Creed and FoundationalValues — hardened design

⚠️ RECORD — superseded by foundational-values-and-creed.md. Known-stale figures a direct reader must not carry forward: "5 components for 80%" (canonical: ~4–5 effective dimensions, 4 reach 79.8% at n=36 / 81.6% at n=34); "1,820 cells / ≈2,650" (canonical: ~1,372–1,596 after the kind narrowing); the ranking-collapse measurement's status as an objection (retired by the operative frame). See the authority's §11 contradiction register.

Date: 2026-07-26 · Status: hardened design, no decision taken Input: creed-and-values-design-as-proposed.md Method: four profile-loaded lenses, read-only — architect (seams/types), doctrine-curator (integrity/exemptions/accreditation), reviewer (adversarial), implementer (feasibility/prototype).

Headline

The structure is sound. The arithmetic is not yet earned.

The design needs no new ArtifactKind, no new NodeKind, and no new Relation — so it touches none of this repo's fail-closed enum drift-guards and none of the three silent kind-drop sites. That is the single most important structural finding, and it is what separates this design from its rejected predecessor.

Against that, two measurements move the centre of gravity:

  • There is one live cost axis, not seven. Negative-cell census over the 36 authored vectors: minimal 19, maintainable 3, agile 2, reachable 2, extensible 1, environmental 0, solvable 0. Total 27 of 252 cells (10.7%), and minimal alone carries 19 of 27 (70%).

  • Creed-weighted ranking collapses toward magnitude ranking. Measured (34 authored vectors, Dirichlet creeds, Spearman vs the unweighted row mean, top-5 overlap vs the flat-creed top-5):

    Creed 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%

    The flat-creed top-5 is exactly the five largest row-sums. external-memory (largest sum, +3.50) appears in the top-5 for 53% of all creeds.

Recommendation: ship the prose layer; gate the numeric layer behind a one-day experiment (§7) whose failure condition is pre-registered. Everything else in this document is the shape to build if that gate passes.

⚠️ Read §0 first. The operator's stated frame — a limited-horizon heuristic that informs an agent, not a measurement — recalibrates most of what follows. §0 records what survives that reframing and what does not. In particular the ranking-collapse measurement above was calibrated to a ranking use and largely loses force under the stated frame; the sign-channel finding does not.

0. The operative frame — and what it changes

Recorded from the operator, and it governs the rest of this document:

Quantitative measures are unreliable, indeed. But a limited-horizon heuristic is a pragmatic approach that is "close enough" for practical purposes. Adjusting the weights of edges between behavioural atoms is exactly how neural networks are trained. Our goal is not to recreate absolute truth or reach a mathematical representation of an unrepresentable fact. It is to formally encode our graph in a way that is "good enough for government work" and improves the LLM's / agent's reasoning about the operator's intent and preferences.

Plus the systems-thinking framing that precedes it: an impacts edge is a causal-loop link — "more of X means more/less of Y" — with polarity and gain, the standard notation of causal loop diagrams.

0.1 What this reframing retires

  • The ranking-collapse measurement (§Headline, §5) loses most of its force. It measures fidelity of a sort order. Under the stated frame the creed's job is context, not sort order — so r = 0.98 against an unweighted row mean is not a failure, because reproducing a sensible global ordering is not what the artefact is for. The measurement stands as a fact; its status as an objection does not.
  • The false-positive finding retires entirely. It measured deriving links from vector-sign opposition. A causal-loop link is authored. Different mechanism; objection withdrawn.
  • "The numbers must be individually defensible" was the wrong bar. A heuristic that informs a reader does not need per-cell defensibility. This was a calibration error carried through the earlier squad rounds.
  • The causal-map notation is already in-repo, unregistered. src/doctrine/templates/diagrams/{mermaid,plantuml}/examples/causal-map-*-template.md ships s (reinforces) / o (undermines) links and states its purpose as "Systems thinking: mapping feedback loops between practices and outcomes." This is the third instance of the pattern (after the three-tier meta model and the AMMERSE definitions): the model shipped as a drawing aid instead of as structure. The design is promoting an in-repo formalism, not importing a foreign one — a materially stronger position than the input doc claims.

0.2 The lever/goal split — a free improvement from the in-repo template

The causal-map template types nodes as Practice / Outcome / Risk. That is the distinction that dissolves the "two dead axes" finding: practices are levers, outcomes are goals. solvable and environmental never going negative across 36 authored vectors is not two dead axes — it is the basis conflating levers with goals. Goal variables sit at the end of causal chains, so nothing pushes them down.

Recommendation: type the axes. A creed weighting over goal-axes and over lever-axes means different things, and a causal-loop model would separate them by construction.

0.3 What survives the reframing — three items

1. The vocabulary: it is a gain, not a correlation. This matters practically, not pedantically, because it determines which composition rule is legal:

  • correlation is symmetric (r(X,Y) = r(Y,X)); a causal link is directed
  • correlation is measured over observations; an authored judgement is an elasticity or gain — a structural parameter
  • correlations do not compose: 0.5 then 0.5 is not 0.25, it is underdetermined
  • standardized path coefficients do compose multiplicatively — but only in acyclic graphs, under linearity and no unmeasured confounding

Causal loop diagrams are cyclic by design. So base + 0.5×first-order + 0.25×second-order is a truncated Neumann series of a feedback loop, valid only while total loop gain < 1. "Limited-horizon" is exactly the right instinct — this names the horizon precisely and gives it a checkable convergence condition instead of an unstated one.

2. The sign-channel finding does NOT retire, and it is the one that bites a heuristic. §2 measured fail_fast at maintainable +0.50 while its own rationale says "leading to potential long-term maintenance challenges." An agent reading the number gets the wrong steer; an agent reading the prose gets the right one. Under a heuristic-that-informs-an-agent frame this is the sharpest remaining risk, because the failure is confidently wrong in the direction that suppresses a real concern — which is worse than absent data, not better.

The floor that makes "good enough" actually good enough: never render a number without its rationale, and treat the rationale as authoritative where they disagree. The corpus shows the prose is right when the number is wrong.

3. The neural-network analogy holds for the conclusion but is missing one property — and supplying it is the design's natural next step. NN weights work despite individual meaninglessness because of three things: a loss signal, gradient updates rather than authoring, and massive redundancy so errors average out. This design has none of the three: weights are authored once by judgement, there is no loss signal, and there is one number per axis per artefact with no redundancy.

So the analogy correctly supports "individual weights need not be right" — but only because NNs have a correction mechanism. Without a feedback signal an authored weight set does not converge; it simply is, carrying whatever bias the author had.

That gap is cheap to close, and closing it is what makes the analogy real: the design already has a candidate signal — override events. When a creed-informed suggestion is rejected or overridden, that is a gradient. A deviation ledger that records (suggestion, override, reason) turns a static authored matrix into something correctable. Recommended as the design's own next iteration, not as a precondition.

0.4 The experiment, re-scoped to the stated goal

§7's arms were built to test ranking fidelity. Under the stated frame the right question is narrower and cheaper:

Does an agent given the creed surface the trade-off rather than resolve it silently?

That is exactly #2538's existing locked metric — surfaced vs silent — and it needs no ranking function, no top-5 threshold, and no complete scoring. Keep §7's perturbation-stability probe (it is cheap and it still tells you whether the instrument is stable), drop the top-5 fidelity threshold as the primary gate, and promote surfaced-vs-silent in its place.

0.5 The cheapest next experiment, unchanged by all of this

Compute the spectral radius of the 7×7 connascence matrix. If it is ≥ 1 the 0.5 / 0.25 truncation is unsound and you learn that before anything ships; the sign of cycle products tells you which loops are reinforcing versus balancing. An afternoon on a 7×7, no schema, no artefact scoring, no interview. Blocked only on the coefficients, which are not in this repository — both in-repo copies defer to the external article.

1. Corrections to the input design's own evidence

Found by independent re-measurement. Two of these are corrections to §5 of the input doc.

  1. Co-location is a convention, not an invariant. The input doc claims every number is paired 1:1 with a rationale. Falsified: the_point_of_dissent.md carries 7 deltas and zero rationale keys; quad_A_test_structure.md the same; TEMPLATE_PRACTICE.md has 7 empty-string rationales. ≥3 of 38 violate it. So a mandatory rationale is a new constraint the calibration corpus fails — it needs the same advisory-vs-hard decision the negative rule needs.
  2. The 39% figure is contaminated. Two of the 14 all-non-negative vectors are not authored artefacts: TEMPLATE_PRACTICE (the authoring template) and quad_A_test_structure (an unfilled stub) — both all-zero. Honest figure: 12 of 34 (35%). Both stubs also sit inside the n=36 PCA and the per-axis means.
  3. §5.3 survives the correction. PCA re-run at n=34 with stubs excluded: cumulative 0.339 / 0.561 / 0.712 / 0.816 / 0.900 — still 5 components for 80%, per-axis signs unchanged. The dimensionality claim is solid and independently replicated. Republish §5.2–§5.6 at n=34 anyway; it costs the argument nothing and removes an easy shot.
  4. "Two axes carry no cost information" understates it. Five of seven axes have ≤3 negative cells. Restate as one cost axis and six benefit axes.
  5. Parts 2 and the connascence matrix have zero grounding in §5. Every measurement in §5 is about the artefact→value delta field. Nothing in the 36-vector corpus calibrates an edge strength or an inter-value coefficient — and 774 DRG edges is 23× the entire calibration set. Split the document so §5's credibility does not travel across the section boundary.
  6. "The interview is the producer" answers ~0.5% of the register. The creed is 7 weights
    • 7 rationales = 14 cells. The design also adds 260 behavioural artefacts × 7 × (delta + rationale) = 1,820 cells, plus 774 edge values, plus a 7×7 matrix ≈ 2,650 cells total.
  7. The design retires nothing. in_tension_with is still 2 edges and is the subject of an Accepted ADR this design does not propose superseding. Net: +5 registers, −0.

2. The 39% question — settled, and the answer is a third option

The input doc says this cannot be resolved from the data. It can. Twelve of the fourteen files were read in full.

Eleven of twelve name an explicit cost in prose while the number sits at or above zero:

Practice Axis delta Rationale (verbatim excerpt)
manual_of_me minimal +0.25 "introduces documentation and cognitive overhead… it is decidedly not minimal"
fail_fast maintainable +0.50 "overhead of maintaining complex validation rules… potential long-term maintenance challenges"
wipe_the_board maintainable 0 "the daily upkeep burden limits long-term sustainability"
external-memory minimal +0.75 "there is a risk of over-complicating the external memory system"
LARS minimal 0 "may introduce complexity in managing emotional aspects"
define_test_boundaries minimal +0.20 "moderated slightly by the complexity introduced"

Compare against controls, where the same author does net correctly: AMMERSE_impact_analysis minimal −0.60 ("introduces significant complexity and effort, conflicting with minimalism"), avoid_gold_plating extensible −0.70, be_a_STARR_at_interviews minimal −0.65.

manual_of_me scores minimal +0.25 for "decidedly not minimal". AMMERSE_impact_analysis scores minimal −0.60 for the same claim. 0.85 apart, opposite signs, near-identical prose.

Verdict: the 35% is not a measurement of authorial honesty — it is a measurement of noise in the sign channel, the one field a mandatory-negative rule reads. Eleven of twelve authors did the honest analysis; the number failed to record it. Exactly one entry is a true slogan (the_point_of_dissent: bare numbers, no rationales) — and a sign-reading rule would pass it while rejecting the eleven careful ones.

This also replicates the earlier squad's other measurement, which §5.3 did not test: sign instability under equivalent input. It held for an LLM across two passes; it holds for a human across a corpus.

Consequences, and they are the design's most important:

  1. minItems: 1 negative must not ship as a schema constraint. Not because the corpus is slogan-shaped (3%), but because it gates on the least reliable field and inverts the outcome it wants.
  2. The constraint belongs on the prose: every rationale must name what adopting this makes worse. Eleven of twelve already comply.
  3. The highest-value new component is a sign-vs-rationale-polarity lint — flag any cell whose rationale names only a cost while delta ≥ 0. It is mechanically checkable by an LLM on prose the design already mandates, and it is the only component of this design with a validation set available today (the twelve rows above). Make it Part 0.
  4. The rationale is the surviving channel. In fail_fast and wipe_the_board the prose is right and the number is wrong. This is strong evidence for the design's own counterpoint that the rationale fields are pivotal — and against the numeric layer.

3. Hardened structure — decisions

3.1 FoundationalValues — not an enum

The "half-open enum" dissolves. Role in src/doctrine/agent_profiles/profile.py is already a proven, lint-clean, Pydantic-wired half-open value object in this codebase — copy it rather than invent a pattern. Openness then lives in a substitutable artefact resolved through the existing built-in → org → project layering (src/doctrine/base.py), and the validation target becomes the active value set rather than a module literal.

⚠️ BLOCKER — do not expose the value-set repository as a DoctrineService property. tests/architectural/test_artifact_selection_completeness.py:55-62 introspects every @property on DoctrineService and demands a matching selected_<kind> on DoctrineSelectionConfig and required_<kind> on OrgCharterPolicy. A property silently converts a 2-file change into a three-way lockstep across two packages. Use a named accessor: resolve_active_value_set().

3.2 Two fields, two types — value_impact vs value_bias

The input design uses one concept for two semantics, distinguished only by which artefact carries it. Split it:

value_impact value_bias
Semantics Δ from not adopting weight — "how much I care"
Arithmetic summable / composable not summable (renormalise or average)
Position in creed · deltas right-hand side left-hand side
Carried by directive, tactic, styleguide, procedure agent_profile, paradigm, and the creed

Make them distinct Pydantic model types, not one model with a mode discriminator, so the type system structurally prevents dotting a profile with a directive (DIRECTIVE_043).

The creed's field is value_bias too — the creed is the project's profile. And paradigm belongs on the bias side: a worldview is a stance, not a Δ-from-baseline.

3.3 The value-bearing kind set — positive, declared once, in code

_VALUE_IMPACT_KINDS = {directive, tactic, styleguide, procedure}
_VALUE_BIAS_KINDS   = {agent_profile, paradigm}   # + the charter creed

OUT, with reasons: toolguide (consulted, not adopted — no Δ-from-not-adopting; zero of 36 corpus entries is tool documentation); mission_step_contract (structural wiring; a vector double-counts the directives its action resolves); glossary_pack (terminology has no adoption delta; also has no JSON schema file at all); template, asset, anti_pattern (align with _NON_AUGMENTATION_ELIGIBLE_KINDS).

Two points that matter more than the membership:

  • Declare it as a named frozenset in artifact_kinds.py, not as a prose sentence. That file's own docstring forbids a second kind enumeration, and there are already ten divergent kind lists in this codebase. A prose sentence becomes the eleventh.
  • Positive, not negative. _NON_AUGMENTATION_ELIGIBLE_KINDS is an exclusion set, so a new kind is silently included by default. For value-bearing that default is wrong — a new kind would silently acquire a scoring obligation nobody authored.

anti_pattern is structurally incapable of carrying this, three ways: its nodes hold only urn/kind/label/tags so there is nowhere for the mandatory rationale; DRGNode has no model_config; and extractor._KIND_MAP lacks it entirely. Negative-only is its natural form — which is exactly why it must be derived as the sign-flip of whatever rejects it, never authored.

3.4 impacts — a field on edges, and in_tension_with survives

Adopt: an optional numeric annotation on DRGEdge, sibling to the existing when and reason. Reject a new Relation.IMPACTS member — it is semantically wrong (an impacts edge with no relation type says two nodes interact without saying how) and it fires the RELATION_DESCRIPTIONS totality gate plus verbatim doc parity.

The load-bearing sub-decision: impacts does not make in_tension_with one band of a continuum, and must not be allowed to. in_tension_with is a typed existence claim with a lifecycle — symmetric, canonically stored, consumed as set membership by _tension_candidate_pairs and both-sides bridging by _tension_reconciled_urns. A band makes pair membership a function of a tunable threshold, and the earlier verdict measured that no threshold separates genuine tensions (4–6) from unrelated pairs (2–4) — they overlap at 4.

{relation: in_tension_with, impacts: "-0.7"} reads "this authored tension is severe." impacts is a strength annotation on an authored, typed edge — never a membership criterion.

scan_unreconciled_tensions is untouched. ADR 2026-07-21-1 is untouched — no superseding ADR needed. Allow impacts on any relation; document it as meaningful on in_tension_with / rejects / refines and ignored elsewhere. No per-relation table, because there is no totality guard for a Relation-keyed table and a table with no guard rots.

If the intent really is to retire in_tension_with, that is a different, larger change and it does need a superseding ADR. Say which reading is meant.

3.5 delta — string on the wire, Decimal in memory

Neither plain string nor float. condecimal(ge=-1, le=1) plus a mode="before" validator that rejects a non-str raw input — because Decimal(float) is lossy, so an unquoted delta: 0.9 (which ruamel yields as float) must be a hard error, not a silent conversion. This round-trips the corpus byte-identically, permits exact arithmetic for the composition formula, and needs no comparison epsilon.

Do not enforce a 0.05 resolution — wipe_the_board authors 0.125, so that rule also rejects the calibration set.

3.6 The connascence matrix — parked, and the composition invariant is a type

Do not ship the matrix in the first commit. The first-order matrix is not in this repository; both existing copies defer to an external URL, and the coefficients are the trademarked party's uncalibrated judgement. That makes it precisely a schema slot with no available producer — a fourth instance of the pattern this design was written to avoid. Define the schema and park it.

When it lands: it lives inside the value-set artefact, not on the charter — a matrix and a value list that can drift apart is the Environmental-definition drift repeated at N² scale. Nested mapping with an after validator asserting squareness, key-set equality with values, and zero diagonal.

The composition invariant, enforced structurally. Authored doctrine already defines base + 0.5×first_order/6 and overall = base + 0.5×fo + 0.25×so. So the authored delta is base, and it is the only persistable term. Guarantee that with a type, not a rule:

Make the computed projection a frozen dataclass with no model_dump(), living in a module no writer imports, carrying a mandatory matrix_id / matrix_version stamp.

Every doctrine and charter writer in this repo serialises through model_dump(). A projection that cannot produce one can never be written back — by anybody, including a future agent who has not read this document. That closes the design's worst failure mode (a computed ripple laundered into an authored assessment, then re-composed against itself) by construction.

3.7 creed on the charter — and the validation seam is not where you think

⚠️ CharterYaml is frozen=True, extra="forbid", but in production it is only ever constructed. Every production reader goes through raw ruamel; model_validate against the on-disk document appears only in tests. So a creed: block on disk would be silently unvalidated — the mirror image of the three silent-drop sites. Add an explicit per-section CreedConfig.model_validate(...) at the sync seam.

⚠️ load_creed_config must be three-state — absent / partial / complete — and the absent state must surface as a consistency-check finding, not a logger.info. The canonical section-loader treats an absent section as "empty config" with an info log, and that is the decay shape that kept Directive.severity inert for 162 days behind green tests.

3.8 Accreditation

  • Required field on the loadable value-set artefact. Copy GlossaryPack.provenance (required, not optional) and SourceEntry. Fields: attribution (required, non-empty), trademark_notice, source{url, accessed_on}, authorization_basis. Everything else points here — do not replicate the notice per artefact.
  • NOTICE at repo root. LICENSE is MIT and says "Copyright GitHub, Inc."; appending a trademark line there would misstate who attributes what. Verify NOTICE ships in the sdist/wheel or the accreditation never reaches consumers.
  • Not the schema $comment. Schemas are generated from the Pydantic models; a hand-added comment is erased on the next regeneration. Use Field(json_schema_extra=...).
  • Unify the two drifted definition copies FIRST. The template copy has lost "impact on nature" from Environmental. Accrediting a misquoted definition is worse than not accrediting. Sequence: unify → reference by id → accredit once.
  • Close the AMMERSE tactic's provenance gap in the same change — it currently carries no attribution at all and cites "the AMMERSE practice article" with no URL. It is where accreditation is actually read.
  • Provenance uses existing machinery. import_candidates already has external_references[].attribution_reason — literally the field for this obligation — and zero authored instances, so the 36-vector import is its first real producer. The design should claim this. Add source_ref + source_digest on each imported vector so a hand-edit of an imported delta fails loudly. That is the check that would have caught the "impact on nature" loss.

4. The interview loop — the sharpest live objection

Part 3 hands question design to the model, and the chain closes on itself:

  1. The seven axes are all virtues. No operator says solvability is unimportant. With no constant-sum budget and no forced choice, independent scoring over seven virtues is a textbook acquiescence instrument, and its predictable output is a near-flat creed.
  2. Near-flat is exactly the regime where the creed is inert — r = 0.980, 97% top-5 identity with ignoring the creed entirely. The elicitation instrument most likely to be produced is the one that makes its own output irrelevant.
  3. The laundering step is provenance, not generation. The model picks the questions, the answers land in creed.yaml, and the same model class then reads it as the operator's authority. Nothing records who supplied each weight. Downstream, "the operator ranked maintainability 0.8" and "the model proposed 0.8 and the operator did not object" are byte-identical.
  4. Prior art says format decides the answer — declared value orderings moved revealed priorities by 0.145 with rankings inverting between elicitation formats. The one variable identified as outcome-determining is the one the design leaves free.

The design's counterpoints do not reach this: the rationale-fields argument defends the consumption side (and is accepted), but this is a production-side failure — and it 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: an authored, versioned question bank (the model administers, does not design); forced-choice or constant-sum (21 pairwise comparisons, or a 100-point budget across 7 — both make the flat creed unreachable); per-weight provenance (source: operator | model-proposed-accepted | default, and refuse 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, with a rationale. That is the one place the rule has unambiguous teeth.

5. Prerequisites that must land first

These are not part of the design; the design breaks silently without them.

# Prerequisite Why
1 extra="forbid" on DRGNode and DRGEdge They have no model_config → Pydantic default extra="ignore" → an authored impacts is silently dropped on load. The tags docstring documents this exact hazard verbatim.
2 _node_to_dict / _edge_to_dict learn the field Explicit field-by-field writers — even a correctly modelled impacts vanishes at graph regeneration. Model + writer + round-trip test in one commit, or do not add the field.
3 extra="forbid" on AgentProfile Has only populate_by_name=True. A value_bias: key loads and silently vanishes at model level. The single most likely way this design ships inert.
4 Unify the two AMMERSE definition copies Already drifted once. Do this before authoring a third.
5 Confirm generate_schemas --check is CI-wired No reference found in .github/workflows/. If unwired, model and schema can diverge.

The dual gate is real: every per-artefact field lands in both the Pydantic model (extra="forbid") and the generated schema (additionalProperties: false) — the latter is enforced only by test_artifact_compliance, not at load. ×4 kinds = 8 files minimum.

6. Sequencing

Three all-surface sweeps, not two — the design does not name its own impacts edge sweep as one:

#2467 (pack tiers, KEYSTONE)
  → #2591 component-type on all artefacts        ← cheap, reversible, additive marker
    → silent kind-drop closure + prereqs §5
      → THE EXPERIMENT (§7)                       ← DECISION GATE
        → [only if it passes] value_impact / value_bias fields
          → [later, separately] impacts on edges, connascence matrix

#2591 goes first: it is a default-open marker with no authoring burden, and the value field's applicability should be expressed in terms of component-type rather than hardcoding a kind-keyed approximation into schemas. Do not batch the two sweeps — one unreviewable bulk-edit occurrence map, and if the population plan is wrong the correction has to unpick both.

7. THE EXPERIMENT — one day, pre-registered, can kill the numeric layer

The proposed prototype (import the corpus, author one creed, implement ranking, eyeball whether top-N differs) is right in scope and wrong in its success signal: "the top-N differ" cannot fail. Two distinct weight vectors dotted against 36 non-identical vectors essentially always differ — that is arithmetic, not evidence. It is the same class of error as the earlier squad's one-axis finding.

The real question is narrower: does creed-weighted ranking do anything a creed-independent quality score does not already do, and is it stable under the wobble a real interview produces?

WP-P1 — corpus import + value set + pre-registration (~3h, no deps) Snapshot the 36 rows with rationales at a pinned commit SHA; author ammerse.value-set.yaml with attribution populated; land the definition-parity test in the same commit (expected red on first run — the drift is known; do not fix by loosening the test); commit two hand-authored creeds and the pre-registered expected top-5 before any ranking code exists.

Creed B differs from creed A on minimal only — the one axis carrying 70% of all cost signal. If flipping the sole informative axis barely moves the ranking, the mechanism is dead and you learn it from one fixture.

WP-P2 — the ranking function (~4h). Pure, no I/O; owns weight normalisation in one place; a missing axis is a hard error, never an implicit 0. Hand-roll Kendall τ-b / top-N / Jaccard (~40 lines; scipy must not become a dependency for a prototype).

WP-P3 — the experiment (~3h). Assert the pre-registered thresholds as a real pass/fail.

Three arms, all required:

  1. Baseline arm — rank by unweighted row-mean. If creed-weighted ranking correlates r > 0.95 with this, the creed is decoration.
  2. Creed-blind arm — the #2538 shape. If the returned set is unchanged, citing creed.yaml in a trace is decoration too.
  3. Perturbation-stability probe — jitter each weight by the wobble a real interview produces and report rank correlation. This is the arm that can kill the design in an afternoon.

Pre-register the failure condition, e.g. top-5 must differ from the flat-creed top-5 by ≥2 items at a moderate creed. Write it down before running.

8. What survives every lens — the minimum shippable thing

If the experiment fails, this is what remains, and it is worth shipping on its own:

A costs: free-text field on the four value_impact kinds — one sentence naming what adopting this artefact makes worse — plus an operator-authored charter statement naming what the project deprioritises, plus the sign-vs-rationale-polarity lint. No numbers. No 7-vector. No matrix. No edge weights. No ranking function.

Every objection in this document lives in the arithmetic: the sign channel is noisy (§2), one of seven axes carries cost signal, weighted ranking collapses toward magnitude, and the register is ~2,650 cells against a 34-vector calibration set. Nothing here touches the prose — and by the corpus's own measurement the prose keeps the only axis that discriminates.

Also surviving intact, and not contested by any lens:

  • The dimensionality claim (§5.3 of the input, independently replicated at n=34). The seven axes carry genuine independent information; the earlier one-dimensionality finding is dead.
  • 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. 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?", authored with a rationale, is cheap, defensible, and has 34 worked examples.
  • Charter-scoping the creed is the right home and cleanly retires the project-specific-data-in-a-reusable-library objection.
  • The replaceable value set genuinely defuses basis lock-in.
  • routing_priority refutes "3-for-3" as a law — 18 authored producers and live consumers from inception. The discriminating variable is a consumer in the same seam, which creed · deltas can satisfy. 3-for-3 is a strong regularity, not a structural impossibility, and should not be carried as a blocker against this design.

9. Open operator decisions

  1. Mandatory negative: advisory lint, not schema. Recommended by every lens. Confirm.
  2. Mandatory rationale: the corpus fails it (§1.1). Advisory or hard?
  3. impacts: annotate or replace in_tension_with? Annotate needs no ADR; replace needs a superseding one.
  4. Does the ranking use arithmetic at all, or does an LLM read creed.yaml as prose context? If the latter, the 1,820 numbers are decorative and the experiment changes shape. The input doc does not say, and the two readings have different costs.
  5. Interview instrument: authored question bank + forced-choice, or keep it model-chosen?
  6. The design's own self-assessment. The corpus already authored a vector for this very practice: AMMERSE_impact_analysis scores minimal −0.60 ("introduces significant complexity and effort") and agile −0.35 ("time-consuming and may hinder agility… the potential indirect support for better decisions is not enough to offset this"). It is the most credible thing available and it is free. Put it in the design.

10. Method limits

Four lenses, read-only, no tests executed — every "breaks silently" claim is a static read. The reviewer sampled 14 of 38 corpus files (the all-non-negative set plus controls), so the sign-inconsistency finding establishes that inconsistency exists at ±0.85 for equivalent prose but does not characterise its distribution; a coherent constant shift, under which sign is meaningless but ordering survives, was not ruled out and deserves the test. The Dirichlet sweep is a sensitivity analysis, not a prediction of real operator creeds, and the claim that LLM interviews yield near-flat creeds is inference from the virtue-framing plus the cited format-inversion prior art — the weakest joint in the strongest argument, and testable in a day by the same test-retest recommended in §4. The 260-artefact and 774-edge counts are greps, directionally right and unaudited. Two lenses diverged on toolguide and agent_profile membership; §3.3 adjudicates on semantics (adopted-vs-consulted; stance-vs-delta) and that adjudication is a judgement, not a measurement.

Edit this page