docs(paper): plain-English 1:1 rewrite of the paper (eli5) - #16
Merged
Conversation
Mirrors docs/paper/mute-map-paper.md one-for-one: same 30 headings in the same order, same 68 carried paragraphs in the same order, nothing merged, dropped, added or reordered. Only the language changes. Verified mechanically before commit: - headings identical and in order (30/30) - per-section paragraph counts match (68 carried + 1 authored header block) - all 91 table rows byte-identical to the source; every table followed by an in-plain-words gloss (12/12) - all 6 figures carried as the real images, same relative paths, captions rewritten (6/6) - 4 lists carried item-for-item (5/5 and 5/5 items in 2.4 and 6.3) - section 8 References carried verbatim, byte-for-byte - no display equations in the source, so no named-form blocks; 8 inline math spans normalized to canonical $...$; the three literal "$0" cost figures escaped as \$0 so they cannot be parsed as math delimiters downstream No numbers, verdicts or claims were altered. No measurement was run. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MwGUzVvUmaiEShFiZfrcW7
… boundaries, and verdict scope in three glosses Adversarial review round 1, findings F1-F3 (all should-fix, all accepted). F1: section 2.3 described the widened oracle as "the model's three most likely word-pieces" -- a top-3 candidate list at one position. The frozen rule in oracle.py is the decoded first-3-greedy span: three pieces produced one after another, most likely at each step. Different objects, and the old wording could not support the "decoded span, stripped of leading whitespace" rule attached to it. Reworded to the greedy continuation. F2: three markdown block boundaries lost their blank line at the file-assembly seams. At line 750 a "---" directly followed paragraph text, which CommonMark parses as a setext heading underline rather than a thematic break -- rendering that whole closing paragraph of 4.5.3 as an H2 on GitHub and injecting it into the auto-TOC. The other two (before "### 4.4" and "## 4. Results") rendered correctly but are the same class. All three fixed; a scan now reports zero. F3: three in-plain-words glosses closed with "all three come out BREADTH-SPECIFIC / LATE-LOCALIZED / MATRIX-SPECIFIC", dropping the off-gate scope. The paper states that 0.5B is never gate-bearing and scopes every verdict to 1.5B AND 3B. Rescoped to the two verdict-carrying models, with 0.5B named as off-gate. Structure re-verified after the fixes: 30/30 headings, 68 carried paragraphs, 91/91 table rows still byte-identical, references still verbatim. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MwGUzVvUmaiEShFiZfrcW7
Owner
Author
Adversarial review — CLEAR TO MERGE (rounds: 2)
Follow-ups (nice-to-have): F4 "barely dents" overstates at 0.5B; F5 country exemplar in §4.5.1; F6 leak-guard described as a scoring rule; F7 dropped correlation condition. All four are in gloss text authored by this PR, all four are cheap one-line edits, and all four are surfaced to Kyle for a pre-publication decision rather than auto-fixed — this file is the input to Coverage: complete (both rounds — the new file in full, the source paper in full, and Round 1: 0 critical, 3 should-fix, 4 nice-to-have. Round 2: 3 verified, 0 reopened, 0 new. Structural fidelity independently confirmed by the reviewer: headings identical in text and order, all 91 table lines byte-identical between the two files, per-section paragraph counts matching once the declared glosses are accounted for. Zero-context reviewer + neutral judge on disputes (no disputes arose); anchored at |
All four are in in-plain-words gloss text authored by this PR, graded nice-to-have by the reviewer and fixed at Kyle's direction rather than deferred, since this file is the input to /paper-gloss and is about to be published as an Artifact. F4: the 4.3 gloss called early/middle removal 'barely dents naming'. True at 1.5B and 3B, false at 0.5B, where 17/28 against a clean 28/28 is a loss of 11 items -- 39%. It also pre-empted the raised-floor discussion two paragraphs later. Now scoped: leaves most naming intact at 1.5B and 3B, costs 0.5B about four items in ten. F5: the 4.5.1 gloss illustrated category-block collateral with 'deleting a country's direction knocks other countries about' -- the exact countries-only framing the next paragraph identifies as an artifact of M3's arm composition. Rewritten to same-category damage across ten categories. F6: the 6.2 gloss described M1's clue-leak guard as a scoring rule about which answers count as hits. It is a load-time validity check on the item set (m1_battery.py:246-256): no word of a clue may start with the concept or control. Restated as a check on clue sentences. F7: the 6.1 gloss asserted the Newcombe-widening argument unconditionally. The source scopes it to positively correlated paired arms, which is what makes the independent-samples interval the conservative one. Condition restored. Structure re-verified: 30/30 headings, 68 carried paragraphs, 91/91 table rows byte-identical, references verbatim. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MwGUzVvUmaiEShFiZfrcW7
… forbidden-forms half (F8) Review round 3, raised on the line round 1's F6 fix rewrote. Nice-to-have; fixed rather than deferred for the same reason as F4-F7 -- this file is about to be published as an Artifact. The gloss called the guard 'a strict word-boundary test'. m1_battery.py:254-265 is a word-*prefix* test: each word of the clue is checked with w.startswith(b) against the concept, the control, and the concept's frozen forbidden_forms. A both-ends boundary test would pass 'ants' for 'ant'; the prefix test catches it. The gloss also explained only the loosening half (the substring test wrongly flagged 'plant' for 'ant') and dropped the forbidden_forms half, which is what the source row means by 'the stricter guard'. A reader was left thinking M1 relaxed the check when it tightened it in one direction while loosening it in the other. Both halves now stated. Verified against the source deviations row (carried verbatim at line 645 of the paper) and m1_battery.py:254-265. The reviewer dispatch cap for this run was spent on round 3, so this fix is author-verified against the code rather than by an independent round -- stated plainly rather than implied. Structure re-verified: 30/30 headings, 68 carried paragraphs, 91/91 table rows byte-identical, references verbatim. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MwGUzVvUmaiEShFiZfrcW7
Owner
Author
Adversarial review — final: CLEAR TO MERGE (rounds: 3)Round 3 verified F4–F7 and raised one new nice-to-have, F8, on the line F6's fix had rewritten.
All five nice-to-haves were fixed rather than deferred. Each is an accuracy error in gloss text this PR authored, each is a one-line edit, and this file is the direct input to One caveat stated plainly: F8 landed after the third and final reviewer dispatch, so it has no independent verification. It was checked by the author against Coverage: complete on all three rounds. Structure re-verified after every commit: 30/30 headings, 68 carried paragraphs, 91/91 table rows byte-identical to the source, references verbatim. |
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.
Adds
docs/paper/mute-map-paper-eli5.md— a plain-English rewrite ofdocs/paper/mute-map-paper.md, produced by the/paper-eli5skill.It is a translation, not a summary: same 30 headings in the same order, same 68 carried paragraphs in the same order, nothing merged, dropped, added or reordered. Every number, verdict and claim is stated exactly as the original states it. Jargon is defined inline on first use (Jacobian lens, projection removal, token, oracle, Wilson/Newcombe intervals, prime/probe, diagonal, gated n, UNDERPOWERED/DEGENERATE/NOT A RESULT).
Verified mechanically before commit
docs/paper/), captions rewritten$…$; the three literal$0compute-cost figures escaped as\$0so a downstream math parser cannot pair them as delimitersNot done here
No measurement was run. No runner, gate, ablation or anchor script was executed. Nothing in
results/,items/or any behavioral file is touched — this PR adds exactly one new markdown file.Next step:
/paper-glosson this file, to produce the interactive glossed HTML artifact.🤖 Generated with Claude Code
https://claude.ai/code/session_01MwGUzVvUmaiEShFiZfrcW7