The Manifesto Tier — doctrine's missing primary-driver layer

Date: 2026-07-26 Status: ⚠️ SUPERSEDED IN PART — the proposed mechanism is REJECTED. Two adversarial squads (8 profile-loaded agents) ran against this document on 2026-07-26. The diagnosis survives and was sharpened; the mechanism (§4c, §7.2–§7.4, §6) is dead, killed by an experiment this document itself nominated as its falsification condition. Read manifesto-tier-verdict-and-handover.md first — it carries the verdict, the corrected facts, and the actionable sequence. Sections below marked ❌ are retained as the record of the investigation, not as proposals. Known factual errors in this document are listed in the handover, §"Corrections". Scope: meta-level. Read-only investigation of the doctrine tier model, run as a complementary lens alongside the grounded architectural work on fix/2934-demock-planning-closeout-test. Sequencing: steps 2–5 of §7 target the 3.3.x doctrine-extensibility / packs band. Step 1 is available immediately.

Summary

Doctrine has no tier for values. Every value-shaped artifact is therefore filed under the nearest available kind — a technique, a rule, or a methodology label. The consequence is architectural, not cosmetic: doctrine resolves to an unordered bag of co-valid rules, so when two of them conflict the arbiter is whatever prior the consuming model brings. For current-generation LLMs that prior is deliver output now. The system delegates its value arbitration by omission.

Three findings make this concrete and actionable:

The omission is inherited and was deliberate upstream — doctrine was imported from a knowledge base that explicitly scoped itself to the middle tier of a three-tier model and excluded values by design (§2). Correct for a human audience; impossible for an agent.
AMMERSE is already in the repository, canonically defined, and completely inert — mis-tiered as an optional tactic with its value definitions buried in step-2 prose (§3).
A manifesto tier does not merely document values; it makes tension derivable instead of declared, turning O(N²) hand-authored reconciliations into O(N) one-time scorings (§4).

1. The slot is already declared, and it is empty

src/doctrine/README.md documents the tier table:

Artifact kind	What it defines
Paradigm	Worldview-level framing (e.g. test-first, DDD)

But schemas/paradigm.schema.yaml has four meaningful fields — id, name, summary, directive_refs — and all 14 instances are named methodologies, not worldviews:

# paradigms/test-first.paradigm.yaml — the entire artifact
id: test-first
name: Test-First Doctrine
summary: Prefer acceptance-first and red-green-refactor loops for
         deterministic delivery quality.

domain-driven-design, trunk-based, atomic-design, execution-lanes, shared-branch-ci — all techniques. A paradigm says which school of practice we follow. None says what we are for, what we would trade away, or over what horizon we are judged.

Three mis-tierings, one root cause

What it actually is	Where it currently lives	Consequence
A 7-dimensional value basis	`tactic:ammerse-impact-analysis`, step-2 prose	Discretionary ritual; inert on the graph
A precedence ordering	`directive:RECONCILE_CHANGE_SCOPE_TENSIONS`	Bespoke per triple; ID breaks the directive contract
Worldview framing	The `paradigm` README row	Slot declared, filled with methodologies

The precedence case is worth reading directly. directives/built-in/reconcile-change-scope-tensions.directive.yaml exists solely to arbitrate DIRECTIVE_024 (Locality of Change), DIRECTIVE_025 (Boy Scout Rule), and tactic:change-apply-smallest-viable-diff:

"When the three pull in different directions on a specific change, resolve in this order: smallest-viable-diff picks the file set and edit size first; Boy Scout Rule then governs cleanup strictly inside that file set; Locality of Change is the brake…"

That is a precedence ordering over commitments, authored as a procedure step inside a rule. Two tells that the type did not fit: its ID breaks ^DIRECTIVE_\d{3}$, and it is bespoke per tension-triple — there is no general ordering to derive it from, and nothing that would catch a second reconciliation directive encoding a contradictory ordering.

2. The omission is inherited and was deliberate

patterns.sddevelopment.be/about declares a sense-making meta model in three tiers:

Tier	Contents	Mutability
Primary drivers	Values, axioms, current context	immutable
Secondary elements	Creeds and behaviors	malleable
Outcomes	Effects resulting from behaviors	—

And it states its own scope explicitly:

This knowledge base explicitly targets "creeds and behaviors" — how practitioners act. It deliberately excludes primary drivers (values, fundamental laws), directing readers elsewhere for philosophical foundations.

Spec Kitty's doctrine was imported from that catalog — a dozen artifacts carry "Adapted from patterns.sddevelopment.be", and AMMERSE came across with them. So doctrine inherited the exclusion along with the content.

This is why the gap exists and why nobody noticed. It was never an oversight. It was a correct scope decision for a human-audience knowledge base whose stated mission is to "help you build judgement" rather than prescribe — so it can reasonably tell a practitioner to bring their own values.

That instruction is coherent for a human reader and impossible for an agent. An LLM has no values of its own to bring; it has a training prior. "Go elsewhere for your philosophical foundations" resolves, for an agent, to "use the default weighting" — which §5 shows is Reachable/Solvable-maximal. A deliberate boundary in the source became a silent defect in the consumer.

Doctrine implements only the middle tier — both ends are missing

A vocabulary census across all doctrine YAML (excluding generated graphs):

Term	Hits
`output`	168
`outcome`	49 — every one means BDD "observable outcome" of a test/step
`stewardship`	4
`horizon`	1 — and it is "horizontal scale"
`long-term` / `long term`	0
`trade-off` / `tradeoff`	0

The meta model's third tier is literally "effects resulting from behaviors", and doctrine has no vocabulary for it. So the value gap and the temporal gap are one shape: doctrine is a faithful port of "creeds and behaviors", with the primary drivers above it and the outcomes below it both absent. An agent needs both ends — the primary drivers to arbitrate with, and the outcomes tier to have anything beyond the current turn to optimise toward.

A corpus that cannot say trade-off cannot arbitrate one.

3. AMMERSE is already in the repository, mis-tiered and inert

src/doctrine/tactics/built-in/analysis/ammerse-impact-analysis.tactic.yaml

The seven canonical value definitions — the arbitration basis the entire doctrine graph needs — currently live inside the description prose of step 2 of a tactic:

  - title: Assign base impact scores per AMMERSE value
    description: >
      Canonical AMMERSE value definitions (use these verbatim as the scoring lens):
      Agile (A): The ability to adapt quickly to changes…
      Minimal (Mi): The focus on simplicity…
      …

And it is gated to discretionary use:

"Reserve the full numeric interaction-matrix analysis for high-stakes decisions" · "Skip this step for low- and medium-risk decisions"

Reachability is two edges: procedure:situational-assessment (requires) and tactic:reference-architectural-patterns (suggests, "for high-stakes decisions"). No artifact in the corpus carries an AMMERSE score. The words "maintainable" / "extensible" appear elsewhere only as English adjectives.

So the value system is present, canonically defined, and completely inert — a ritual that may optionally be performed on one decision, with zero purchase on the graph that governs every decision.

4. Why a manifesto solves precedence, and better than a ranked list

The Agile Manifesto's grammar is "we value A over B — while there is value in B." A preference relation with acknowledged cost. AMMERSE generalises it, and improves on a ranked list in three ways:

(a) It is a basis, not a total order. An artifact's position is a vector in 7 fixed dimensions, not a rank. This answers the obvious objection to a global ordering — that real engineering trade-offs are not totally ordered. Tensions stay local to a decision-class.

(b) Cost is structural, not narrative. A negative component is the acknowledged cost. A manifesto cannot be authored as a slogan, because a value-vector with no negative components is visibly a lie. Shape enforces what a prose convention could not.

(c) Tension becomes DERIVABLE instead of DECLARED. This is the architectural payoff:

Two artifacts are in tension wherever their impact vectors carry opposite signs on the same value.

Authoring cost drops from O(N²) hand-authored pairs to O(N) one-time scorings — and, far more importantly, it surfaces the tensions nobody noticed. The corpus today has 2 in_tension_with edges across ~200 artifacts; the unnoticed conflicts are the whole problem, and per-pair authoring can only ever record conflicts someone already spotted.

Checked against the one real example in the corpus:

DIRECTIVE_024 Locality of Change → Minimal +, Reachable +, Extensible −
DIRECTIVE_025 Boy Scout Rule → Maintainable +, Minimal −, Reachable −

Opposite signs on Minimal and Reachable. The tension that required a hand-authored in_tension_with edge plus a bespoke reconciliation directive falls out of the vectors. AMMERSE's own first-order matrix already encodes the structural reason: Minimal→Extensible is −0.75, Extensible→Reachable is −0.75.

Reconciliation then derives from the project's declared weighting of the basis rather than being authored per pair. RECONCILE_CHANGE_SCOPE_TENSIONS stops being a bespoke node and becomes a worked example of a general rule — which is also the honest test of this design: if the manifesto cannot reproduce that directive's stated ordering, the manifesto is wrong.

5. The payoff: the output-bias gets an exact name

"Current-generation LLMs optimise output over outcomes" is a true but unactionable intuition. In the AMMERSE basis it becomes a specific point in value space:

Value	LLM default weight	Why
Reachable	maximal	Achievable now, within the turn's constraints
Solvable	maximal	Address the problem in front of me
Maintainable	~0	Consequences land after the turn ends
Extensible	~0	Same — no future to optimise for
Minimal	ambiguous	Favours a locally-small diff, tolerates a globally-duplicated seam
Environmental	unweighted	No standing in the objective
Agile	apparently high, actually low	Optimises the current turn, not adaptability across turns

This is no longer a vague bias but a nameable, correctable weighting defect. And AMMERSE's own matrix predicts the failure rather than merely describing it: Reachable is negatively impacted by Extensible (−0.75), so a Reachable-maximising agent will structurally sacrifice extensibility. Not a quirk — a consequence.

Re-read the charter's nine Quality & Tech-Debt Standing Orders in this light: every one is a patch on a Maintainable / Extensible deficit. Their own stated throughline — "never trust a green check, a clean diff, or a confident summary" — is nine symptom-patches for one weighting defect, which is why the list keeps growing. Adding more rules cannot fix this: every rule enlarges the bag and makes the implicit arbitration more load-bearing.

Two things follow immediately:

Deviation becomes measurable. A repository manifesto declaring Maintainable > Reachable for a decision-class makes "this diff scores Reachable +0.8 / Maintainable −0.7" a statable observation instead of a reviewer's hunch.
The licensed non-delivery move gets a reason. Today every profile's avoidance-boundary is a scope limit ("not my job"), never a stop condition. All 18 built-in initialization-declarations are first-person job descriptions ("I am X. I do A, B, C. I do not do D"), so the only licensed outcome of an invocation is a deliverable — and the agent duly delivers one. In the basis the stop condition is expressible: escalate rather than deliver when the deliverable's vector opposes the manifesto weighting beyond a threshold.

6. The three scopes have precise meaning

Organizational manifesto — the weighting over the basis: what this organization trades for what. AMMERSE's "context", made durable.
Repository manifesto — weighting plus the horizon it is judged over; may override org weighting per decision-class. This is also where the missing temporal vocabulary (§2) enters the system.
Agent profile manifesto — the profile's characteristic value bias. This is literally "who am I and how do I see the world?", as a vector.

The profile case has an immediate practical dividend: squad composition becomes computable.

Renata: Maintainable +, Solvable +, Reachable −
Priti: Reachable +, Agile +, Extensible −
Paula: Extensible +, Minimal +, Reachable −

Adversarial-squad lens selection is judgement today ("profile-by-lens", left to taste). With vectors a squad can be checked for coverage gaps — if no member weights Environmental, that lens is structurally absent — and for redundancy, if two members' vectors are near-collinear. That is the first concrete thing the tier buys that nothing else does.

7. Proposal

Sequenced so each step is independently valuable.

Re-tier AMMERSE (pure refactor, no new kind — do this first). Lift the seven canonical definitions out of the tactic's step-2 prose into a first-class, queryable basis. The tactic remains, as the procedure for scoring against the basis. That is the correct split: basis = declarative value tier; scoring ritual = tactic. Nothing is lost and the mis-tiering is corrected. Correct the paradigm README row to "named methodology" in the same pass.
A manifesto node kind, three scopes (§6), holding the weighting over the basis, the horizon, and the observable signals that the manifesto is being violated.
Optional value_impact vector on directives / tactics / styleguides. Author incrementally — the graph must work partially scored, so score the artifacts that appear in known tensions and on hot paths first.
Derive in_tension_with from opposing signs, keeping hand-authored edges as explicit overrides. Add a validator that flags a reconciliation whose ordering contradicts the declared weighting — the consistency check that is impossible today.
Profile worldview.value_bias + escalates_rather_than_delivers, the latter expressed as a threshold in the basis (§5.2).

Naming. Use manifesto — it carries the Agile Manifesto lineage, says statement of what we value, and maps onto the meta model's primary-driver tier. Do not overload paradigm: that repeats the primary/merge footgun exactly (one token, two installed senses, 14 instances, live DRG edges). Do not use creed: in the source meta model that word names the secondary tier doctrine already implements.

A sixth item, deferred but named. §2 shows the outcomes tier is missing too. A manifesto gives an agent something to arbitrate with; an outcomes tier gives it something to optimise toward. It needs its own analysis, and the value tier is its prerequisite — you cannot judge an effect without a value to judge it against. On the record here so it is not rediscovered later.

8. Risks, and one non-negotiable

The non-negotiable. The existing AMMERSE tactic says its numbers are "informed guessing" and "conversation prompts, not deterministic verdicts", and that inputs "reflect the assessor's current mental model, not ground truth." That caution must survive the port. If manifesto scores become a gate that auto-decides, this fails as precision theater and deserves to. The manifesto's job is to make the trade explicit and the deviation visible, and to demand a rationale — never to compute the answer.

Remaining risks:

Scoring burden. N artifacts × 7 values is real cost, and a half-scored graph yields half-derived tensions. Partial population must be a first-class state, never an error; score tension-participants and hot paths first.
Basis lock-in. Seven values is someone's ontology. If an organization needs an eighth, the basis must extend — it should be pack-overridable, which is precisely why this belongs with the pack work. (Pleasing recursion: the basis itself should score well on Extensible.)
Second-order matrices are the weakest thing to port. Those coefficients are calibrated judgement, not measurement. Port first-order only to start; the standard path in the existing tactic is already first-order.
Derived tension could be noisy. Opposite-sign-on-same-value will produce false positives on artifacts that are merely unrelated. It needs a magnitude threshold and probably a decision-class scope, or reviewers will learn to ignore it.
Vector-washing. An agent asked to score its own output against the basis will score it favourably, for exactly the Reachable-maximising reason in §5. A deliverable's score cannot be self-attested by the profile that produced it.
Additive-only failure. A tier at the top of an already-large bag only helps if it replaces injected prose with an ordering. If it is purely additive it makes the context problem worse, not better.

9. Framing note for whoever specs this

§2 establishes that the primary-driver tier was deliberately out of scope upstream, for a human audience. That makes this a scope extension, not a bug fix — Spec Kitty is consuming a knowledge base outside the audience it was scoped for. Framing it as an inherited defect would be unfair to the source and would also understate the work: the tier has to be authored, not ported, because the upstream catalog deliberately never wrote it down.

Provenance

Prior art: AMMERSE Impact Analysis — the seven values, the −1..+1 scale, and the first/second-order interaction matrices.
Prior art: Pragmatic Penguin Patterns — About — the three-tier sense-making meta model and its stated scope exclusion (§2).
Secondary: Structured Knowledge Sharing.
In-repo evidence: src/doctrine/README.md (tier table), src/doctrine/schemas/paradigm.schema.yaml, src/doctrine/tactics/built-in/analysis/ammerse-impact-analysis.tactic.yaml, src/doctrine/directives/built-in/reconcile-change-scope-tensions.directive.yaml, src/doctrine/drg/models.py (relation vocabulary), src/doctrine/drg/query.py (resolve_transitive_refs — the unordered bag), src/doctrine/agent_profiles/built-in/*.agent.yaml, .kittify/charter/charter.md (Standing Orders).