Home / Projects / Orchestrata Ultimate / Decisions / ADR index / ADR-0022
ADR-0022: Interpretation mutation must be adversarially separated; promote the model-free detector
| field | value |
|---|---|
| Status | Accepted |
| Date | 2026-07-24 |
| Register | factory (docs/adr, 4-digit ids) |
| Source | docs/adr/0022-c2-independence-and-promote-c4.md — permalink at bdafa201a (unchanged on master since that commit) |
| Wiki register entry | ADR-0022 (wiki status: Accepted) |
The record
- Status: Accepted
- Date: 2026-07-24
- Spec: SPEC-033 §6
Context
Interpretation mutation (C2) generates alternative readings of a requirement and checks whether the ratified tests/anchors discriminate them. As specified, the same model generates the readings and authored the anchors meant to discriminate them. The cheapest way to "pass" C2 is therefore to produce anchors keyed to one's own generated alternatives — the detector satisfies itself.
Two further limits, stated plainly rather than assumed away:
- C2 is generator-relative. It cannot surface a reading it does not sample. Its silence is a lower bound on ambiguity, never an upper bound on coverage — and its silence is anticorrelated with its usefulness: quietest exactly when the model's posterior is confidently concentrated on one (possibly wrong) reading, noisiest when the spec is merely vague, which the cheap lexical detectors already catch.
- C4 — under-constraint detection over a declared formalization — is the model-free sibling of C2, and the spec buried it as a "cheap logical detector" while promoting the model-dependent one.
Decision
- Anchors are ratified and frozen before C2 runs. C2 may never influence the artefact it is tested against.
- Readings are generated by a different tier, under adversarial framing — the generator's job is to find a reading the anchors fail to exclude, not to confirm the author's.
- Inter-generator divergence is itself a signal: when independent generators produce materially different reading sets, the requirement is ambiguous regardless of whether either set is discriminated.
- A reading counts only if it is material — it must differ on some AC-observable output; otherwise C2 can manufacture unlimited unfalsifiable readings and block indefinitely.
- Promote C4 to first-class, ahead of C2 where a formalization exists: it is deterministic and has no generator-coverage limitation.
- State C2's ceiling in the spec with the same prominence T5 states mutation testing's.
Alternatives rejected
- Same-model generation — self-satisfying, as above.
- Rely on C2 alone for intent — generator-relative; promoting C4 gives a model-free floor.
- Unbounded reading generation — makes DoR unsatisfiable at the generator's discretion.
Consequences
- C2 costs a second model tier per requirement, and cannot run until anchors are frozen — an ordering constraint on the decision phase.
- The materiality criterion needs a ratified constant (ADR-0015) or it becomes a matter of taste.
- C4's promotion implies specs should carry a declared formalization wherever practical — a real authoring cost with a deterministic payoff.
Provenance
Round-3 Opus (#2, #3): the self-satisfying critique, the generator-relative ceiling, the anticorrelated silence, and the observation that the spec promoted the unsound detector over its model-free sibling.
Related ADRs
- Also mentioned: ADR-0015
Rulings cited
None.