; (2) a model empty on some batch tasks needs a re-run — if the empty slot MOVES between tasks it's a transport flake (demote to supervised), if it stays put it's a genuine capability gap. They require opposite routing responses. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
…retire older bold verdicts
Updating the top summary after a default/order flip leaves older evidence
blocks with their own bold imperative verdicts ("KEEP X AS DEFAULT") intact
lower in the file — a future reader, or a weaker model executing the file
top-to-bottom, can hit the stale verdict first and follow the superseded
order. After any flip, grep the whole file for the superseded term and tag
each stale block explicitly rather than deleting it.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
…ording All four contributed rules are behavioral and shipped without the README covenant's required label — each now carries an in-body \`unprobed\` (private incident as shape) marker plus a dated Provenance entry naming the incident shape and the owed executable probe. Wording: gtg rule-2 "hours on in"→"hours later in"; d-a-r §3 failure-claim rule ties the re-run to the isolated-copy discipline and adds the cannot-re-run path (unverified claim, recorded as a gap, never acted on as a verdict); d-a-r §1 differential names the rule-above connective and softens the flake disposition to supervised-or-retry-wrapped vs dropped; s-a §4 sweep names itself the instruction-file analog of op-rigor §3's call-site sweep and drops the glyph mandate from the tag template. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…cted on in-file precedent Convergent pair (grok F1=luna F2): the cannot-re-run RED now binds fail-closed (gate UNKNOWN — no revert, no clean ship, escalate). Convergent pair (grok F2b=luna F3): ladders refine the diagnosis, never the pinned-string adverse binding — unverified work still does not route. grok F2a lead reworded; grok F3+luna F5: shifting-empty hedged to intermittent-cause + per-task parser rule-out + two-runs-floor; luna F1: reuse re-proof names both valid solutions + known-bad; luna F4: controlled differential (same canary/params, route attribution); luna F6 + grok N5: alias grep + read verdict blocks (§5 cross-ref) + realistic meet-the-stale-verdict story; grok N4: ladder-A exit dispositions (your-side-first, UNRESOLVED never dead). Rejected with counter-evidence: grok N6/luna N7 (marker period+placement matches GTG rule-3 sibling verbatim), luna N8 (plural "incidents" factual — two incidents — and verbatim-matches both adjacent §1 markers). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…eutralization Convergent (grok F2 = luna F6): the SUPERSEDED tag now rewrites the stale verdict LINE in place per s-a §3's own correct-in-place rule — the old imperative stops being one; example updated to match. grok F1: ladder-B disposition subordinated to the property binding + "usable but unreliable" replaces the capability claim. luna F1: the ON-RECORD baseline carries per-reference scores + exact invocation. luna F2: escalation exempted from the re-run prerequisite (it is the unresolved case's outlet). luna F3: re-run in the environment the gate's contract targets; env split = a finding. luna F4: per-task rule-out widened (limits, collection loss). luna F5: target attempt must be evidenced — silence alone leaves UNKNOWN, not dead. sol verdict pending; its findings triage against this state. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…already closed by the grok+luna fold
Three-family convergence note: sol F6 = luna F6 = grok F2 (in-place
neutralization) — three independent finds of one defect, closed once.
sol F1: reuse baseline is per-CASE outcomes + invocation + corpus
identity, and a missing record STOPS reuse (fresh two-sided proof first).
sol F4: controlled differential now isolates the TARGET ROUTE, with the
attributed status deciding model-outage vs per-model block on your side
(entitlement/quota/params) — remedies differ. sol F5: batch ladder
classifies per TASK, never per run ({A,B} then {A,C} = stable A +
intermittent B/C). sol F2 fold: unexplained environment split leaves the
gate UNKNOWN (three words on the luna-F3 fix). sol F3 = luna F2 fixed
same round (escalation exempt from re-run prerequisite).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
sol F1: reuse baseline records invocation SHAPE (ephemeral values normalized) + a deliberate-change re-baseline path (fresh two-sided proof) — literal-identity read would have stopped every CI reuse. sol F2: one green re-run refutes a deterministic RED only; RED-then-green unexplained = flake finding, never noise. sol F3 + luna F1 (convergent on the ✅ artifact): example now route-isolates and checks the attributed status before "outage"; luna's angle folded — a silent alternate leaves the differential incomplete (own block possible) → UNRESOLVED, "your side" claim dropped. sol F4: Provenance no longer equates repetition with capability gap (stable per-task failure) nor pre-isolation "model-specific outage". Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…-fold residue) sol F1: a completed work-path differential now outranks the auxiliary gateway screen (a /models-restricted key must not veto route isolation). sol F2: the ✅ example rests the outage verdict on positive evidence (a server-side failure signature) vs an account/request block incl. parameters; nondiagnostic status stays UNRESOLVED. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Integrates firaen22's #62 (ground-truth-gates rule 2: reuse-time grader
re-validation), #63 (delegation-and-review §3: reproduce FAILURE claims
too), #64 (delegation-and-review §1: empty-output differential diagnosis),
and #65 (skill-authoring §4: superseded-verdict sweep) as one combined
branch (ascending merges preserve the contributor's commits; precedent
PR #59), plus maintainer integration commits.
Integration fixes (beyond the four originals)
no
unprobedlabel — each now carries the in-body marker plus a datedProvenance entry naming the incident shape and the owed executable
probe (README covenant, second branch). Probe debt: these four join the
round-5 probe-suite candidates.
cannot-re-run RED path binds fail-closed (gate UNKNOWN, no revert, no
clean ship, escalation exempt from the re-run prerequisite); the
differential ladders subordinate to §1's pinned-string adverse binding
(diagnosis, never the binding); per-TASK batch classification; the
reuse baseline records per-CASE outcomes + invocation shape (ephemeral
values normalized) and a missing record stops reuse; the SUPERSEDED
tag rewrites the stale verdict line IN PLACE per s-a §3's own
correct-in-place rule.
Cross-model gate (owner-directed cadence: luna ultra + grok high every
round; sol max on every 3rd round and the pre-commit round)
~25 findings folded (every one reproduced before fixing; fixes authored,
never pasted); 3 format nits rejected with in-file counter-evidence
(marker punctuation/placement matches the GTG rule-3 sibling verbatim;
plural "incidents" is factual and matches both adjacent §1 markers).
Three-family convergences: RED-cannot-re-run fail-safe (grok r1 = luna
r1), in-place neutralization (grok r2 = luna r2 = sol r2 — three
independent finds of one defect), the ✅ HTTP-000 example (sol r3 = luna
r3).
python3 .github/checks.pygreen at every fold commit. No versionbump (contributor doctrine batch; precedent #37/#38/#59).
Merging this branch marks #62-#65 merged by reachability.
🤖 Generated with Claude Code
https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx