Reviewer lens — adversarial pass and the 39% verdict

⚠️ EVIDENCE — figures here predate the authority. "The 39% verdict" is canonically 35% (12/34); "five of seven axes carry no cost signal" is canonically two never-negative, six with ≤3 negatives. Authority: foundational-values-and-creed.md §11.

Squad report, 2026-07-26. Profile-loaded (reviewer-renata), read-only. Evidence base for creed-and-values-design-hardened.md. Tactics applied: reverse-speccing (the load-bearing one — reconstruct the intended semantics of the ammerse field from the corpus alone, then compare against the design's stated invariants), code-review-incremental; DIRECTIVE_041 generalized from tests to schema slots (a gate that passes for the wrong reason), DIRECTIVE_030, DIRECTIVE_024, DIRECTIVE_032. All corpus files fetched read-only; all statistics recomputed independently.

Findings

[CRITICAL] :113 — "Every number is paired 1:1 with its rationale sentence — co-location is already the invariant." Falsified by the corpus. Of the 14 files fetched, the_point_of_dissent.md carries 7 deltas and zero rationale keys; quad_A_test_structure.md carries 7 deltas and zero rationale keys; TEMPLATE_PRACTICE.md carries 7 empty-string rationales. Co-location is a convention with ≥3/38 violations, not an invariant. This directly undercuts §6's row "Ungrounded scores are unfalsifiable → Addressed": the corpus offered as proof the pairing works contains the exact unpaired case. → Stop citing co-location as established practice. If rationale is mandatory it is a new constraint the calibration corpus fails, needing the same advisory-vs-hard decision §5.6 demands for negatives.

[CRITICAL] :189-192 — The 39% figure is contaminated and is the wrong quantity. Two of the 14 named all-non-negative vectors are not authored artefacts: TEMPLATE_PRACTICE is the literal authoring template (all-zero, empty rationales) and quad_A_test_structure is an unfilled stub (all-zero, no rationale keys). Honest denominator 34, numerator 12 → 35% — and both stub rows sit inside the n=36 PCA of §5.3 and the means of §5.4. → Republish §5.2/§5.3/§5.4/§5.6 at n=34 with the stub-exclusion rule stated. I re-ran it: the conclusions survive (PCA cumulative 0.816 at PC4; per-axis signs unchanged), so this costs the argument nothing and removes an easy shot.

[CRITICAL] :150, :164-168 — "two axes carry no cost information" materially understates the measured result. Negative-cell census (n=36): agile 2, minimal 19, maintainable 3, environmental 0, reachable 2, solvable 0, extensible 1. Total 27 of 252 cells (10.7%), and minimal alone carries 19 of 27 (70%). Five of seven axes have ≤3 negative cells. → Restate as "one cost axis and six benefit axes." It changes what a creed weighting can mean, and it is the honest reading of the design's own data.

[CRITICAL] :68 — Part 3's retrieval degenerates to a magnitude ranking at any plausible creed. Because six of seven axes are near-monotonically positive, a weighted sum is dominated by row magnitude. Measured (34 authored vectors, Dirichlet creeds on the positive simplex, Spearman vs the unweighted row mean, top-5 overlap vs the flat-creed top-5):

Concentration	Spearman r	Same top-5
near-flat (α=50)	0.980	97%
mild tilt (α=20)	0.958	90%
moderate (α=8)	0.913	81%
opinionated (α=3)	0.825	67%
uniform (α=1)	0.671	48%
extreme (α=0.3)	0.486	34%

The flat-creed top-5 is external-memory, efficient_async_communication, define_test_boundaries, do_the_important_things, easy_to_change — exactly the five largest row-sums. external-memory (largest sum, +3.50) appears in the top-5 for 53% of all creeds. → Before building this, specify the ranking function and state the null model. If the answer is a linear dot product, gate it on "top-5 must differ from the flat-creed top-5 by ≥2 items", or it is measuring authorial enthusiasm and labelling it the operator's values.

[MAJOR] :68 — The creed score carries no query term, so it is not retrieval. It imposes one global preference order over the whole corpus regardless of what the operator is doing: the flat-creed top hit for any question, including a testing question, is external-memory. Any usable system must combine it with a topical filter and the design does not say how. Worse, the topical filter selects a small, topically-coherent candidate set whose vectors are more similar to each other than the corpus average (the design's own maintainable × extensible r = +0.41), so the creed's discriminating power falls exactly where it gets used. (Last step is reasoning, not measured.)

[MAJOR] :206 — §6's answer to the 3-for-3 decay precedent is a category error. "Part 3 (the interview) is the producer" is true for the creed (7 weights + 7 rationales = 14 cells). It is false for everything else. Measured in this repo: 260 behavioural doctrine artefacts (27 directive, 126 tactic, 21 styleguide, 14 toolguide, 14 paradigm, 22 procedure, 18 agent, 17 step-contract, 1 glossary-pack) × 7 axes × (delta + rationale) = 1,820 authored cells; 774 DRG edges across 9 populated graph files × 1 impacts value; plus a 7×7 matrix ≈ 2,650 cells. The named producer covers ~0.5% of the register it is offered as the answer for.

[MAJOR] :56-60 — Part 2 and the ValueConnascence matrix have ZERO grounding in §5. §5 is titled "Verified grounding measured against this design" and every measurement in it concerns the artefact→value delta field. Nothing in the 36-vector corpus calibrates an edge strength or an inter-value coefficient, and 774 edges is 23× the entire calibration set. §6 row 11 concedes availability of the first-order matrix but not calibration. → Split the document. Inheriting §5's credibility across the section boundary is the main rhetorical risk here.

[MAJOR] :63 — The interview's elicitation format is delegated to the model, and the design's own provenance chain says format inverts the answer. The handover records prior art in which a declared value ordering moved revealed model priorities by 0.145 normalized, "with rankings inverting between elicitation formats." Part 3 makes format a free model choice — putting the operator's value ranking under the control of the single variable prior art says inverts it. → Fix the instrument: authored question bank, constant-sum or forced pairwise comparison (never independent Likert over seven virtues), and per-weight provenance (source: operator | model-proposed | default).

[MODERATE] :32-52 — The design retires nothing. Part 2 says impacts is added "rather than the black-or-white absolutist in_tension_with" but the verb is added, and §6 row 2 explicitly preserves authored edges. Verified: in_tension_with is still 2 edges (src/doctrine/directive.graph.yaml:92, :105) and is the subject of an Accepted ADR this document does not propose superseding. Net: +5 registers, −0.

[MODERATE] :45-49 — By the calibration corpus's own scoring, adopting this design's central practice costs minimal −0.6 and agile −0.35. AMMERSE_impact_analysis's authored vector: minimal −0.6 ("introduces significant complexity and effort, conflicting with minimalism"), agile −0.35 ("time-consuming and may hinder agility… The potential indirect support for better decisions is not enough to offset this"). minimal is the design's own single live discriminator. → Put that vector in the document as the design's self-assessment. It is the most credible thing available and it is free — the corpus already authored it.

[MODERATE] Nothing in the design detects the defect the corpus actually has. A sign-vs-rationale-polarity lint — flag any cell whose rationale names only a cost while delta ≥ 0 — is the single highest-value gate here, is mechanically checkable by an LLM on the prose the design already mandates, and exists in no part. Make it Part 0. It is the only proposed component with a demonstrated defect population to validate against.

[MINOR] :110 — "0.125", "0", "1", "-0.25" all appear. The string/number coercion boundary and the "0" vs "" vs absent-key distinction need a stated rule — and per §5.6 that rule decides whether the_point_of_dissent is "complete."

[MINOR] :43, :49, :58, :65 — Four vocabularies for arguably two concepts: creed "weighted score", artefact "delta", edge "impacts", interview "normalized score" (DIRECTIVE_032). The advisory homonym warning is the precedent — resolve the vocabulary before rendering any of it into agent context.

The 39% verdict

Neither (a) nor (b). The rule is right about the content and wrong about the instrument it reads. I read the rationale prose for 12 of the 14 named files plus 3 negative-carrying controls.

1. Two of the 14 are not authored artefacts. TEMPLATE_PRACTICE is the authoring template — all-zero, seven empty-string rationales. quad_A_test_structure is an unfilled stub — all-zero, no rationale keys. Real population: 12 of 34 (35%).

2. Eleven of the remaining twelve name an explicit cost in prose. Not slogans, not costless:

practice	axis	delta	rationale (verbatim excerpt)
`manual_of_me`	minimal	+0.25	"introduces documentation and cognitive overhead… it is decidedly not minimal"
`wipe_the_board`	maintainable	0	"the daily upkeep burden limits long-term sustainability"
`fail_fast`	maintainable	+0.5	"the overhead of maintaining complex validation rules can detract from this, leading to potential long-term maintenance challenges"
`fail_fast`	reachable	0	"the overhead… can slightly detract from its effectiveness"
`fail_fast`	agile	+0.1	"may introduce some rigidity… can limit quick adaptability"
`LARS`	minimal	0	"may introduce complexity in managing emotional aspects"
`wax_on_wax_off`	agile	0	"may introduce slight rigidity"
`external-memory`	minimal	+0.75	"there is a risk of over-complicating the external memory system"
`ten_minute_tasks`	minimal	+0.25	"requires some preparation and planning to be effective"
`define_test_boundaries`	minimal	+0.2	"moderated slightly by the complexity introduced"
`impact_oriented_communication`	reachable	0	"some discomfort remains for newcomers"
`communication_channel_compression`	extensible	+0.1	"may need refinement as organisations scale"

3. Exactly one is a true slogan — the_point_of_dissent: seven bare numbers, no rationale keys, solvable = 1, four axes at 0. Un-examined by construction — and the mandatory-negative rule would not catch it either, because a rule that reads signs cannot detect an absent argument.

4. The floor-clamping hypothesis is wrong, and the truth is worse. Three negative-carrying controls show competent net arithmetic: AMMERSE_impact_analysis extensible 0 = "supports future extensibility but may initially hinder it" (a correctly netted zero); agile −0.35 = "the potential indirect support… is not enough to offset this"; avoid_gold_plating extensible −0.70, maintainable −0.25. The scale is not clamped — 27 cells go negative.

What is actually happening is sign inconsistency for equivalent prose across artefacts:

manual_of_me minimal = +0.25 for "introduces documentation and cognitive overhead… decidedly not minimal"
AMMERSE_impact_analysis minimal = −0.60 for "introduces significant complexity and effort, conflicting with minimalism"

Near-identical claims, 0.85 apart, opposite signs. And:

fail_fast maintainable = +0.50 for "overhead… leading to potential long-term maintenance challenges"
safe_to_fail maintainable = −0.15 for the milder "minor impact… does not significantly contribute"

A 0.65 swing in the opposite direction from the prose.

Therefore: the 35% is not a measurement of authorial honesty. It is a measurement of noise in the sign channel — the one field the mandatory-negative rule reads. Eleven of twelve authors did the honest analysis; the number failed to record it.

And this is the finding §5.3 stopped one step short of. The design used the real corpus to refute the earlier squad's dimensionality claim, correctly. It did not check whether the earlier squad's other measurement — a Solvable sign flipping between the same agent's two passes an hour apart — also replicates. It replicates, in the human-authored corpus. The rebuttal is selective: one finding was tested and refuted, the adjacent one was not tested and survives.

Recommended decisions:

minItems: 1 negative should be rejected as a hard schema constraint — not because 35% of the corpus is slogans (it is 3%), but because it gates on the least reliable field and would reject eleven honestly-analysed practices while passing the only real offender.
The constraint belongs on the prose: every rationale must name what adopting this makes worse. Eleven of twelve already comply; the_point_of_dissent and quad_A_test_structure fail, correctly.
Ship the sign-vs-prose-polarity lint instead, validated against the twelve rows above — the only component of this design with a falsification set available today.

Fakeable-success register

Part	The fake success	Why it passes review	Detector that must ship in the same commit
P1a `FoundationalValues` type	Pydantic round-trips; enum half-open; tests green	This is `Directive.severity` verbatim — green for 162 days proving round-trips	A test that fails when zero readers exist. Assert an importer outside `src/doctrine/` consumes it
P1b `creed` on charter	`creed.yaml` exists with 7 weights + 7 rationales; operator recognizes their words	Authored once at charter time, never revisited; nothing re-opens it	(i) A creed whose max−min weight spread is below a floor is the flat creed → warn; (ii) log every read and assert non-zero reads per mission
P1c per-artefact deltas (1,820 cells)	All values at ±0.5 — or the corpus's real tell, all values at 0 (2 of 38 files are exactly this)	A complete 7-vector on every artefact looks like coverage	Distribution gate: flag if >X% of cells fall in {0, ±0.5}; flag all-identical-magnitude vectors; the polarity lint
P1c' rationales	Seven paraphrases of the artefact's own summary, one per axis	Reads fluent; every cell populated	Assert the 7 rationales are not near-duplicates of each other or of the description
P1d ValueConnascence 7×7	Imported wholesale and never revisited	It is a constant; constants never break	If the project copy is byte-identical to upstream forever it is a citation, not a model. Require ≥1 cell to differ with a rationale, or delete it and link the URL
P2 `impacts` on 774 edges	Field defaults to `0.0`; schema green; "partial population is first-class"	`GovernanceConfig.enforcement = {}` exactly	Coverage keyed to non-default values only. `0.0` counts as absent. Ratchet the non-default count upward
P3a the interview	It runs; the operator feels heard	No ground truth for "the operator's real values", so nothing can be wrong	Test-retest: same operator, two formats, report rank correlation. Prior art predicts inversion. If it inverts, the instrument is speaking
P3b retrieval	The LLM cites `creed.yaml` in its trace	Citation reads as causation	A/B against a creed-blind arm. My measurement predicts 81–97% top-5 identity at near-flat-to-moderate creeds

The two I would rank first: P1c's polarity lint (a real defect population exists today) and P3b's creed-blind arm (it decides whether the load-bearing consumer does anything).

The interview loop

Yes, it is a closed loop, and it is the sharpest objection available against this design. The mechanism is more specific — and more damaging — than generic anchoring.

The seven axes are all virtues. No operator says solvability is unimportant. Part 3 specifies neither a constant-sum budget nor forced choice — only "a normalized score on an easy-to-reason-with scale." Independent scoring over seven virtues with no budget is the textbook acquiescence instrument, and its predictable output is a near-flat, uniformly-high creed.
Near-flat is precisely the regime where the creed is inert. At α=50, r = 0.980 with simply summing the deltas and 97% of the top-5 is the flat-creed top-5. The elicitation instrument most likely to be produced is the one that makes its own output irrelevant. That is not rhetorical — it is measured, and it is the most compact statement of the problem.
The laundering step is provenance, not generation. The model chooses the questions, the questions determine which axes become salient, the answers land in creed.yaml, and the same model class reads it as operator authority. Nothing records who supplied each weight. Downstream, "the operator ranked maintainability at 0.8" and "the model proposed 0.8 and the operator did not object" are byte-identical.
The design's own provenance chain says the format decides the answer — 0.145 normalized, rankings inverting between formats. The one variable prior art identifies as outcome-determining is the one the design leaves free.
The counterpoints do not reach this. The rationale-fields argument defends the consumption side, and I accept it — prose does survive where arithmetic does not. But this is a production-side failure. And "graceful degradation… better than not doing it" does not apply: the failure is not weaker performance, it is a false attribution that gets stronger with a better model, because a more fluent interview produces a creed the operator is more likely to endorse and less likely to have originated.

Fixes, all cheap: authored question bank (the model administers, does not design); forced-choice or constant-sum (21 pairwise, or 100 points across 7 — both make the flat creed unreachable, which per (2) is the whole difference); per-weight provenance, refusing to render the creed as operator authority when the operator count is zero; and apply the mandatory-negative rule to the creed itself — require the operator to name what they are deprioritising. That is the one place in the entire design where the rule has unambiguous teeth. Note: the earlier verdict reserved the charter Purpose rewrite for the human precisely because an agent drafting it is the self-scoring failure. Part 3 as written re-delegates to an agent the one thing that verdict reserved for the human.

What survives me

§5.3 is correct and I verified it independently. PCA cumulative variance replicates to three decimals, and holds with the stubs removed (0.339 / 0.561 / 0.712 / 0.816 / 0.900). Only two pairwise correlations exceed |0.35|. The strongest thing in the document and it is solid. My attack lands on the sign channel, not the dimensionality.
The corpus is not slogan-shaped. avoid_gold_plating extensible −0.70, AMMERSE_impact_analysis minimal −0.60 for its own method, be_a_STARR_at_interviews minimal −0.65, breaking-conditions reachable −0.50. Specific, costly, unflattering self-assessments — a genuine calibration corpus authored by someone doing the work.
minimal is a real discriminator. Mean +0.024, stdev 0.402, 19 of 34 negative, range [−0.65, +0.85]. A single field asking "what does adopting this cost in simplicity?" is cheap, defensible, and has 34 worked examples. Nothing in my attack touches it.
Co-locating a rationale with every number is right, and the corpus proves it in the way the design did not anticipate. In fail_fast and wipe_the_board the prose is correct and the number is wrong. The rationale is the surviving channel.
Charter-scoping the creed is the right home. Not contested.
The half-open, replaceable value set genuinely defuses basis lock-in. Not contested.
"Better than not doing it" is a legitimate posture — for the prose. No counter there. Only to the arithmetic.

The minimum I cannot argue away:

A costs: free-text field on behavioural doctrine artefacts — one sentence naming what adopting this makes worse — plus an operator-authored charter statement naming what the project deprioritises. No numbers. No 7-vector. No matrix. No edge weights. No weighting function. No retrieval claim.

Everything I attacked lives in the arithmetic. Nothing I attacked touches the prose. If the design were resubmitted as prose-only I would have almost nothing left — and by §5.4's own measurement it would keep the only axis that discriminates.

Concession

Sampling. 14 of 38 files plus 3 controls — the all-non-negative set and the stubs, because that is what settles the question. I did not sample the 22 negative-carrying files systematically. The sign-inconsistency finding rests on 3 controls against 12 subjects: enough to establish inconsistency exists at ±0.85 for equivalent prose, not enough to characterise its distribution. I cannot rule out that the scale is coherently shifted by a constant, in which case sign is meaningless but ordering survives intact — which strengthens "gate on prose, not sign", but I did not test it.

The retrieval model is mine, not the design's. Part 3 does not specify a ranking function. I assumed a linear dot product. If the real intent is an LLM reading creed.yaml as prose context with no arithmetic, my α-sweep does not apply — but then "look for artefacts that align" is doing no retrieval work and the 1,820 numbers are decorative. Either horn is a finding; the design should say which was meant.

Dirichlet-uniform is not a model of real operator creeds. The sweep is sensitivity analysis, not prediction. My claim that LLM interviews yield near-flat creeds is inference from the virtue-framing plus the cited prior art. It is the weakest joint in my strongest argument, and it is testable in a day.

Repo claims are static greps. No tests run. The 260-artefact and 774-edge counts are find/grep; the 774 sums relation: occurrences across nine graph files and may double-count. Directionally right; not audited.

Out of lens: the trademark/authorization question and licensing generally. Also whether a non-linear consumer could rescue Part 3.

Verdict

The design is materially stronger than its predecessor and its central empirical claim is sound: I independently replicated §5.3 and the seven axes really do carry independent information — the one-dimensionality finding is dead. On the question the input doc says cannot be resolved from the data, it can, and the answer is a third option: the fourteen all-non-negative vectors are neither costless nor slogans — two are unfilled stubs, one is a genuine slogan with no rationale at all, and eleven of the remaining twelve name explicit costs in prose while the number sits at or above zero, with manual_of_me recording minimal +0.25 for prose reading "decidedly not minimal" against AMMERSE_impact_analysis's −0.60 for the same claim. So minItems: 1 negative must not ship as a hard constraint — not because the corpus is dishonest, but because the rule gates on the sign, and the sign is the one field the corpus demonstrably cannot hold steady; it would reject eleven careful analyses and pass the only real offender. The correct constraint is on the prose, and the correct new component is the polarity lint, the only piece with a validation set available today. Against that, the numeric layer does not currently earn its cost: five of seven axes carry ≤3 negative cells in 34 vectors, so a creed weighting is one cost axis against six benefit axes, which makes the Part 3 ranking collapse into a magnitude ranking — measured at r = 0.98 and 97% top-5 identity with ignoring the creed entirely at a near-flat creed, and 81% even at a moderate tilt — and the interview as specified is the instrument most likely to produce exactly that near-flat creed while giving a model's prior operator provenance in a file no consumer can audit. Part 2 and the matrix carry none of §5's grounding, §6's "the interview is the producer" answers 0.5% of a ~2,650-cell register, and the design retires nothing: net +5 registers, −0. Recommendation: split it. Ship the prose layer — a one-sentence costs: field plus an operator-authored deprioritisation statement plus the polarity lint — which is cheap, has 34 worked examples, and survives everything above. Hold the numeric layer, the edge impacts, and the matrix behind two gates: a creed-blind A/B on the retrieval claim, and a test-retest on the interview.