Codesign SEAM 2: regime rules as recommendations — the five-knob priors table

[aarontrowbridge]

Important

Problem — The Aug-16 campaign's core physics finding (the whitening/ensemble/diagonal crossover is set by the measured noise/bias ratio; the twelve VENDOR_PROFILES encode per-platform drift scales, trust geometry, and measurement kinds) is trapped in the internal tier. Dr. Okafor's tune-up campaign needs those findings as RECOMMENDATIONS — calibration knobs proposed with confidence and provenance, scoped to the platforms the evidence covers — while the engine that derives them stays internal (the A1 boundary: prior values + provenance strings ship; the rule engine + profile internals never do).
Approach — a distilled priors table for the FIVE named calibration knobs (tr_frac, β, y_goal, GLS weighting, min_contrast), keyed by platform family with the profile census in every provenance string, served through the recommendation surface, with the off-profile audit query as F2's mechanical sensor.
Scope — in: the priors table (values + provenance + scope), the serving path, the audit query, tests · out: the rule engine, VENDOR_PROFILES internals, any Intonatissimo code (the distillation is generated content, regenerated by an internal-env script — the regeneration wiring to the freshness cadence is a named follow-up, not this slice).

Acceptance Criteria

• regime_rec_priors_live == 5 — one prior entry per NAMED knob (tr_frac, β, y_goal, GLS weighting, min_contrast), each servable through the recommendation surface (the amico_recommend query path or an amico CLI verb — read the code, pick the seam that serves the session's recommendations, and say which in the PR)
• Every prior's provenance names its profile scope (the platform families + the census it distilled from, with the public-scale caveat the source profiles carry) and its evidence chain (the campaign fixtures / skills / cards it cites)
• The priors table carries the census stamp (families + count + date); the dynamic-census contract is stated in the table's schema (a new profile changes the census → the table is stale → the audit surfaces it), and the audit query FAILS when a prior is applied outside its profile scope without the caveat surfaced
• Recommendation provenance is attribution-free in the A1 sense: prior VALUES + provenance strings only — no crossover logic, no profile internals, no vendor-confidential data (the profiles' sources are public-scale arXiv/meeting citations, which ARE shippable; the distinction is recorded in the PR)

Testing Decisions

Unit tests over the priors table schema + the serving path (vitest, following the calib_chain/rehearsal test patterns); the audit query is tested with a violation fixture (a prior applied off-scope without a caveat → the query fails); the off-scope WITH-caveat case passes (the caveat is the point).

Key Decisions

• Platform-FAMILY keying (spin / transmon / atom…), not vendor keying: the recommendation surface is coarse (the interview's platform axis); the vendor profiles aggregate into family priors, with the census + sources in the provenance.
• The values are recommendations with confidence + provenance — the existing recommendation mechanics own confidence capping (ledger-sourced confidence caps at medium; static priors state their confidence explicitly).
• Outcome events feed amico's own events.jsonl (the flywheel's input side) — the amico-side feed, not Telaio's spine (SEAM 4's).

Constraints & Invariants

• The A1 boundary: attribution never ships. The public-scale caveat rides every provenance string (the source profiles' own rule: "do not cite as device data").
• No Intonatissimo import is possible from amicode (private package) — the priors table is generated content committed in the amicode repo; its regeneration script may live internal-side (named follow-up wires it to the cadence).

Prior Art

The spec's SEAM 2 section (design-of-record: specs/spec-20260831-120000-amicode-outside-lab-codesign.md); Intonatissimo's VENDOR_PROFILES (the census + values, read in an internal env for the distillation — the table cites, never includes, it); the amico_recommend tool mechanics (propose/query/outcome, confidence capping); the hardware-loop skill's strategy-knob doctrine; the campaign's H-series fixtures.

Source

Part of #679 (the codesign PRD) · design-of-record: the spec's SEAM 2 · no blockers.

Notes

Second-wave decomposition of the codesign PRD (after #680/#681 landed). The priors' VALUES are distillation work: ground them in the campaign evidence (the H-series verdicts, the min_contrast 2.5 fixture validation, the measurement-kind → weighting mapping) — the PR must show the evidence chain for each value, not present it as taste.