Mission Specification: FoundationalValues and Creed Band

Created: 2026-07-26 Status: ⏸ NOT YET SPECCED — deliberately. Scope sketch only. Programme: Mission D of five — see the programme record doctrine-canonical-structure-remediation-01KYEYSD. Order: Blocked by A only. Independent of B1, B2 and C. Design authority: foundational-values-and-creed.md · Sequencing authority: manifesto-program-delivery-sequence.md

Why this is a sketch and not a specification

Operator guidance, 2026-07-26: do not over-specify or plan any mission after phase A. Missions run one at a time, each specced and planned only after the previous phase is finalized.

This mission's blocking design decisions are now all settled — D-1, D-5 and ADR-D8 were answered on 2026-07-26 and are recorded under Carried constraints below. What still argues for waiting is mission A: it adds extra="forbid" to AgentProfile and the zero-producer lint that both of this mission's new profile fields must satisfy, so the shape of that work is only knowable after A lands. D-4 (the interview instrument's question design) also remains operator-only and unanswered, though it gates only a slice.

Scope sketch

Doctrine resolves to an unordered bag of co-valid rules. When two conflict, the arbiter is whatever prior the consuming model brings — which for current-generation LLMs favours delivering output now. The system delegates its value arbitration by omission. This mission builds the value layer that closes that, under the operative frame: a limited-horizon heuristic that informs, never auto-decides, and degrades gracefully.

> ### The corpus is a reference, not a target (operator, 2026-07-26) > > The AMMERSE practice corpus is reference material, not a corpus to chase. Four consequences, > and the last cuts against us: > > - D-5's recorded consequence softens. An earlier revision required I9's import to supply or reject the ≥3-of-38 corpus entries lacking rationales. It does not: the corpus need not satisfy our rules — our authored artefacts must. Import it labelled as reference. > - I5 narrows to internal consistency. Unifying the two drifted definition copies is still real — the tactic says "use these verbatim" and routes to different text on a live shipped path — but it is a self-consistency fix, not fidelity to upstream. > - I1a keeps its calibration set, changes its subject. The 12 known rows stay useful for tuning the lint; the lint then runs against our artefacts rather than policing the corpus. > - ⚠️ Releasing the corpus as a target also releases it as a check. The "1,372 cells against a 34-vector calibration set" objection dissolves — and so does the only external referent the numeric layer had. That leaves I8's perturbation probe and I1a's polarity lint as the sole instruments. Thinner than before; know it going in.

  • I5 — unify the two drifted AMMERSE definition copies (tactic + template) with a parity test.
  • I6 — accreditation: required non-empty attribution, a root NOTICE, packaging inclusion.
  • I9 — corpus import + value-set artefact + a non-AMMERSE N ≠ 7 fixture.
  • I13 — N-generic first-order connascence matrix and its validator. ADR-D8 settled 2026-07-26 — unblocked.
  • I8 — perturbation-stability probe (~20 lines). Can falsify the mechanism; run it early.
  • I1a — sign-vs-rationale polarity lint, advisory, against the 12-row validation set.
  • I16 — advisory-homonym unification, 8 vocabularies → 2. Its own occurrence map (C2).
  • I4-WP01/02 — additive ResolvedContext partition; un-vacuum walker.py:507-509 (the RED is the deliverable).
  • I4-WP03 — Required/Suggested render grouping. Independent merit only; its arm-B purpose died with G5.
  • I10 — ranking function. Unparked by the D-3 arithmetic ruling.
  • Charter interview slice — see below.

Estimated 7–13 agent-days plus the interview slice, ~60–100 files.

The charter interview slice (operator ruling, 2026-07-26)

The creed creation band must include a charter interview slice. A creed that cannot be created is a schema slot with no producer — precisely what mission A's zero-producer lint exists to reject, so shipping the value layer without a creation path would fail A's own gate.

The design authority puts the creed on the charter (§7.7): project-specific data carrying value_set: <id> plus N weights, each with a rationale. So creation runs through the charter interview, and §7.7 already names the hazard:

> ⚠️ The charter's on-disk sections are not model-validated in production — every production > reader goes through raw ruamel; model_validate appears only in tests. So a creed: block would > land silently unvalidated.

That makes the slice concrete rather than vague:

  • Extend the charter interview to elicit a creed: value_set pointer plus N weights with rationales.
  • An explicit per-section validate at the sync seam, because the production readers do not validate.
  • A three-state loader — absent / partial / complete — whose absent state surfaces as a consistency-check finding, not an info log.

D-4 is settled (2026-07-26), and its answer shapes this slice. The original framing — authored question bank versus model-chosen questions — bundled two independent failures:

  • Flatness comes from the response format, not the authorship. "Rate Maintainable 0–1" yields ~0.8 for every value whoever asks it. The ruling is a constant-sum budget ("distribute 100 points across the N values"), which prevents a flat creed structurally. §13's own objection concedes this by saying "with no budget".
  • Provenance laundering is a separate concern with a separate fix: record per-weight provenance — operator / model-suggested-operator-confirmed / default. Required regardless, because §9 names override events as the only available gradient and a deviation ledger of (suggestion, override, reason) is what makes that real.
  • An authored question bank was rejected on a third ground the framing missed: it is basis-specific and breaks ADR-D10's N-genericity. A bank of 21 AMMERSE pairs cannot serve a consumer's 5-value basis, so shipping one makes every non-default basis second-class — the failure the N≠7 fixture exists to prevent. Elicitation questions live on the value-set artefact, authored by whoever authors the basis.
  • I8 runs before the instrument is built. It reports how much elicitation precision is worth buying: stable under wide wobble → cheapest honest instrument; stable only under tight weights → forced-choice rigour earns its cost; unstable everywhere → the numeric layer is dead and I10 dies with it.

The interview is a general doctrine-gap elicitation surface, not a creed-only tool (operator, 2026-07-26). Once it exists it is plausibly the cheapest route through several things currently priced as authoring sweeps — FR-018's 55 laundered suggests verdicts, this mission's ~126 bias cells, and artefact vectors generally.

What is in scope versus what stays closed. The plumbing above is in scope. I17's interview instrument stays closed — that is the question-bank design question (D-4: authored bank with forced choice, versus model-chosen questions), it is operator-only, and the design authority is explicit about why it cannot be agent-drafted: model-chosen questions over N virtues with no budget predictably yield a near-flat creed — the regime in which the creed is inert — and nothing records who supplied each weight, so a model's prior silently acquires operator provenance.

The 3.3.x milestone scoping of charter-interview work is not final and should be revised. It was assigned before the creed became a charter-resident artefact with a creation dependency. Re-decide the milestone when this mission is specced; do not inherit 3.3.x as settled.

What Phase 0 closed — and the caveat on it

Gate G5 failed: the #2538 experiment rig does not exist. Verified twice, independently, and re-verified against the fork: every in-repo hit for 2538 is a document; the rig's unique markers ("Reproduction threshold", "consulted-arbitration", "add CSV export", "reporting module") return zero hits across all 115 fork branches; git log --all -S returns only the commits that added the design documents. The investigation branch origin/docs/manifesto-tier-analysis is fully reconciled with main — nothing was lost in transfer. The rig was never built.

Therefore gate G3 is unreachable, and I14 (value fields — 14–18 code files plus ~1,372–1,596 authored cells, the programme's entire unquantified authoring tail) and I17 (the interview instrument, as distinct from the charter slice above) are CLOSED, not deferred.

*Be precise about why I14 closed. It closed on G3 being unreachable — not on a measured null, and not* on authoring cost, which the D-4 interview ruling plausibly collapses. "We could not run the experiment" is a weaker reason than "we ran it and it read null." Do not restate this as settled evidence when speccing.

Caveat, stated plainly: this proves the rig is not in any repository or fork branch reachable from here. If it exists in the operator's own environment, these closures reverse. #2538 itself is milestone 3.3.x and stays open.

Carried constraints — do not lose these when speccing

1. an explicit creed/bias matrix: a list of {value, bias, rationale} entries; 2. a creed description: free-text, multiline.

  • ADR-D8 — SETTLED 2026-07-26. I13 is unblocked. The asymmetric maintainable ↔ extensible pair is confirmed an error and symmetrised to +0.75 in both directions; the adopted claim is that the two values reinforce each other. Headline residual 4.70% → 4.37%, per-step gain 0.3892 → 0.3766. The amendment lands on ADR 2026-07-26-3.
  • Basis: the operator authored the AMMERSE analysis procedure this repository implements, and holds prior written consent from J.B. Crossland to use the AMMERSE idea and publish that procedure under his own IP. Crossland's value system is the acknowledged ideological root (§3). The adjudication is therefore the operator's own, on his own authority. Not provisional, not pending anyone.
  • No upstream report is owed. Earlier revisions treated one as an outstanding operator-only obligation gating the matrix — superseded. Accreditation and mention discharge the obligation in full, which makes the required non-empty attribution and the root NOTICE (I6) load-bearing rather than ceremonial.
  • Divergence record — keep it, but as provenance, not a gate. Our matrix differs from the published one at exactly one cell. Record which and why, with a source_digest so a later upstream change is detected. Good engineering; no longer a precondition on shipping.
  • Do not edit _ammerse-connascence-first-order.json or _ammerse-second-order.json. They are the published upstream data and the reproduction script's inputs. The repair is applied at consumption, in the value-set artefact — editing the source data would destroy the ability to reproduce the discrepancy for the upstream report.
  • D-1 — SETTLED IN FULL by operator ruling, 2026-07-26.
  • toolguide is OUT. A complimentary artefact, not a behavioural one, so there is no Δ-from-not-adopting to record. Confirms §7.4 ("consulted, not adopted"). _VALUE_IMPACT_KINDS stays {DIRECTIVE, TACTIC, STYLEGUIDE, PROCEDURE}.
  • *agent_profile is IN, on the bias side, and carries two fields* — more than §7.4 anticipated:

_VALUE_BIAS_KINDS = {AGENT_PROFILE, PARADIGM} (+ the charter creed) is confirmed.

Authoring tail this creates: 18 built-in profiles × N values = ~126 bias cells at N = 7, plus 18 creed descriptions. Real, but two orders of magnitude below I14's closed ~1,372–1,596 cells, and it is not a reopening of I14: this is the bias side (who is weighing), not the impact side (what an artefact costs). I14 was gated on G3; the bias side never was.

Both fields are new schema slots on AgentProfile, so C-009 binds them: producer plus coverage gate in the same commit, or they ship inert and mission A's zero-producer lint fails. Mission A also adds extra="forbid" to AgentProfile, which is a further reason this cannot precede A.

Consequence to handle at import: the calibration corpus fails a hard rule (≥3 of 38 entries lack a rationale). I9's corpus import must therefore either supply the missing rationales or reject those entries — it cannot import them silently. Decide which when speccing; do not let the import quietly weaken the rule to advisory.

  • D-5 — mandatory rationale is HARD on value descriptions and deltas. Settled by operator ruling, 2026-07-26, with a deliberate boundary: the numeric value is a warning signal; the rationale is the tie-breaker and the real value-add, so a number without one is the failure mode the layer exists to prevent. Hard on value descriptions and delta entries; not on DRGEdge relationships, unless the rationale there can be trivially inferred.
  • Do not unify into the compiled charter bundle. .kittify/charter/ is a regeneration target; divergence there is a staleness symptom, and a fix written there is deleted by the next regeneration.
  • The gain check must be start-vector-independent. Plain power iteration is fail-open on two-camp bases — a true gain of 1.0 reads as 0.33 from an all-ones start. That is the dangerous direction.
  • Never renormalise per tier — that is power iteration, and every artefact converges to the same attractor. Scale once at the end, from the sup-norm (1/(1+r∞) = 0.75 for AMMERSE), not the spectral radius, which overflows to 1.074 on an adversarial ±1 base. And no fixed constant bounds a sum — two real corpus vectors already sum past 1.7.
  • I10 must not rebuild what already failed measurement: creed-weighted ranking collapses to row-sum at r ≈ 0.98, and vector-derived precedence scored 0 reproductions of 6 — worse than chance. Run I8 first; it can falsify the arithmetic reading cheaply.
  • Land I5 before I9 so the parity test is green on arrival. If I5 slips, land it as xfail(strict=True) with the ticket ID inline — never as a plain red, which is indistinguishable from the honest P0 reds and invites green-washing.
  • I16 needs its own occurrence map — C2 forbids batching two all-surface sweeps.
  • Everything is N-parameterised (ruling (d)); axis identity is by id, never position. The N ≠ 7 fixture is what makes that a fact rather than a claim.
  • Do not expose the value set as a DoctrineService @property — an architectural test introspects every property and demands matching selected_<kind> / required_<kind> config fields. Use resolve_active_value_set().