docs(m4): start-of-stage brief — vocabulary collateral strip, D19–D22 for freeze - #10
Merged
Conversation
… for freeze M4 added by Kyle's post-M3 pick (2026-07-28): close M3's stated bound (collateral among 12, not across the vocabulary) before write-up + /seed-hunt. S1/S2 stretches declined and banked as idea #13 in the j-lens-proj-ideas backlog. Brief contents: 12 primes x 180 probes frame (D19), single-clause VOCAB-SPARING level gate on the never-measured non-subset pool (D20), treatment of the four cross-mention confound cells found while drafting (D21), D18 bars widened to all 60 scored words (D22). Realized ns computed from recorded artifacts with the frozen oracle; both inherited obligations dispositioned explicitly. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01ANtkrPFv6i9CgCasojqcoZ
F1: own the 0.5 gate constant as new, uncalibrated and deliberately lenient rather than "carried from M3" — the statistic changed (12-fold conjunction vs per-cell; 0.5 on the conjunction ≈ 0.944 per-cell under independence) and the status changed (D17 qualifier that could never rescue a claim → sole dispositive gate). The per-cell equivalence and the deletion-count dependency go into GATE_WORDING itself; new deviations row. F2: own D9(b)'s span-truncation residual, which re-enters a gate-bearing arm for the first time since M1 now that probes widen to all 60 — 0 / 3 / 4 span-filling gate-arm items (1.5B trumpet-1, beetle-1, butterfly-1; 3B trumpet-1/2/3, butterfly-1), biasing toward the gate where the 1.5B margin is one item. Adds an instrument fact, a power-table row, a deviations row and a pre-registered residual-conservative recomputation (never dispositive), and corrects the claim that D22 pins the oracle's premises — it pins the span premise, not the scope premise. oracle.py untouched. F3: restate zero-gating (25 / 7 / 5 concepts) as competence-selection, not tokenizer geometry — under D9(b) the gate is token-count-insensitive up to 3 and the recorded clean spans are wrong answers, not truncations. States the upward bias on the floor and connects it to F12's selection-enrichment. F4: correct "cells no milestone has ever measured" to 486 / 844 / 993 (of 492 / 852 / 1,008); the remaining 6 / 8 / 15 are M1-recorded — the very cells used as the proxy — so they are in-sample and cap the gate arm at 35/41, 69/71, 82/84 before any forward pass. Power-table row split, ceilings stated as pre-registered fact, out-of-sample framing tempered. HANDOFF.md mirrors the F1 and F4 corrections. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
Both are defects of the F2 fix, not of the finding it addressed.
F11: the residual-conservative selector was degenerate and its counts were
wrong. "The span fills the 3-token window" matches every recorded cell —
greedy_3 IS the 3-token span. The real condition is narrower: the span, after
stripping leading whitespace, equals the concept's spelling with nothing
following it, so no boundary is observed. On that reading the gate-arm
residual is 0 / 2 / 2, not 0 / 3 / 4 — 1.5B beetle-1, butterfly-1; 3B
trumpet-3, butterfly-1 — which is exactly the set oracle.py's frozen docstring
already names ("beetle-1, butterfly-1 x2 arms, trumpet-3"). trumpet-1 (1.5B)
and trumpet-1/-2 (3B) record 'Trumpet<|im_end|>' and <|im_end|> closes a word
under oracle._BOUNDARY, so they carry no residual; generated 'Trumpet' is two
tokens where bare 'trumpet' is three, which leaves room for the terminator.
Corrected in all three places (instrument fact, power table, deviations row).
F12: the conservative read's denominator was left to the runner, and the two
readings straddle the bar (same numerator: wilson(43,71) lower 0.489 fails,
wilson(43,69) lower 0.505 passes). Pre-registers fail-in-place — the arm stays
at its as-scored n (41 / 71 / 84, preserving the power table's cross-check)
and residual-affected items score as failures — and names and rejects the
un-gating alternative as the less conservative of the two.
F13-F15 (nice-to-have) remain FOLLOW-UP pending Kyle's pull-in call.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
Kyle's call at the freeze gate: pull all eight in before D19-D22 freeze,
since four of them get byte-frozen into code once he signs.
F5: re-scan the cross-mention confound with D5's OWN rule (prefix +
forbidden_forms, not whole-word) — five pairs, not four. Adds Egypt->beetle-2
("Ancient Egyptians carved amulets of the scarab..."), which a whole-word scan
misses because the clue inflects the prime. beetle-2 is ungated on all three
subjects ('Scarab'/'scarab'), so it carries no gate-bearing cell today; it is
listed so a future re-gate cannot silently acquire one. D21 retitled, list
frozen in its text, deviations row and power-table row updated.
F6: replace the proxy-only 0.5B prediction with the in-statistic number.
M3's recorded 0.5B subset reads 19/28 = 0.679 under M4's own gate statistic,
Wilson lower 0.4934 — it ALREADY FAILS M4's bar. So there is no "tension"
between M3's clean 0.5B floor and an M4 failure: M3's floor was clean only
under the cluster-mean per-cell statistic, and switching to the conjunction
flips it. Sharpest illustration of D20's point that the constant carries no
meaning across statistics.
F7: move the realized ns outside the underpowered conditional — the proposed
wording literally read "41 / 71 / 84 < 20". Now abstract, with realized n in a
trailing parenthetical, matching m1_battery.GATE_WORDING's shape.
F9: add the re-certification precondition to GATE_WORDING. The single-clause
level bar, read alone, is satisfiable by a no-op intervention (~100% survival
prints VOCAB-SPARING). The guarantee existed in D19's design and the runner's
exit code but not in the sentence a write-up quotes; now the sentence carries
its own precondition.
F10: PROJECT.md Next action and README status now point at M4 pending freeze
instead of contradicting HANDOFF in the same tree.
F13: name all three zero-gating mechanisms, not one — answers something else,
answers correctly behind a modifier the opening-word rule refuses ('Peking
duck', 'Electric guitar', 'Insect Beetle'), or misses on morphology. The
upward-bias conclusion is unchanged.
F14: the 0.5B ceiling (35/41, Wilson lower 0.716) is far above the bar and is
not a reason to expect failure; the reasons are M3's in-statistic 19/28, the
0/6 proxy and M1's 33/69 control cell. Removes a "so" that contradicted the
preceding sentence.
F15: re-wrap the paragraphs the earlier fix commits broke.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
F16 [should-fix]: the F11 fix swapped a selector matching EVERY cell for one matching NONE. "Equals the scored concept's spelling exactly" is case-exact; recorded residual spans are 'Beetle' / 'Butterfly' / 'Trumpet' while roster spellings are lowercase, so a verbatim runner would select zero cells and turn the sole mitigation for a bias-toward-the-gate into a no-op. Verified: over all 540 cells x 3 subjects, case-exact selects 0; case-insensitive selects exactly the pre-registered 0 / 2 / 2. The selector now states case-insensitive comparison "exactly as oracle.says_concept_prefix compares (re.IGNORECASE)", and says why the case rule decides the set rather than leaving it as a detail. F17: the F10 fix rewrote each file's forward-looking prose but not the status lines above it, leaving PROJECT.md and README.md self-contradictory (each was internally consistent before). PROJECT.md's Status paragraph and README's Status line now both carry M4-in-flight, and README's milestone table gains an M4 row with S1/S2 relabelled banked. F18: "exactly the set oracle.py's docstring names" was four cells against the oracle's stated six — the docstring enumerates across arms (butterfly-1 x2 arms: clean and control_late). Restated as the gate-arm members of that set, with the arm/subject scoping spelled out. F15 re-fix: the :383 wrap was missed and the F13 fix introduced a new orphan at :105. Both repaired; full re-scan of all four files is clean. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
Kyle froze all four decisions 2026-07-29 at the review gate: D19(a) full 12 x 180 strip + clean re-run (2,340 cells/subject), with the embedded 255-cell M1 and 468-cell M3 re-certifications graded first. D20(a) the single-clause survives-all-12 level gate as written. D21(a) all five cross-mention pairs kept in the gate-bearing pool with named per-cell reporting. D22(a) both run-time instrument bars over all 60 scored words + the 12 direction words. F8 (the verdict-label choice, deliberately left open as a design question rather than fixed as a defect) resolved: the label stays VOCAB-SPARING and the realized survival proportion rides INSIDE the verdict string — "VOCAB-SPARING (k/n survive all 12 = <rate>; Wilson 95% lower <lo>)". This is M3's ON A DAMAGED FLOOR pattern applied to a level bar: the label is what a write-up quotes and it is frozen into every results JSON, so the number that qualifies it travels with it rather than living in prose. The brief now states why — at the 1.5B pass point 27 of 71 gated items (38%) are damaged by at least one deletion, and at the nominal 0.5 it would be half. Freeze note added to the brief per the M3 precedent; full DECISIONS.md entries D19-D22 land with the M4 code PR, per the M0-M3 pattern. PROJECT.md next action moves to building the runner; HANDOFF.md records the freeze. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
Line-by-line patching kept shifting the break point; re-flowed both whole paragraphs at 78 columns instead. Also drops a line in HANDOFF that still said "code only after freeze" now that D19-D22 are frozen. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
F19: the freeze commit updated PROJECT.md and HANDOFF.md but not README.md, so
the front page still said "decisions not yet frozen" while the other two said
FROZEN — in a repo whose whole discipline is freeze-before-code, that answered
"were D19-D22 frozen before the run?" with "no". Both README spots corrected.
F20 [should-fix, amends the frozen D20 wording]: the verdict-string clause as
frozen carried only the AS-SCORED proportion, so the two pre-registered reads
that can flip which number is honest — residual-conservative fail-in-place and
the concept-level collapse, each closed in the brief with "the ... numbers are
the honest ones to quote" — stayed in prose. That is precisely the failure the
F8 resolution was adopted to prevent ("prose is not what gets quoted; the label
is"). The window is live: at 1.5B the bar needs k >= 44 (wilson(44,71) lower
0.50342), and fail-in-place removes the 2 residual items, so at an as-scored k
of 44 or 45 the JSON would have printed VOCAB-SPARING ... lower 0.503 while the
brief's own pre-commitment said 42/71 (lower 0.47539) was honest. M3's
precedent is a CONDITIONAL qualifier attached by the runner (m3_matrix.py:179);
M4 had borrowed the sentence shape and dropped the mechanism. D20 now carries a
pre-declared AS-SCORED ONLY qualifier, attached whenever a conservative read's
Wilson lower bound is below 0.5 while the as-scored read's is not, naming each
such read and its number. Per D17's carried rule the qualifier scopes a claim
and can never create or rescue one; the gate, its 0.5 bar, its arm and its
re-certification precondition are all unchanged.
F21: "the verdict, whatever it is" was followed by a template hard-coding the
pass label, and no failing label appeared anywhere in the brief. NOT
VOCAB-SPARING is now named in the wording.
Recorded in the brief's freeze note as an amendment POST-FREEZE, PRE-RUN — the
M3 precedent (its own round-4 F15/F16 amendment) — and FLAGGED FOR KYLE'S
RATIFICATION IN THE MERGE BRIEF.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
Kyle ratified the AS-SCORED ONLY qualifier amendment 2026-07-29 ("I ratify the
F20 amendment"), so the freeze note now records it as ratified rather than
flagged-pending. The amendment itself is unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
…e same clause F22 [should-fix]: the F21 fix invented NOT VOCAB-SPARING as the failing label. All three predecessors emit the lineage's pre-committed null "not shown" (m1_battery.py:662, m2_depth.py:857, m3_matrix.py:1124), and for good reason — failing a Wilson LOWER bound does not establish the contrary, so an assertive negative cannot be earned by it (at 1.5B, k=40 has a point estimate of 0.563 above the bar with a straddling interval, yet would have printed the negative). The wording now emits "not shown", and scopes the string: the gate verdict is the AND over the two gate-bearing subjects, while 0.5B is reported in the same shape under its standing frame and is never a gate claim — the brief had pre-committed the string "per subject" while the same block says 0.5B is never gate-bearing and the power section expects 0.5B to fail. F23 [should-fix]: HANDOFF.md never got the round-4 amendment — still "three adversarial rounds", still summarizing the pre-amendment string F20 found defective, still reading round 4 as pending, and still "D21 decides their treatment" after D21 froze. Fourth recurrence of the F10/F17/F19 propagation class; the whole M4 paragraph is rewritten and the file's Last-updated line moved to 2026-07-29. Three nice-to-haves are fixed here rather than deferred BECAUSE THEY LIVE IN THE SENTENCES F22 FORCED ME TO REWRITE — leaving known defects in a clause I was rewriting anyway would be perverse, and all three are in the frozen wording that byte-freezes into code. Flagged for Kyle rather than assumed: - F26: "the verdict, whatever it is, additionally carries the qualifier" had no precedence exclusion, so verbatim it attached AS-SCORED ONLY to DEGENERATE / UNDERPOWERED / NOT A RESULT — contradicting the D17 rule the same sentence cites. Now explicitly bar-level only, matching m3_matrix.py:1109-1125, and the D20 precedence paragraph says so too. - F25: the frozen example had dropped "survive all 12" from the template, the qualifier had an example but no template, and the both-reads-fire rendering was unspecified — against the brief's own "stated once and implemented verbatim" rule. Both templates now stated explicitly, with a fixed read order. - F24: the "k of 44 or 45" window quoted only k=44's numbers. Both are now worked (44 -> 0.503 vs 42/71 0.475; 45 -> 0.518 vs 43/71 0.489). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
F27 [should-fix]: the F22 label swap reached every use of NOT VOCAB-SPARING except the amendment note that documents it, leaving the retired label as the only surviving assertion of itself in the repo — inside the paragraph the DECISIONS.md D20 entry gets built from. Fifth recurrence of the F10/F17/F19/F23 partial-propagation class. F28 [should-fix]: a427b2d was a SECOND material amendment to the byte-frozen GATE_WORDING block, but the freeze note recorded only the round-4 amendment and carried only its ratification quote ("I ratify the F20 amendment"). The unratified block therefore read as covered by Kyle's ratification. It is now recorded as a separate, explicitly NOT-YET-RATIFIED Amendment 2, enumerating what changed against eef9a27: (i) the failing label became the lineage null "not shown"; (ii) 0.5B is scoped inside the wording; (iii) the qualifier attaches to claim-level verdicts only; (iv) both string templates are stated. The claim that the gate, its 0.5 bar, its arm and its precondition are unchanged is re-verified byte-for-byte against 90b994c and holds under both amendments. HANDOFF carries the same split. F29: HANDOFF's "reviewed across five rounds" was true only until this round's triage; replaced with a pointer to the mailbox, which cannot go stale. Amendment 2 is flagged for Kyle's ratification; nothing merges until he rules. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
Kyle ratified the second post-freeze amendment 2026-07-29 on the recommendation that enumerated all four of its changes: the failing label became the lineage null "not shown"; 0.5B was scoped inside the wording; the AS-SCORED ONLY qualifier attaches to claim-level verdicts only; both string templates are stated explicitly. The amendment itself is unchanged — only its status. With this, both amendments to the frozen D20 wording package are ratified, and the gate, its 0.5 bar, its arm and its re-certification precondition stand exactly as frozen at 90b994c. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015YkaGZYHwRoGa4e6AsLNLL
Owner
Author
Adversarial review — CLEAR TO MERGESix reviewer rounds over Rounds 4, 5 and 6 were Kyle-authorized overages beyond the three-dispatch cap, each granted because the preceding round had found a defect inside the previous round's fix. Disposition
Waivers, verbatim
— on a recommendation that (1) ratified Also recorded: "I choose option b. I authorize a fourth reviewer dispatch", "I agree with your recommendations. (a) across the board", "I ratify the F20 amendment and approve a fifth round", "one more round like you mentioned". Decisions frozen in this PRD19–D22, (a) across the board (Kyle, 2026-07-29). Two post-freeze, pre-run amendments to the D20 wording package, both ratified; the gate, its 0.5 bar, its arm and its re-certification precondition are unchanged from the freeze at The pattern worth carrying into the code PREvery defect after round 1 was the same shape — a rule stated in prose that breaks when read literally by a runner: F11's selector matched everything, F16's matched nothing, F20's string carried the number the brief calls dishonest, F22's label asserted what a lower bound can't establish. Follow-ups carried: none. Every finding is verified, waived, or closed. |
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
1 participant
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.
What this is
The start-of-stage brief for M4 — the vocabulary collateral strip, the close-out stage Kyle picked on 2026-07-28 (post-M3 decision session) over closing immediately and over the S1/S2 stretches (declined and banked as idea #13 in the j-lens-proj-ideas backlog). Docs only — no runner code until Kyle freezes D19–D22.
What M4 is
M3's Honest limits states the bound: the matrix measures collateral among 12 concepts, not across the vocabulary. M4 closes it — the 12 characterized primes run against all 180 M1 items; the new content is the never-measured non-subset pool (48 concepts' gated items × 12 deletions: 492 / 852 / 1,008 cells per subject).
Decisions proposed for freeze
Both inherited obligations dispositioned explicitly:
oracle._BOUNDARYnot triggered (no non-ASCII vocabulary), and the conjunction-degeneracy carry-forward discharged by a deliberately single-clause gate whose one surviving arm is named in the wording.All ns realized from recorded artifacts with the frozen oracle (gate arm 41 / 71 / 84; worked bar at 1.5B: ≥ 44 of 71, computed with the frozen ruler). HANDOFF updated to the new state.
🤖 Generated with Claude Code
https://claude.ai/code/session_01ANtkrPFv6i9CgCasojqcoZ