Skip to content

delegation-and-review §2: scope a sweep by every generator, not one grep pattern - #50

Merged
F-e-u-e-r merged 7 commits into
F-e-u-e-r:mainfrom
firaen22:mine-sweep-scope-generator-surfaces
Jul 22, 2026
Merged

delegation-and-review §2: scope a sweep by every generator, not one grep pattern#50
F-e-u-e-r merged 7 commits into
F-e-u-e-r:mainfrom
firaen22:mine-sweep-scope-generator-surfaces

Conversation

@firaen22

@firaen22 firaen22 commented Jul 21, 2026

Copy link
Copy Markdown
Contributor

What

delegation-and-review §2 (The dispatch packet): scope a find-and-fix-every-
instance sweep by every surface that can generate the target pattern, not by
one search pattern.

Why

Private incident: a 53-file styling-token sweep, three review rounds, and a
merged fix all missed the actual defect — it lived in a shared global utility
class the token grep pattern never touched. Each follow-up round's "still
broken?" re-ask surfaced a different category (a color-tier band, a
class-emitting helper function) that the prior round's search pattern
structurally excluded — the sweep converged only by exhausting the user's
patience for re-asking, not by a scope definition that covered the space up
front.

Rule

(this Rule summary tracks the current diff; original filing in edit history)

  • Branch first: textual invariant (string absence from an explicitly named corpus = the packet's readable scope) -> the scoped search IS the gate, claiming corpus-level absence only. Behavioral target -> generator SEARCH scope hunted per §3's miss-is-costly loop (finders quoted verbatim) with at least one producer/effect axis mandatory before closure; per-round query/result ledger in the packet; searching stays inside the packet's readable scope.
  • Value family closed only by a verified-finite source (sealed enum/const union at file:line), else bounded; unclosable producer surfaces or variation axes -> non-exhaustive outcome; every-instance claims with unobserved members are false. WRITE scope stays the owned files; outside generators reported, never edited.
  • Proof gate (behavioral): observed effect per inventoried surface across declared variation axes, in a side-effect-contained harness with per-invocation authorization at fire time; equivalence classes need verified equivalence including the producing implementation; unobserved = reported unverified.

Ships unprobed per the README covenant.

Checks

  • python3 .github/checks.py — all pass.
  • Diff (as restructured across review rounds — see trail comment): §2 scope +
    proof-gate amendments, one §3 cross-reference line, one Provenance paragraph.

firaen22 and others added 7 commits July 21, 2026 22:13
…rep pattern

A find-and-fix-every-instance sweep needs the same rigor as any other
owned-scope field. Private incident: a 53-file styling sweep, three review
rounds, and a merged fix all missed the real defect — it lived in a shared
global utility class the token grep pattern never touched — and each
follow-up round's "still broken?" surfaced a different category (a
color-tier band, a class-emitting helper function) the prior round's search
structurally excluded.

Rule: enumerate every generator (literals, shared/global definitions, class-
emitting helpers) and the full value space before dispatching or accepting
such a sweep, and gate acceptance on observing the effect rather than on the
search returning zero. Ships unprobed per the README covenant.
…al review

PR F-e-u-e-r#50 round-1 gate (grok-4.5 high: FIX 1,2,3,4; gpt-5.6-sol max:
FIX 1-5). Reproduced and addressed:
- search scope vs write scope split: inspecting a generator does not
  own it; out-of-scope generators are reported, never edited on
  discovery (codex F2)
- generator inventory from definition surfaces, never the token grep;
  unresolved discovery gaps block exhaustive claims (grok F2 / codex F1)
- acceptance moved into the Proof gate field and bound to every
  inventoried surface and tier, declared partitions the only shortcut;
  zero-hit search demoted to a report (grok F1, F4 / codex F3)
- open-ended families get a declared bound per S3's bounded-sweep
  clause instead of a forbidden-sampling mandate (grok F3 / codex F4)
- pos/neg pair added at the proof gate; dangling 'not that the defect
  is' parallel completed (grok n5, n6)
- S3 sweep bullet now points at the S2 sweep fields so the review path
  lands on the acceptance rule (grok F4)
- codex F5 (contributor Checks self-vouch) rejected-with-reason:
  factual process self-report, independently re-verified by this gate;
  packet rubric clarified to stop this recurring false-positive class

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…y closure via S3's own loop

PR F-e-u-e-r#50 round-2 gate (grok-4.5 high: FIX 1,2,5; gpt-5.6-sol max:
FIX 1,2,3,4,5,7). All reproduced and addressed:
- inventory gains a concrete build procedure (axis-diverse searches)
  and a closure criterion borrowed from S3's discovery loop (two
  consecutive empty passes); the token-grep ban is now aimed at the
  completeness INFERENCE, not at using search - fixes the direct-
  literal flip (grok F1, F2 / codex F1)
- partition shortcut replaced: equivalence-class observation with the
  shared-outcome reason stated; unobserved = reported unverified,
  never folded into exhaustive (codex F2)
- S3 pointer now quotes the canonical report-not-the-gate sentence
  verbatim instead of paraphrasing it (codex F3)
- pos example covers every inventoried surface class (codex F4)
- WRITE-scope referent pinned to the explicitly listed owned
  files/modules (grok F5 / codex F7)
- PR body Checks line to be updated to disclose the S3 amendment
  (grok n4 / codex F5); codex n6 (in-packet unverifiability of
  checks.py claim) recorded - the gate-runner re-runs checks.py
  independently each round

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…axis, verbatim loop cite, textual carve-out

PR F-e-u-e-r#50 round-3 gate (grok-4.5 high: FIX 1,2; gpt-5.6-sol max:
FIX 1,2,3,4). All reproduced and addressed:
- at least one inventory axis must never match the target's spelling
  (effect-source definitions/constructors) - the axis set was
  spelling-seeded and could not reach the cited utility class (grok F1)
- S3 anchors cited by real name with the loop quoted verbatim ('Stop
  only after two consecutive empty rounds; one clean round is not
  convergence'), a round defined as every-axis-ran, dedup imported
  (grok F2 / codex F1 / grok r3 n5 of r2)
- equivalence classes need verified branch-free shared paths, not a
  stated reason (codex F2)
- textual-invariant carve-out: an exact-spelling sweep over a declared
  corpus is gated by the scoped search itself; effect-gate scoped to
  behavioral targets (codex F3)
- bounded/gap-carrying sweeps return a non-exhaustive outcome; scope
  reduction needs the dispatcher's say; every-instance claims with
  unobserved members are false (codex F4)
- inventory neg example added; pos example names per-surface
  observations (grok n6 / codex n5); target forms glossed (grok n4)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…y, equivalence evidence, contained observation

PR F-e-u-e-r#50 round-4 gate (grok-4.5 high: FIX 1,2; gpt-5.6-sol max:
FIX 1,2,3,4). All reproduced and addressed:
- textual carve-out guarded: only when the declared corpus covers every
  inventoried surface AND no inventoried generator can emit the target
  without the literal; token sweeps with emitters stay on the effect
  gate (grok F1 / grok n6)
- value family closes only via its named canonical source (file:line)
  or the same discovery loop, else bounded - the color-tier axis of the
  incident (grok F2)
- rounds must add a broadened or new query; identical reruns count for
  nothing (codex F1)
- equivalence requires evidence across inputs, backing data, and
  downstream context; branch-free control flow alone is not
  equivalence (codex F2)
- observation runs inside an authorized side-effect-contained harness;
  undrivable outward effects reported unverified + escalated, never
  fired for the gate (codex F4)
- 'behavioral target' qualifier restored at both categorical zero-hit
  statements and the S3 pointer, which now also carries the positive
  half of acceptance (codex F3 / grok n4); scope-side pos example
  added (grok n5)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…vs target, ledgered hunt, clean textual discriminator

PR F-e-u-e-r#50 round-5 gate (grok-4.5 high: FIX 1,2,3; gpt-5.6-sol max:
FIX 1-6). All reproduced; the fix removes guarantee surfaces instead
of adding conditions:
- TARGET defined as the defect/effect, spellings demoted to probes;
  trigger gains task-phrasing synonyms (grok F1, F3)
- generator hunt delegated to S3's miss-is-costly loop as written
  (axis-diverse finders, dedup, stop rule) instead of restating a
  two-axis floor with its own round semantics (codex F1 / grok F3)
- per-round query/result ledger (empty rounds included) carried in the
  packet; seed inventory + worker-continues resolves the packet-
  closure-vs-delegation tension (codex F5 / grok r5 n6)
- value-family closure needs a verified-finite source; extensible
  registries/configs are never closed (codex F2)
- gate covers declared variation axes where the outcome can differ;
  untested combinations reported unobserved (codex F3)
- textual branch rewritten to a single discriminator (string-absence
  IS the invariant; nothing produces the outcome without the exact
  spelling) - removes the self-contradicting emitting-generator clause
  (codex F4 / grok F1)
- S3 pointer reduced to a pure pointer, no acceptance restatement
  (grok F2); unprobed marker added on the proof-gate half (grok n5);
  53-file unit drift fixed in both examples (codex F6)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
…mandatory producer axis, honest closure

PR F-e-u-e-r#50 round-6 gate (grok-4.5 high: FIX 1; gpt-5.6-sol max: FIX 1-7).
All reproduced and addressed:
- textual-vs-behavioral branched FIRST; textual gate claims corpus-
  level absence only, corpus = packet's readable scope named
  explicitly, never the seed hit list (codex F3 / grok n4-of-r6)
- producer/effect axis made a MUST before loop close; spelling-only
  axis sets non-compliant regardless of rounds (grok F1)
- S3 finder list quoted verbatim; dedup + stop rule cited, not
  restated (codex F4)
- closure honesty: producer surfaces/variation axes not closable from
  a verified-finite source -> non-exhaustive outcome (codex F1);
  verified-finite wording covers const unions (grok n5-of-r6)
- equivalence requires the producing implementation itself; unproved
  divergence -> per-member observation (codex F2)
- search bounded to the packet's readable scope; outside surfaces are
  reported gaps (codex F6)
- outward effects carry per-invocation authorization at fire time
  (codex F7)
- PR-body Rule sync next (codex F5)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GhLu6NCGzM9HxnqNb87MAx
@F-e-u-e-r
F-e-u-e-r merged commit 3864137 into F-e-u-e-r:main Jul 22, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants