PR12.12 — Conservative Domain-Sufficient Directional ClaimEvaluation v1¶

Status: implementation contract
Architecture issue: #52
Base: main @ 31449efebeb675f37dcc5cc166a1e5f0c680f291

Purpose¶

PR12.12 is the first generic PR12 layer allowed to emit a real claim-wide ClaimEvaluation.

It does not introduce a new learned evaluator, score, confidence threshold, majority rule, or policy-specific conflict resolver. It deterministically materializes the direction already implied by:

1. the complete PR12.9 evidence-disposition universe;
2. the PR12.10 lineage/non-inference audit;
3. one exact runtime-authorized PR12.7 domain policy;
4. one exact HUMAN-approved PR12.11 requirement mapping;
5. the frozen PR12.6 sufficiency semantics.

The frozen sufficiency rule remains:

all required policy requirements explicitly COVERED

PR12.12 does not reinterpret it.

Authority chain¶

exact PR12.7 runtime-authorized policy
        +
complete exact PR12.9 dispositions
        +
exact PR12.10 lineage/non-inference basis
        +
complete PR12.11 mapping proposal
        +
terminal HUMAN APPROVE
        +
current process-local PR12.11 review admission
        +
exact PR12.11 application receipt
        ↓
full governed replay
        ↓
deterministic PR2 ClaimEvaluation
        +
strict PR12.12 audit receipt

PR12.12 adds no new HUMAN decision and no new runtime permission capability. Its conclusion is a pure deterministic consequence of already-governed inputs.

A serialized PR12.12 receipt is audit data, not authority. Validation fully replays the upstream governance chain.

Anti-cherry-picking invariant¶

The most important new boundary is:

PR12.11 requirement mapping
!=
directional evidence selection

PR12.11 answers which evidence semantically covers which policy requirement. It is not allowed to choose which relevant evidence counts directionally.

PR12.12 therefore places every exact PR12.9 disposition into the emitted ClaimEvaluation.evidence_assessments.

Consequences:

mapped SUPPORTS + unmapped CONTRADICTS -> conflict remains visible
unmapped CONTRADICTS                   -> cannot be discarded
unmapped SUPPORTS                      -> cannot be discarded
NOT_RELEVANT                           -> contributes no direction
INDETERMINATE                          -> relevant but non-directional

This prevents the HUMAN semantic mapping boundary from becoming a hidden cherry-picking mechanism.

Deterministic conclusion table¶

Let:

complete = PR12.11.required_requirement_coverage_complete
support  = any complete PR12.9 disposition is SUPPORTS
contra   = any complete PR12.9 disposition is CONTRADICTS

Then:

No v1 path emits ConflictStatus.RESOLVED_BY_POLICY.

PR12.6 defines no conflict-resolution rule, so PR12.12 must preserve a real support/contradiction conflict instead of deciding a winner.

Complete + no direction¶

PR2 requires SUFFICIENT_FOR_CLAIM to contain at least one relevant assessment.

That is compatible with the table. PR12.11 forbids NOT_RELEVANT evidence from covering a requirement. Therefore a valid complete mapping with neither SUPPORTS nor CONTRADICTS necessarily contains at least one mapped INDETERMINATE assessment. An all-NOT_RELEVANT universe cannot legitimately obtain complete required coverage.

Reliability and lineage¶

PR12.12 deliberately ignores reliability when choosing direction:

LOW SUPPORTS
HIGH SUPPORTS
UNASSESSED SUPPORTS

all contribute the same boolean fact:

has_support = true

Likewise PR12.10 lineage is mandatory upstream audit context but is not converted into:

• independent evidence counts;
• replication counts;
• weights;
• confidence;
• majority voting;
• positive independence.

Shared lineage may be important to a later policy version, but PR12.6 defines no such directional rule in v1.

Real PR2 ClaimEvaluation¶

The emitted evaluation is a normal immutable PR2 ClaimEvaluation.

Its fields are governed as follows:

claim_id
    = exact PR12.11 claim

policy_ref
    = exact runtime-authorized admitted policy

evaluator_ref
    = RULE:capability_lab:pr12_12_domain_directional_rule_v1

evaluated_at
    = exact terminal PR12.11 HUMAN mapping review reviewed_at

evidence_assessments
    = complete exact PR12.9 dispositions

coverage
conflict_status
conclusion
    = deterministic table above

rationale
    = fixed PR12.12 rule rationale

No caller parameter exists for any of those directional decisions.

Deterministic identity¶

The ClaimEvaluationId is derived from a domain-separated SHA-256 digest of every stored evaluation field except the id itself.

This means:

same exact governed basis -> same evaluation content -> same evaluation id

Callers cannot mint semantically duplicate evaluations by selecting a different timestamp, evaluator, rationale, conclusion, or arbitrary id.

The full evaluation also receives a separate domain-separated SHA-256 audit digest.

PR12.12 receipt¶

ClaimDomainPolicyDirectionalEvaluationReceipt binds:

snapshot_sha256
claim_id
subject_ref
concept_ref
claim_scope
as_of
policy_ref
specification_sha256
disposition_coverage_sha256
lineage_dependence_sha256
requirement_application_sha256
mapping_review_id
mapping_review_sha256
mapping_reviewed_at
evaluation_id
evaluation_sha256
coverage_status
conflict_status
conclusion

The receipt has strict canonical dict/JSON serialization:

• exact key set;
• exact scalar types;
• duplicate JSON object keys rejected;
• unknown/missing fields rejected;
• noncanonical JSON rejected;
• nonstandard JSON constants rejected;
• exact enum values required.

The receipt is not a capability or authorization token.

Replay validation¶

validate_claim_domain_policy_directional_evaluation_v1(...) does not trust the supplied evaluation or receipt.

It:

1. strict-validates their stored types;
2. fully replays PR12.11;
3. thereby replays PR12.9, PR12.10, PR12.7 authority and policy applicability;
4. revalidates the exact current HUMAN mapping-review admission;
5. rebuilds the expected evaluation;
6. rebuilds the expected receipt;
7. requires exact typed equality.

A forged conclusion, forged application-completeness boolean, forged evaluation digest, changed claim, stale policy authority, or stale mapping-review authority therefore fails closed.

Persistence boundary¶

PR12.12 creates a real ClaimEvaluation, but it does not silently mutate an EpistemicRecordSet.

Persistence remains a separate append-only operation under PR11.3.

Likewise PR12.12 does not:

select a preferred historical evaluation
supersede an older evaluation
derive PersonalCapabilityState
approve a state transition
derive progression
emit readiness/mastery
authorize an action
render Player Window presentation

Those boundaries remain downstream and separately governed.

Public API¶

The PR12.12 surface is exported from capability_lab.evaluation_policy:

ClaimDomainPolicyDirectionalEvaluationReceipt
DomainPolicyDirectionalEvaluationError
InvalidDomainPolicyDirectionalEvaluation

build_claim_domain_policy_directional_evaluation_v1(...)
validate_claim_domain_policy_directional_evaluation_v1(...)

claim_domain_policy_directional_claim_evaluation_sha256_v1(...)
claim_domain_policy_directional_evaluation_receipt_sha256_v1(...)

claim_domain_policy_directional_evaluation_receipt_to_dict(...)
claim_domain_policy_directional_evaluation_receipt_from_dict(...)
claim_domain_policy_directional_evaluation_receipt_to_json(...)
claim_domain_policy_directional_evaluation_receipt_from_json(...)

It is deliberately not exported from the package root.

Explicit non-goals¶

PR12.12 != LEARNED EVALUATOR
PR12.12 != CONFIDENCE MODEL
PR12.12 != EVIDENCE WEIGHTING
PR12.12 != RELIABILITY THRESHOLD
PR12.12 != MAJORITY VOTE
PR12.12 != RECENCY WEIGHTING
PR12.12 != REPLICATION COUNTING
PR12.12 != POSITIVE INDEPENDENCE INFERENCE
PR12.12 != CONFLICT RESOLUTION POLICY

PR12.12 != EVALUATION SUPERSESSION
PR12.12 != CURRENT-EVALUATION SELECTION
PR12.12 != PersonalCapabilityState
PR12.12 != DERIVATION
PR12.12 != PROGRESSION
PR12.12 != READINESS
PR12.12 != MASTERY
PR12.12 != SCORE
PR12.12 != PERMISSION
PR12.12 != PlayerWindow

Release criterion¶

PR12.12 is complete when the repository can demonstrate that one exact PR12.7–PR12.11 governed basis produces one deterministic PR2 ClaimEvaluation, preserving all relevant contradictory evidence and refusing to infer any strength, preference, state, or progression beyond the frozen domain-policy semantics.