Manifesto tier — verdict, corrections, and handover

⚠️ RECORD. This verdict killed the derived-tension mechanism; a successor design exists and is the build target: foundational-values-and-creed.md (authored impacts edges, creed on the charter — none of the mechanisms rejected here). Its "~47,900 candidate pairs" is C(310,2) over DRG nodes; the authoring denominator is 260 behavioural artefacts — different measures (authority §11).

Date: 2026-07-26 Method: two sequential adversarial squads, 8 profile-loaded agents, all read-only. Round 1 (grounding): researcher-robbie, architect-alphonso, doctrine-daphne, paula-patterns. Round 2 (dialectic): reviewer-renata (prosecution), designer-dagmar (defence), debugger-debbie (experiment), planner-priti (sequencing). Supersedes the mechanism proposed in manifesto-tier-primary-drivers.md.

THE CALL

Do not build the tier. Convert it into a measurement, and run the decision gate that already exists (#2538).

The diagnosis was right. The mechanism is dead. And the diagnosis was right about the wrong layer.

What killed the mechanism

The proposal named its own falsification condition: "if the manifesto cannot reproduce [RECONCILE_CHANGE_SCOPE_TENSIONS]'s stated ordering, the manifesto is wrong." Nobody had run it. Round 2 ran it.

Result: 0 reproductions out of 6 derivations (2 independent scoring passes × 3 weightings), producing 4 distinct orderings — worse than chance, which would yield ~1 hit in 6. The authored ordering appeared zero times, including under the weighting derived from this repository's own charter.

The failure is categorical, not calibration. RECONCILE_CHANGE_SCOPE_TENSIONS is not a precedence over values. It is a pipeline of three operator types:

Artifact	Role	Function
`tactic:change-apply-smallest-viable-diff`	generator	picks the file set
`DIRECTIVE_025` Boy Scout Rule	transformer	operates within that set
`DIRECTIVE_024` Locality of Change	guard	vetoes growth of the set

A scalar weighting has exactly one output type and cannot express "A determines the domain over which B ranges." Note the ordering's direction is ambiguous before you even start: "last" means most authoritative here (the brake vetoes), where in a precedence ranking last means weakest.

Feasibility analysis confirms it structurally: with d1 = v_SVD − v_025 and d2 = v_025 − v_024, we find d1 ≈ −d2, so the reproducing set is a measure-near-zero cone requiring ~zero weight on Minimal and Reachable — the two values the triple is about. The mechanism can reproduce the ordering only via a manifesto that declines to have an opinion on the trade the ordering exists to resolve.

The defence reached the same conclusion independently and conceded it before the experiment's results were shared. Two lenses, one categorical answer.

Three more independent failures

False positives: 5 of 5 (100%). Five deliberately unrelated pairs (a git-signing directive vs a docs-accessibility styleguide, etc.) all flagged. Opposition counts overlap: genuine tensions 4–6, unrelated 2–4, overlapping at 4. No count or magnitude threshold separates them.
The corpus is ~one-dimensional. It holds two archetypes — impose rigour (Ma+ E+ Mi− R−) and remove friction (Mi+ R+ Ma−) — so every cross-archetype pair is automatically "in tension" whether or not it can ever co-apply. This falsifies §4(a)'s "vector in 7 fixed dimensions" by measurement, corroborating the prior-art finding that AMMERSE's non-zero matrix off-diagonals mean the dimensions co-vary and are therefore not a basis.
Author-dependence in the deciding cell. DIRECTIVE_024's Solvable sign flipped between the same agent's two passes one hour apart — and that is precisely the cell determining whether a threshold is needed at all.

At 310 DRG nodes the mechanism emits ~47,900 candidate pairs, on which the one real tension does not stand out. The load-bearing control is the co-application filter, and that filter is exactly the human judgement the mechanism was supposed to automate away.

The finding that decides the cost question

The experiment surfaced two genuinely unrecorded tensions: DIRECTIVE_043 × smallest-viable-diff, and DIRECTIVE_044 × DIRECTIVE_024. Both carry zero tension edges. Both are already resolvable by machinery that ships today — required beats lenient-adherence.

The cheap path resolves them. The expensive path only flags them.

The re-frame that matters most

This investigation began from "current-generation LLMs optimise output over outcomes." The repo's own history says: yes — and the same bias already operated on the governance machinery itself.

Three arbitration registers shipped as Reachable-maximal slots — schema now, producers later:

Register	Born	Fate
`Directive.severity`	`8b25f444f` 2026-02-15	13/13 uniform `warn`, 162 days, zero producers, zero readers
`GovernanceConfig.enforcement`	`8b25f444f` (same commit)	`{}` — only consumers are a test asserting it is empty and a round-trip probe
`routing_policy` / `task_fit`	`45a451a16` 2026-07-04	inert for ~20 of 21 verbs, remediated the same day by copying one artifact's scores four times

All three shipped with green, passing tests the entire time — the tests proved Pydantic round-trips, not that anything populated or read the field. git log -S'severity: error' returns nothing: never populated once, ever.

Three for three: in this repo, a scoring register without a producer and a coverage gate in the same commit does not degrade gracefully — it goes silently inert and passes review.

§8.1's mitigation ("partial population must be a first-class state, never an error") is the Reachable-maximal move, and it is falsified by measurement, not argument. The manifesto proposal was itself an instance of the bias it diagnosed. That is the real finding.

What survives

The core diagnosis, inverted and sharpened. Doctrine does not lack ordering — it discards the orderings it already has. resolve_context walks requires unbounded and suggests bounded, then destroys the distinction: all_artifacts = scoped | required | suggested (src/doctrine/drg/query.py:119), returning a frozenset. Measured:

Action	union	scope	requires	suggests	soft share	depth-2-only
specify	4	3	1	0	0.0%	0
plan	26	7	12	7	26.9%	3
tasks	25	7	8	10	40.0%	3
implement	89	37	27	25	28.1%	8
review	79	30	39	10	12.7%	2

suggests is the majority relation (332 edges vs 259 requires). At implement, 8 artifacts arrive as a suggestion-of-a-suggestion and are presented identically to a hard transitive prerequisite.

§4(b) — mandatory negative components. The proposal's one contribution with no prior art, and it survived all four Round 2 lenses. It is why detection worked at all: the shape forced honest negatives where a positives-only rubric would have vector-washed all eight pairs to "aligned." It needs no kind, no vectors, no weighting, no complete scoring. It survives as an authoring constraint and review heuristic.
Detection on an already-co-applying set. Both passes flagged both authored tensions and missed neither. Real, useful, and requires none of the dead machinery.
A twelfth arbitration register, inert because of the union. src/specify_cli/calibration/walker.py:507 sets known_irrelevant = resolved_scope - required_scope, making the over-broad half of the R-005 inequality vacuously true by construction. The in-code comment is the confession: the harness needed a blanket absorber because the union makes a hard prerequisite indistinguishable from a depth-2 suggestion. Fixing query.py:119 un-blinds an existing, CI-collected instrument.
A new always-on surface nobody had identified. src/specify_cli/tool_surface/profiles/_render_helpers.py:47-65 renders to .claude/agents/<id>.md and twelve harness siblings, unconditionally, outside the 32k token budget, reaching every dispatched subagent on every harness with a cold context. Confirmed first-person: it is the only doctrine in a subagent's system prompt. Nothing else in the repository has that property.

What is dead

§	Claim	Why
§4(c)	Derived ordering reproduces the authored one	0/6, worse than chance; categorically impossible (sequencing ≠ precedence). Killed independently by two lenses
§7.4	Derive `in_tension_with`, keep authored edges as overrides	Option C of ADR `2026-07-21-1`, Accepted 2026-07-21 — explicitly rejected 5 days before the proposal was written. Plus 5/5 false positives on a one-axis corpus
§7.3	`value_impact` vector per artifact	Reintroduces field-authored relationships that `test_built_in_relationships_authored_in_drg` asserts are gone; collides with `#2216`'s `component-type` all-artefact sweep; its incremental-population plan is the exact recipe that produced `070edbd4f`
§7.2	`manifesto` node kind, three scopes	41–59 files / 2.2–3.8 kLOC measured from three real precedents (`d54470c83`, `ce9d20e6c`, `1e3dc8d2c`). Unjustifiable pre-gate. Also undeliverable as scoped: the kind universe is a triple-mirrored fail-closed literal, so an "org manifesto" means editing a library that declares itself project-independent
§6	Computable squad coverage / collinearity	Zero prior art; depends on the dead vectors. Was billed as "the first concrete thing the tier buys"
§5.2	The licensed non-delivery move needs the basis	Premise false. `tactics/built-in/stopping-conditions.tactic.yaml` already exists, is activated, and carries six inbound DRG edges including `scope` on `implement` and `review`. Grammar: "When X happens, I will Y"; worked example escalates to the human-in-charge. The capability was never missing — and its presence has not fixed the bias, which is the measurement the proposal needed and never took
§5's promise generally	Naming the defect makes it correctable	Prior art measured a declared value ordering moving revealed model priorities by 0.145 normalized, with rankings inverting between elicitation formats. §5.1 (deviation becomes statable) is supported. §5.2 (agent obeys) is not

Honest scope: diagnosis and review support. Not agent correction.

Corrections to the original document

Errors found by independent verification. Several corrections make the case stronger.

The vocabulary census is wrong on two rows. Claimed trade-off = 0 and long-term = 0. Actual: 35 and 15. Cause was my own regex bug — grep -E "trade-off\|tradeoff" searches for the literal string trade-off|tradeoff. The rhetorical close "a corpus that cannot say trade-off cannot arbitrate one" is falsified. The single-term rows stand: output 168, outcome 49 (all BDD "observable outcome"), horizon 1 (= "horizontal"), stewardship 4.
^DIRECTIVE_\d{3}$ is not the directive ID contract. It is ^[A-Z][A-Z0-9_-]*$ (schemas/directive.schema.yaml:40); RECONCILE_CHANGE_SCOPE_TENSIONS conforms. That pattern lives only at paradigm.schema.yaml:62. The real defect is sharper: because its ID is not DIRECTIVE_NNN, no paradigm can list it in directive_refs and no profile can cite it — verified zero inbound references corpus-wide. It is unreachable, not malformed. It is also itself enforcement: advisory — the arbiter is the most ignorable item in the bag.
"Completely inert" is wrong. The AMMERSE tactic is charter-activated (.kittify/config.yaml:53), ships in the default pack, and is a requires target of procedure:situational-assessment. True statement: activated and procedure-mandated, yet zero scores exist anywhere.
"Nothing catches a contradictory reconciliation" is wrong. scan_unreconciled_tensions() (src/charter/consistency_check.py:977) is live, canonical, flags half-reconciled pairs, and has a non-vacuous suite naming this exact triple. Missing piece is narrower: presence is validated; ordering semantics are not.
Counts: 13 paradigm DRG nodes, not 14 (test-first.paradigm.yaml sits outside built-in/ so is not a DRG node — a separate drift defect). 310 DRG nodes, not ~200. 6 artifacts carry "Adapted from patterns.sddevelopment.be", not a dozen — and the AMMERSE tactic carries no sddevelopment provenance, so "AMMERSE came across with them" is inference. in_tension_with = 2 ✓. 18 profiles ✓.
"All paradigms are methodologies" is overstated. brownfield-onboarding.paradigm.yaml is a genuine epistemic stance. Defensible claim instead: the paradigm schema has no field able to hold a value, weighting, or ordering, so worldview content degrades into prose.
§9 "must be authored, not ported" is partly refuted. The three-tier meta model is already in-repo as unregistered diagram templates — templates/diagrams/{plantuml,mermaid}/examples/structure-meta-model-*.md carry Primary Drivers, {{Derived Concept}} (e.g., Creeds, Principles), {{Observable Behavior}}, {{Effects}}. Better framing: the meta model shipped as a drawing aid instead of as doctrine structure.
The AMMERSE definitions are already duplicated and already drifted. Second copy at templates/architecture/ammerse-analysis-template.md; the template's Environmental definition has lost "impact on nature". Step 1 is a DIRECTIVE_044 unification of two drifted copies — a debt payment, not a speculative refactor.
The first-order interaction matrix is not in the repository. Both copies defer to an external URL. §8.3's "port first-order only" mitigation names the one component that is unavailable and is the trademarked party's uncalibrated judgement.
enforcement is not wholly unrendered (correcting Round 1). It is emitted — but only by _format_full_artifact_payload_body as a json.dumps(sort_keys=True) blob reached via --include/fetch recovery. Sharper claim: enforcement reaches the agent only after the agent already went looking, alphabetized inside a JSON dump — never on the line where the directive is first presented.
tactic.schema.yaml has no references.type enum (correcting Round 1). It is a free string, and TacticReference.type is ArtifactKind including glossary_pack. So such a reference validates at both schema and model level and is then silently dropped by extractor._KIND_MAP (11 entries; missing glossary_pack, asset, anti_pattern). Strictly worse than a rejected enum value.
There is no feature terminology guard. test_no_legacy_terminology forbids exactly two retired legacy nouns (see _FORBIDDEN_TERMS; not reproduced here — the guard scans docs/, so quoting them in prose reds it), and _SCAN_ROOTS excludes .kittify/. Do not sell the Purpose rewrite as gate-verified — nothing checks it.
#2917 is CLOSED (corrects standing project memory). The release blocker is #2934 plus 14 open P0s. #2537 is CLOSED → ADR 2026-07-21-1 accepted.

THE SEQUENCE — what to actually do

File unconditionally, regardless of the manifesto's fate

Close the three silent kind-drop sites. ~3 source files, ≈150–300 LOC:

#	Site	Failure
1	`src/doctrine/drg/query.py:231-241`	`resolve_transitive_refs` buckets all 16 `NodeKind`s, reads out 10 → new kind computed, silently dropped
2	`src/charter/context.py:670-682`	`_classify_artifact_urns`, four `elif`, no `else` → silent drop at the injection boundary
3	`src/doctrine/drg/migration/extractor.py:133-145`	`_KIND_MAP` has 11 entries; `.get()` → `None` → edge vanishes at extraction time

It is a shared enabler for four already-open issues (#2468, #2847, #2862, #2829), belongs under #2466, and retires the proposal's "step 1 is a pure refactor, available immediately" claim outright.

The first mission — make the thesis falsifiable, build no new tier

WP	Scope	Dep
WP01	Partition `ResolvedContext` additively — add `scoped_urns`/`required_urns`/`suggested_urns`; keep `artifact_urns` as the union so all 3 consumers are untouched. Publish the strength breakdown through the existing `test_surface_report`. Campsite: delete `Directive.severity` + `GovernanceConfig.enforcement` (dead since 2026-02-15); fix the `src/doctrine/README.md` kind table (10 rows vs 12 kinds, 2 non-kinds, 4 omissions)	—
WP02	Un-vacuum `walker.py:507` → `known_irrelevant = ctx.suggested_urns - required_scope`. Expect RED. The red is the deliverable. Do not re-widen the absorber to get green	WP01
WP03	Group Action-Doctrine under `Required:` / `Suggested:`. Render enforcement for the non-`required` minority only — the census is 22 required / 3 lenient / 1 advisory, and at `implement` 15/2/0, so rendering the field wholesale distinguishes 2 of 17 while adding a token to all. Critically: put it in `_format_inline_directive_body`, which feeds the always-on profile block and so escapes the 4-of-24 Action-Doctrine ceiling	WP02

⚠️ Before rendering the token advisory, resolve the homonym. It already means three incompatible things ({required, lenient-adherence, advisory} in doctrine; {advisory, enforcing} in contracts; Literal["advisory","blocking"] in runtime), with four more severity ladders alongside. Rendering it into agent context as-is is a new primary/merge-class footgun. Collapsing eight vocabularies to two is the real unification mission, it is independently valuable, and it is the prerequisite for anything the enforcement render could buy.

THE DECISION GATE — `#2538`, and it already exists

#2538 — "experiment: does missing tension-modeling cause bad deferment at depth?" · OPEN · P2 · milestone 3.3.x · under epic #2466 · "Rig is standing. Run and results pending."

It is a pre-registered controlled rig on the real specify → plan → tasks loop with six seeded deferment/ambiguity forks. Its metric is exactly right — "surfaced vs silent, because silently resolving a tension is the failure even when the pick is right" — and its threshold is locked: reproduced if, across a majority of runs, ≥2 forks are decided wrong or resolved silently. Its own triage says it should run before any tension build work and that "a null result is a legitimate outcome."

Do not build a second rig. Run the standing one as an A/B: arm A = today's flat context, arm B = WP03's strength-ordered context. Post the arm-B pre-registration as a comment before the run.

If arm B does not increase surfaced on ≥2 forks across a majority of runs, §7 steps 2–5 close as won't-do.

Two caveats for whoever runs it: specify resolves only 4 artifacts — there is nothing there to reorder, so weight the plan (26) / tasks (25) forks or it reads null for a trivial reason. And verify the rig still runs before starting WP01 — "rig is standing" is the author's claim, unverified, and the entire gate decays with it.

Prior art predicts arm B ≈ null (0.145 effect size). Running the gate first therefore has a positive expected value of roughly one avoided 41–59-file mission — the trade the proposal never priced.

Two human decisions — not agent tasks

The AMMERSE trademark question. "AMMERSE, AMMERSE Method, AMMERSE Theory, AMMERSE Value System are all trademarks of J.B Crossland." Shipping a trademarked value system as a built-in doctrine kind in a distributed CLI needs an answer. Cheap to get; a "no" kills the branch regardless of #2538.
The charter.md:16-18 Purpose rewrite. This is a declaration of the operator's own values and trade posture. An agent drafting it is precisely the self-scoring failure §8.5 names, and nothing verifies it (correction 12). Operator authors; an agent may then propagate mechanically. Apply §4(b) as the constraint: it must name something the project gives up. The material already exists — the throughline at charter.md:49-52, the dated horizon ("target 0 by 4.0") at :481-491, the never-auto-decide posture at :29-30.

Explicitly parked — with reasons

Parked	Reason	Unpark on
§7.3 `value_impact`	Field-authored relationships + `#2216` collision + 3-for-3 decay precedent	`#2538` positive and a coverage gate designed before the schema
§7.4 derived tension	Option C of an Accepted ADR	A superseding ADR, not a plan doc
§7.2 `manifesto` kind	41–59 files, unjustifiable pre-gate	`#2538` positive and `#2467` KEYSTONE landed and silent-drop closure landed
§6 squad composition	Zero prior art; needs the dead vectors	Only if the kind exists for other reasons
§7 step 6 outcomes tier	Downstream of an ungated prerequisite	Value tier shipped
Re-tier AMMERSE onto `glossary_pack`	Landing kind is 4 days old and not in `_KIND_MAP` — the edge would silently vanish. Also the basis still has no consumer	Silent-drop closure landed and a consumer exists
Populating `severity` / `governance.enforcement`	Permanently — in favour of deletion. Populating either reproduces `070edbd4f` voluntarily	Never; a consumer must be designed first, at which point it is a new field

Debt tickets worth filing now (independent of everything above)

RECONCILE_CHANGE_SCOPE_TENSIONS has zero inbound references and is itself advisory — delete-or-wire.
AMMERSE definitions duplicated and drifted (tactic vs template; "impact on nature" lost) — DIRECTIVE_044 unification. Home: #2080.
src/doctrine/README.md kind table drift. Home: #2080.
DIRECTIVE_043 × smallest-viable-diff and DIRECTIVE_044 × DIRECTIVE_024 carry zero tension edges while both directives are required and both mandate non-local work.
Nothing detects a schema field with zero producers — three registers proved it. A lint for that is worth more than any of the above.

Handover note to the parallel session (`fix/2934-demock-planning-closeout-test`)

What this was: a deliberately meta-level lens run alongside your grounded work, per the operator's framing. It is now finished and has produced a verdict, not an open question.

What you need from it — three things, in priority order:

src/doctrine/drg/query.py:119 is a real defect on your side of the fence. It unions scoped | required | suggested into a frozenset, destroying an authored strength ordering. Measured impact: 40% of the tasks doctrine surface and 28% of implement (25 of 89 artifacts, 8 of them depth-2-only) are soft suggestions presented identically to hard prerequisites. suggests is the majority relation in the graph (332 vs 259). If you are anywhere near context assembly or planning-closeout behaviour, this is upstream of you.
src/specify_cli/calibration/walker.py:507 has a vacuous gate. known_irrelevant = resolved_scope - required_scope makes the over-broad half of the R-005 inequality true by construction — the in-code comment admits it. It is a live instrument that currently measures nothing, and it went vacuous because of (1). If your work touches calibration, treat this as a known-red-when-fixed.
The three silent kind-drop sites (query.py:231-241, _classify_artifact_urns missing else, extractor._KIND_MAP) silently discard data with no error. Any work adding or plumbing a doctrine kind hits these.

What you should NOT inherit: no new doctrine kind, no value_impact field, no derived tension, no manifesto schema. All rejected above with reasons. If a future prompt suggests any of them, this document is the refusal.

The one transferable discipline, and it generalises well beyond this investigation:

A schema slot without a producer and a coverage gate in the same commit goes silently inert and passes review. Three for three in this repo, one of them inert for 162 days behind green tests. If your work adds a field, add its producer and its coverage assertion in the same commit, or do not add the field.

Where the artifacts are: this note and manifesto-tier-primary-drivers.md (marked superseded, kept as the investigation record) live on branch docs/manifesto-tier-analysis, worktree .worktrees/docs-manifesto-tier, based on main. Nothing was landed on either active mission branch and the parallel checkout was never touched.

Method note and limits

Eight agents, all profile-loaded, all read-only, no HEAD-moving git. Round 2 was assigned adversarial positions (prosecution / defence / experiment / sequencing) specifically because Round 1 converged uniformly deflationary and four reviewers each optimising for their own Reachable value is exactly the bias under investigation.

Honest limits on the verdict:

The experiment is one instrument sampled twice, not two raters. The 4-distinct-orderings spread is a lower bound on real disagreement. The categorical argument (sequencing ≠ precedence) does not depend on the numbers and is the load-bearing part.
Experiment 3's "unnoticed tensions" are confirmation, not discovery — the agent picked the pairs by reading, then scored them. A real discovery test requires blind scoring and a ranking in which unknown tensions rise above ~47.9k pairs of noise.
The 5/5 false-positive rate has a wide interval — 5 pairs of 47,900, chosen as expected-to-be-unrelated rather than adversarially hard.
A non-linear reconciliation could plausibly reproduce the ordering. What was falsified is the linear weighting the proposal specifies. A typed generator/transformer/guard composition might work — but that is a small composition algebra, not a manifesto, and it is a different proposal.
The scoring agent was a current-generation LLM scoring a rubric designed to detect current-generation LLM value bias, and said so: it gave the two scope-limiting artifacts the highest Reachable scores in its table. Whether that is correct reading or its own bias rendered as data cannot be determined from the inside.
Nothing was executed. All "what breaks" claims are static reads. No test was run by any agent.
The recommended blind experiment before any schema work: 3 independent scorers × 20 artifacts (10 known-co-applying, 10 known-unrelated) × 7 values, blind to the authored edges. Outputs: Krippendorff's α per value, precision/recall against the 2 authored + 2 nominated edges, and a PCA rank of the 20×7 matrix to settle the one-axis question. ~1 day. It cannot test §6, which needs profile vectors that do not exist.