Repository path: workshop/experiments/E-20260725-tierD-ladder/critic.md · rendered 2026-09-09

Independent pre-run critic pass — Tier D perturbation ladder

Experiment discipline step 2. An independent adversarial critic (a separate agent instantiation, not the orchestrator that drafted the design) pressure-tested design.md v1 and the on-disk instrument before any API call. Verdict: NEEDS-REDESIGN, on twenty-one blockers.

This is the third design this project has put through an independent critic pass and the third to come back NEEDS-REDESIGN. Two of the blockers (B10, B11) are regressions: defects that a previous critic pass raised, that were accepted, and that this design reintroduced. That pattern is the most useful thing in the critique and is recorded as such at the bottom.

The critique measured the files rather than reasoning about them, and it is right about essentially everything it measured. Its single most consequential finding is D21: the sham arm — the control that makes every positive result mean anything — was measurably archaizing, and on the design's own §7.3 would probably have voided the entire run after the money was spent. Its second is E26: the pre-registered wave-2 gate arithmetically forbade wave 2 from ever running, in an experiment titled "the perturbation ladder".

design.md was rewritten as v2. A run may proceed only against v2.

Critique (verbatim)

Independent pre-run critic pass — E-20260725-tierD-ladder

I read design.md (v1, mtime 07:25), the charter §§2–8, wiki/goodness-senses.md, config/models.md, config/budget.md, wiki/program.md Slate A, T-berman-tendances-R04-v1, all three anchor directories, both prior critic passes, and the S014 raw run data. I also read the un-reviewed instrument that already exists on disk: materials/perturbations.md, materials/items.json, materials/build.py, materials/*.txt, tools/run_tierD.py, tools/verify_tierD.py.

The materials moved under me during the review — perturbations.md, items.json, build.py and every variant .txt were regenerated at 07:38:05, after I had measured them at 07:34. All numbers below are as of 07:38:05. That fact is itself finding B10.

Everything numbered below marked "measured" was computed from the files, not inferred.

---

A. Materials

A1. Passage word counts in §3 are both wrong. LESSER. Measured: Shaw §I is 314 whitespace tokens, not 318. The lead's Genji §1‑3 is 325 words, not 380 (§1‑3 + §1‑4 together are 600). The materials are correctly cut — A-source-ja.txt (805 chars) is exactly 蜘蛛の糸 §一 and ends at 「まっすぐにそれを御下しなさいました。」 matching Shaw's §I ending; B-source-classical-ja.txt (494 chars) is exactly 蓬生 §1‑3 and matches B-ref-lead.txt. So this is a documentation defect, not a build defect — but it is the second time in three sessions that a design's stated file facts did not survive checking, and 380 vs 325 is a 17% error in the number that justifies passage comparability.

A2. The Beowulf line range in §3 is wrong. LESSER. The design says "ll. 2015–2040". Every stored file — beowulf-ingeld-oe.txt, gummere-1909.txt, kirtlan-1913.txt, the lead's T-beowulf-ingeld-R04-v1 — is ll. 2015–2070a (56 lines), and that is what C-source-oe.txt (369 words) contains. Nothing is trimmed to 2040. Fix the design text; do not trim the files (a 2040 cut falls inside Kirtlan's chapter break XXIX and mid-sentence).

A3. §3's justification for choosing Genji as the lead-provenance reference is false. LESSER. "Of the five lead translations available, the Genji passage is the only literary narrative prose into English whose source is stored beside it." T-beowulf-ingeld-R04-v1 is literary narrative, rendered as English prose, with its source stored in the same anchor directory — and this design uses that source in arm C. A-garnett-vanka also supplies a stored published RU→EN reference that would have bought a third language pair for free. The choice of B may still be right; the stated reason is not a reason.

A4. Provenance headers, footnote markers and chapter headings are correctly stripped. FINE — say so. Measured on the built files: C-gummere.txt contains zero {28a}–{28e} markers, C-kirtlan.txt contains zero [55]–[59] markers and no XXIX heading, A-ref-shaw.txt carries none of the === PROVENANCE & LICENSE === block that names Glenn W. Shaw, and no file carries a title line. This was the single largest available authorship leak and the build handles it. It is also the one thing the design document never mentions, which is why I checked it rather than assuming it.

A5. Kirtlan's stored text is missing the rendering of half-line 2015a, and Gummere's is not. BLOCKER for arm C‑GK. beowulf-ingeld-oe.txt line 2015 opens Weorod wæs on wynne. Gummere renders it ("The liegemen were lusty"); the lead renders it ("The company was in joy"); kirtlan-1913.txt's own header records that the clause "sits at the end of the preceding sentence and is quoted in the anchor page rather than here" — so C-kirtlan.txt opens at 2015b ("Nor ever have I seen greater joy…"). A juror scoring accuracy against the source sees the first clause of the source untranslated in one text and translated in the other. That is an extraction artefact of the anchor page presented to the jury as a translation difference.

---

B. Design and statistics

B6. Detection threshold 0.75 on n=12 has no stated statistical basis and cannot survive multiplicity. BLOCKER. 0.75 of 12 votes = 9/12. Exact one-sided binomial under p=0.5: P(X≥9) = 0.0730 — not significant at 0.05 even for a single cell. The firing rule is evaluated on six cells (4 operators heavy + 2 light). Under a global null with independent votes, P(at least one cell fires by chance) = 0.365. The design states no null, no CI, no correction, and no significance test anywhere. Compare 10/12 (0.833): P = 0.0193. The threshold is one vote away from a rate that would at least be nominally defensible, and the design's own §12 says "margins of ≤ 0.17 (2 votes) are reported as indistinguishable" — i.e. it already concedes the resolution is coarser than the threshold it set.

B7. The 12 votes are not 12 independent observations, and the design knows this and pools anyway. BLOCKER. The calibration-v1 critic raised exactly this (A4, "pseudo-replication from the two orderings"), it was accepted, and the S014 analyzer implements it — analysis.json reports n_units: 15, i.e. orderings averaged within juror→span before pooling. This design reverts to raw pooling of 12 (juror × ordering × passage) and verify_tierD.py's agg() divides by len(votes). The independent units here are 6 (juror × passage). At n=6, P(X≥5)=0.109 and P(X≥4)=0.344. Either average orderings first and state a rule on 6 units (e.g. "reference preferred in both orderings for ≥5 of 6 juror×passage units"), or justify the pooling. This is a process regression on an already-dispositioned point.

B8. The specificity threshold of 0.75 scale points has zero empirical basis, and the project has never run this scoring format. BLOCKER. Every prior jury run in this repo is winner-per-sense, not numeric. I checked the S014 raw responses: {"accuracy":{"winner":"1","confidence":"high","rationale":…}}. §10 calls S014 "an equivalent per-sense scoring task"; it is a different task. So the project has no data on how P1/P2/P5 use a 1–7 integer scale — no mean, no SD, no evidence they use more than 2–3 of the 7 points. A 0.75-point differential could be enormous or trivially attainable, and nothing in the design says which. This is the headline metric per charter §5.4. A 2-item, 2-ordering, 3-juror pilot costs ≈ $0.12 and would fix it. Not running it is the largest un-derisked element in the design.

B9. Specificity's baseline includes naturalness, so the design's own naturalness prediction mechanically inflates the margin it is supposed to be independent of. BLOCKER. verify_tierD.py:specificity() computes drop[target] − mean(drop[s] for s ≠ target), and the "others" set includes naturalness. The design predicts (§8.6) that naturalness rises for O2 and O3 — i.e. drop(naturalness) < 0 — which lowers the baseline and inflates specificity for exactly the two operators where the design most wants a specificity result. The probe and the metric are not independent. Report specificity both with and without naturalness in the baseline, and pre-register which one fires the rule.

B10. The instrument was authored after the design, is unfrozen, and changed twice during this critic pass. BLOCKER (repeat of a dispositioned defect). The anchor-verification critic's E6 — "the claims file, the actual instrument, is authored after the critic pass and is never frozen" — was accepted and fixed there by committing claims.json before any call and recording its SHA in every run file. Here perturbations.md, items.json and build.py are untracked, were written after design.md, and were regenerated at 07:38:05 mid-review. At 07:34 the log declared B-O4-heavy site 1 as elder brother → elder sister; the built file contained elder brother, and two edits present in the file (might be cut back → should be left to grow; put off → drawn / bleakness → richness) were absent from the log. verify_tierD.py did not catch it: its check only tests that logged edits are present, never that present edits are logged. After the 07:38 rebuild the log check passes (92 edits, 0 problems), but nothing prevents the same drift recurring. Required: commit the materials in their own commit before any API call, record the items-file SHA in every run file, and add an unlogged-edit check (diff variant against reference, assert every hunk maps to a logged site).

B11. run_tierD.py reuses cached responses keyed only on (item, order, juror). LESSER→BLOCKER given B10. if os.path.exists(stem + ".json"): spent += …; continue. There is no payload hash. Given that items.json has already been rebuilt twice today, a re-run after a rebuild will silently score old responses against new texts. This is the anchor-verification critic's E7, accepted and fixed there, reintroduced here.

B12. parse_scores rejects 6.0 and "6". LESSER, cheap. It requires isinstance(v, int). A juror emitting 6.5, 6.0 or "6" burns a retry at full cost, and §9 voids the run at >10% failures. Coerce integral floats and numeric strings before rejecting.

---

C. Confounds

C13. Length is genuinely well controlled — say so. FINE. Measured word counts: A variants 313–326 against a 314-word reference; B variants 316–322 against 325. No variant is identifiable by length. Given that O2 instantiates Berman's allongement, this was the obvious failure mode and it did not happen.

C14. O2 is identifiable without reading, by counting full stops. LESSER (intrinsic, but must be reported). Measured sentence counts: A reference 14 → A‑O2‑heavy 18 (+29%); B reference 10 → B‑O2‑heavy 18 (+80%). B‑O2 also drops all 3 semicolons and the single em-dash to zero. This is what the operator is, so it is not an artefact — but it means a pooled "worse" verdict on O2 is compatible with a jury that never read either text, and the specificity claim for style-correspondence cannot distinguish reading from counting.

C15. A‑O2‑heavy is the only A variant with 9 paragraphs; every other A text has 10. LESSER. Site 1 merges "It was morning in Paradise." into the preceding paragraph. A per-item formatting fingerprint. Restore the paragraph break or apply the same merge to the reference.

C16. O3 injects accuracy damage while the design claims "Meaning preserved". BLOCKER. Measured edits in the built O3 files: - the golden pistils and stamens in their centers → the gold-colored centers — an omission (pistils, stamens). - as through a sterioptiscope → as if through a window — a different image. - In the eighth month of a year → In September one year — the eighth lunar month is not September, and the lead's own frozen log records rejecting "typhoon" for 野分 as inaccurate. - the herd-boys → the local kids — a different referent (総角 is a class-marked term for herd-boys).

O3 and O4 are therefore not orthogonal, and the O3 cell cannot support a claim about voice specifically.

C17. A-O3-heavy swaps 1930 register for contemporary colloquial: "No, no, as small as this thing is, it, too, has a soul" → "Hold on now, small as it is, it's got a soul too". BLOCKER, and it interacts with C19. This is Berman's tendency 4 (vulgarisation), not 10/11. Dropped into Shaw's elevated-archaic 1930 prose, it is the sharpest possible register cue — and §6 predicts naturalness will hold flat or rise here.

C18. O3 also removes all transliterated Japanese from the A variant. LESSER (intrinsic). the Sanzu-no-Kawa and Hari-no-Yama → the river of the dead and the hill of needles. Presence/absence of romaji is a one-glance discriminator. Unavoidable given the operator, but it should be named as a limitation rather than left for a reader to find.

---

D. Controls

D19. The naturalness probe is run without naming a register, in direct contradiction of the typology it cites. BLOCKER. wiki/goodness-senses.md states as a rule that "an evaluation must say period-idiomatic or contemporary-vernacular, not 'naturalness in general'". run_tierD.py's SENSE_DEFS["naturalness"] says only "Reads as fluent, idiomatic English prose". S014 already measured what happens: naturalness was the only sense to reach 0.80 and config/models.md records it as register-cued; the calibration-v1 critic's C1 was accepted on exactly this ground. The design calls this probe "the sharpest discrimination the design contains" and then runs it on the one sense with a documented register artefact, with the register unspecified, against a 1930 reference and a variant containing "Hold on now… it's got".

D20. The naturalness prediction has no threshold and no consequence. BLOCKER (third occurrence of a dispositioned defect). §8.6 says "naturalness does not fall for O2 and O3" — what counts as falling, and what happens if it does, is nowhere stated. Calibration-v1 critic B3 ("cannot fail in any defined way") and anchor-verification critic E4 ("not a prediction, measured while deciding nothing — reintroducing it is a process regression") both flagged this shape and both were accepted. Give it an integer rule or drop it.

D21. The sham is measurably directional, not quality-neutral. BLOCKER — and it is a global veto. Every A‑sham and B‑sham edit, measured by diff:

A‑sham	B‑sham
on → upon (×2 sites)	no one → nobody
all → every one	over → on
Presently → Before long	rarely → seldom
which → that	so much as → once
noticed → caught sight of	east and west gates → east and **the** west gates
looked → glanced	around → about
took up X in his hand → took X up in his hand	so much as → even
	nobody → no one

Four of eight A‑sham sites push archaic-formal (upon twice, every one) or clunkier (took the spider's thread up in his hand); one shifts sense (glanced ≠ looked). Three of eight B‑sham sites push archaic or clunkier (about the grounds, the east and the west gates, not even **on** some small matter after "called on her"), and one replaces a phrase the lead's frozen log records as a deliberate register choice (never so much as occurred to him). The asymmetry is structural, not incidental: the reference is an unedited, considered text and any eight substitutions drift downward. My prediction is that the sham lands above 0.70 for the reference and voids the entire run under §7.3 — after the money is spent.

D22. Even a perfectly neutral sham voids the run ~15% of the time. BLOCKER. The band [0.30, 0.70] on n=12 admits counts 4–8. Under a true null with independent votes, P(inside band) = 0.854, so P(spurious global void) = 0.146 — before adding juror clustering and the slot bias S010 measured at 0.63–0.85. A single control arm of 2 items is carrying a veto over 14 items. Either widen the band (counts 3–9), state it as an equivalence test with a stated tolerance, or give the sham more reps than any single targeted cell.

D23. The held-out arm compares 1909 alliterative verse against modern prose and calls it a same-quality control. BLOCKER. Measured on the built files:

	words	non-empty lines	paragraphs	archaic tokens
C-gummere.txt	403	56 (verse half-lines)	1	5 (thou, thee, thy, canst, ere)
C-kirtlan.txt	489	2	2	17 (goeth, boasteth, carrieth, dieth, lieth, walketh, cometh, escapeth, exhorteth, bringeth, shouldst, whilst…)
C-lead.txt	504	7	7	0

C‑GK is verse vs prose with a 3.4× archaism gap. C‑GL is 1909 verse vs 2026 plain prose with a 5-vs-0 archaism gap — and the lead's own frozen log states the plainness was chosen deliberately to be the plain pole of the study's scale (log §3a). §5 claims C‑GL "probes whether the jury systematically prefers published over lead provenance": no probe is needed, the two texts are separated by a century of English at first glance. Charter §5.3 and Slate A item 4 both specify the held-out arm as "an independent same-quality translation, chance expected"; the design silently relaxes this to "the margin should be markedly smaller", which is (a) not a pre-registered rule, (b) unlikely to hold, and (c) a change to a mandatory control that under charter §2.8/§8 should be opened as a decision page, not asserted inside the design that benefits from it. This arm costs 4 payloads ≈ $0.17 and, as built, can only produce an uninterpretable number.

D24. The dose ladder is confounded: light is a nested subset that retains the most salient edit. BLOCKER for the ladder's interpretation. §4 says only "Light = 3 edit sites. Heavy = 8 edit sites" — nothing about nesting or selection. Measured, every light set is a subset of its heavy set, and the subsets keep the loudest edits. A-O4-light retains it, too, has a soul → it has no soul, producing "as small as this thing is, it has no soul: it would be rather a shame to recklessly kill it" — an internal self-contradiction detectable with the source closed, present at both doses. B-O4-light retains Not so much as a servant stayed on. → A few servants stayed on.. Predict: no dose effect for O4, for a reason that is a property of the site selection, not of the jury. State the selection rule and either match salience across doses or draw the light set at random from the heavy set.

D25. O3's declared target contradicts the project's own frozen mapping, and the co-damaged sense is not scored. BLOCKER. T-berman-tendances-R04-v1's mapping table — frozen before this design and cited by it as authoritative — assigns tendencies 10 and 11 to voice, cultural-mediation. The design declares O3's target as voice alone and scores a five-sense set that excludes cultural-mediation. O3's most visible effect (removing Sanzu-no-Kawa, Hari-no-Yama, 禅師, 陸奥紙, 総角) is squarely cultural-mediation. Either score six senses or re-target O3 — as written the specificity test for O3 is measuring the wrong sense by the design's own catalogue.

---

E. Budget and operations

E26. The pre-registered wave-2 gate arithmetically forbids wave 2 from running, on the design's own point estimate. BLOCKER. §10: headroom $1.600774; wave 1 estimate $0.995; wave 2 estimate $0.332; gate = "wave 2 runs iff (headroom − wave 1 actual) ≥ 2 × wave 2 estimate" = $0.664. Maximum permissible wave-1 actual = 1.600774 − 0.664 = $0.936774 < $0.995. Wave 2 can run only if wave 1 comes in ≥6% under estimate — against a project record of S010 (est $1.9–3.0 → $4.49), S014 jury (est $0.90 → $1.83), S014 probe (est $0.25 → $0.40). The experiment is titled "the perturbation ladder" and its own budget rule makes the ladder unrunnable. Either restructure (run light before or interleaved with heavy), drop the held-out arm to buy headroom, or defer the dose axis explicitly rather than by arithmetic accident.

E27. The $0.04146 per-payload assumption is a mean transplanted from a smaller, different task. BLOCKER. The figures are exactly the S014 per-juror means (verified: P1 $0.01524, P2 $0.01609, P5 $0.01013). But (i) S014's task was winner-per-sense, this one is 10 integer scores + preference (B8); (ii) payloads are larger — measured source+two-translations: item A 4,307 chars, item B 3,970, item C 7,734–7,794, against S014's ≈2,000 with mean prompt_tokens 1,224; (iii) using S014's per-call maxima instead of means ($0.01738 + $0.02365 + $0.02364 = $0.06467), wave 1 alone = $1.552 — above the $1.45 abort and just under the $1.601 headroom, and all 32 payloads = $2.069. (iv) P5's measured mean cost $0.01013 is ~3× what config/models.md's list prices predict from its token counts, so there is unmodeled billing in the one juror the estimate leans on for cheapness. Restate the estimate as a range with a worst case, as E-20260725-anchor-verification did (the only run so far to land inside its estimate).

E28. Dispatch order puts the global gate last, so a budget abort destroys the run. BLOCKER, one-line fix. run_tierD.py builds payloads = [(it, o, j) for it in items …] in items.json order, which for wave 1 is: 8 heavy targeted → A-sham, B-sham → C‑GK, C‑GL. The pre-dispatch guard if spent + max(worst, 0.05) > cap: break therefore drops the sham and held-out arms first. Per §7.3, without sham data no sense may be declared calibrated whatever the other arms show — so an abort at 80% of budget yields a run worth exactly nothing. Reorder: sham → heavy → held-out. Better still, run the sham arm as a standalone $0.17 gate before committing to the rest (this also derisks D21 for two-thirds of a cent per call).

E29. The abort cap ($1.45) sits above the point at which the run can complete, and below headroom. LESSER. With headroom $1.601 and a plausible wave-1 actual of $1.2–1.55, the cap will fire mid-wave-1 rather than cleanly between waves. Either set the cap at a value that guarantees a complete wave (and shrink the wave to fit) or make the guard arm-aware so it never truncates an arm.

E30. --max-tokens 9000 is undocumented in the design and is the S010 failure surface. LESSER. S014 used 10,000 and P5 averaged 3,079 completion tokens on a smaller payload. Truncation → parse failure → retry → double billing, and §9 voids the run at >10% failures. State the value in the design, state the retry budget, and state what happens to a truncated call's cost in the ledger.

---

F. Charter compliance

F31. Relaxing a charter-mandated control without a decision page. BLOCKER. Charter §5 step 3 and Slate A item 4 both make the held-out arm "an independent same-quality translation, chance expected". §5 of the design overrides this ("Not 'chance expected'") inside the experiment page. Charter §2.8: "Value-laden choices are opened as decision pages and ratified cross-session… A session never ratifies a decision it opened." Weakening the gate experiment's mandatory control is exactly such a choice.

F32. The config/models.md revisit trigger is declared discharged in the wrong file, and on the wrong evidence. BLOCKER. config/models.md currently reads "FIRING as of 2026-07-25 … Re-probe with tools/panel_probe.py before the next jury-heavy run — cheap, and Tier D depends on the instrument being current." The design declares the trigger's premise false and discharges it in §6 of an experiment page, leaving the config page saying FIRING; CLAUDE.md rule 6 makes config/models.md the sole configuration locus. Separately, the substitute evidence offered — that four of five were "exercised at length" in S015 — is a factual-adjudication task that config/models.md itself records as one that "does not bear on calibration". Either edit config/models.md with the reasoning and the evidence, or run the ~$0.05 probe.

F33. The order-flip failure threshold is set just above the one juror already known to be unstable. BLOCKER. §9 down-weights any juror with flip rate > 0.50. Measured in analysis.json: P1 0.194, P2 0.167, P3 0.167, P4 0.139, P5 0.444 — and config/models.md records P5 as "down-weight until reps increase". P5 is one of three jurors here, i.e. a third of every cell, and the pre-registered rule cannot fire on it. Also, with 16 items each juror yields at most 16 order pairs, so the flip-rate estimate itself has a 95% interval roughly ±0.25 — a hard threshold on it is not measurable at this n. Fix: pre-commit a P5-excluded sensitivity analysis (the anchor-verification design's accepted disposition for the identical objection, C2), not a threshold P5 cannot cross.

F34. Charter §5.4 makes cross-sense specificity the headline metric, and the design predicts its own headline will fail with no fallback decision. LESSER but worth stating. Prediction 3 says specificity fails on ≥2 of 4 senses, and S014's floor contrast already showed the "worse across the board" pattern (record-fit 0.98 / 1.00 / 0.87 / 0.77 / 0.80 / 0.92 on a blatantly worse translation). §7 names the outcome ("detects damage, not sense-calibrated") but never says what Tier D's gate status becomes if every sense lands there — which is the whole point of running it. Pre-commit that.

F35. Compliant where I checked. FINE. Order swap mandatory and implemented; dispatch strictly sequential (§6, charter §5); raw JSON preserved including .attemptN.json discards (the S015 fix); pre-dispatch worst-case cost guard (the S015 fix, anchor-verification G4); slugs resolved from config/models.md at run time, tools referring to roles P1/P2/P5; Berman read in French and correctly ledgered in wiki/base/consulted.md; freeze ordering verified (commit 0111d38 predates design.md, whose mtime is 07:25); the "none of the twelve tendances targets accuracy" finding is real and correctly used to license O4 under Slate A item 2; the kimi exclusion figures are accurate (measured P4 mean $0.09089, i.e. 5.96× P1 and 8.97× P5); no private-texts/ content is reproduced in the public materials; the absence of a senses: field on a page that evaluates the jury rather than a translation is defensible.

---

The single most important weakness

Three of the four things this run is meant to buy are unbuyable as designed. Detection is nearly a foregone conclusion — S014's floor contrast already showed 0.77–1.00 record-fit against a genuinely worse translation — so the detection leg mostly re-purchases something the repo has. Specificity, the charter's headline, is measured on a 1–7 scale that has never been piloted with these models, against a 0.75-point threshold with no empirical grounding, using a baseline that the design's own naturalness prediction mechanically inflates. The dose ladder cannot run, because the pre-registered wave-2 gate ($0.937 permissible against a $0.995 estimate) forbids it. And the one arm that would make any positive result mean something — the sham — is a measurably archaizing, clunkier edit set that is more likely to fail its own band than to sit in it, carries a ~15% spurious-void rate even if perfect, and is dispatched after everything it gates.

The cheap path forward, in order: (1) commit the materials and add an unlogged-edit check; (2) run the sham arm alone, 12 calls ≈ $0.17, as a gate — it simultaneously pilots the 1–7 scale (B8) and tests D21; (3) only if it lands in band, rebuild the firing rule on 6 independent units with a stated null, fix the naturalness prompt and rule, re-target or re-score O3, drop or replace the held-out arm, and re-plan the budget so the ladder can actually run.

VERDICT: NEEDS-REDESIGN

Disposition of every point

design.md v2 implements the critique's own recommended sequence. Twenty-one blockers and eight lesser points; 26 implemented, 3 accepted-and-recorded-as-limitations, 0 rejected.

The pattern worth keeping

Three of the blockers — B7, B10, B20/D20 — are defects that an earlier critic pass in this project raised, that were accepted, and that this design reintroduced. The project's critic passes are working; its memory of them is not. Neither workshop/experiments/README.md nor continue-prompt.md carries forward the specific dispositions that previous passes established, so each new design re-derives them from scratch and sometimes gets them wrong. That is a process defect, and the fix is a standing checklist rather than a better designer. It goes to NEXT.md.