ADR: Proposal-only quality evidence protocol

Hosts need to compare candidate AI interaction contracts on correctness, grounding, cost, tokens, and latency without moving provider execution or optimization into Templiqx. Extending run_eval would conflate deterministic fixture evaluation with externally measured, host-attested evidence. A weighted winner would also allow low cost or latency to hide an unsafe quality result.

External evidence is assertive, not self-authenticating. Templiqx can verify structure, cardinality, immutable bindings, comparability, and deterministic math, but cannot prove that a host evaluator, model, clock, tokenizer, or billing source was truthful.

Decision

Add one actor-neutral operation, assess_quality_proposals, across the canonical service and thin transports.

Templiqx loads the current package and base contract, validates every full candidate source, computes package/contract/fixture/candidate/policy/request identities, and distinguishes them from claimed evaluator, scorer, model, and measurement profile identities.
The current base contract’s fixtures define a fixed logical trial grid. Candidate-quality failures remain in semantic denominators; infrastructure failures remain in coverage, infrastructure, and resource calculations.
Integer-only, checked fixed-point aggregation follows the normative denominator and rounding rules in the quality proposal contract.
Every binary scorer has one nonzero same-metric gte ratio floor and is also a Pareto objective. All floors run before equal multi-objective Pareto comparison. The 950_000 and 850_000 reference floors are policy data, not hard-coded domain types.
The operation returns deterministic assessments and Pareto fronts. It never chooses a winner, mutates a package, persists evidence, publishes, applies, or promotes a candidate.
Existing run_eval, test_package, EvalFixture, and diff_contract behavior remains unchanged.

Ownership

The host retains model calls, provider retries, optimizer and evaluator execution, authentication, signatures, authorization, approval, evidence retention, selection, and promotion. Templiqx structurally validates claimed profiles but does not certify them.

Ax is permitted only as an external development-time optimizer. BAML informs contract and prompt design but is not integrated. Effect may be evaluated later as optional host orchestration tooling. This decision adds no Ax, BAML, Effect, provider SDK, optimizer, or new host-policy port dependency.

Consequences

A cheap or fast proposal below any hard quality floor cannot be Pareto-ranked.
Equal and incomparable eligible proposals remain visible; product policy must make the eventual selection explicitly.
Reports are reproducible and bound to exact artifacts, while evidence authenticity remains a host trust decision.
The public surface gains one additive operation and corresponding transport and SDK work, but no persistent migration or second runtime path.
A valid operation may return no eligible candidates and still have OperationEnvelope.ok = true.

Alternatives rejected

Extend run_eval: mixes external attestation with existing deterministic fake-eval semantics.
Embed Ax, BAML, or Effect: crosses the host boundary and adds runtime coupling without improving deterministic evidence semantics.
Weighted or automatic winner: hides tradeoffs and can let resource metrics compensate for unsafe quality.
Store or promote from assessment: creates authority and lifecycle semantics that belong to the host and bypasses the existing CAS mutation boundary.

Follow-ups

Hosts may define authenticated evidence/signature policy, build an external Ax runner, and expand sanitized benchmark corpora after the protocol is stable. Those follow-ups do not alter this proposal-only operation.