VALUE-03 — Configuration decision

[daniele21]

Goal

Unlock the next user-visible decision after VALUE-02: choose the best model + quantization + supported configuration for one use case/device, with every searched value frozen into immutable Run identity and evidence.

Delivery / REAL_ENVIRONMENT policy

Software slices may be integrated incrementally into dev once their repository-owned deterministic gates pass. REAL_ENVIRONMENT acceptance is intentionally deferred until the planned software modifications are integrated and dev is the consolidated candidate. VALUE-03D remains required before final real-device claims, but it no longer blocks integrating VALUE-03A/B/C software.

Current gap on dev

• Find best setup already has the Optimization step and fixed/quick/standard/thorough/custom strategies.
• application planning already exposes supported_generation_parameters, bounded_generation_parameter_ranges and configuration_search_options.
• bounded_generation_parameter_ranges is currently always empty and only FIXED is executable.
• canonical GenerationConfig currently owns request-level generation fields (max_output_tokens, temperature, top_p, seed, stop, response_format).
• the product contract explicitly forbids inventing ranges; quantization remains candidate identity; runtime/model-load parameters may be searched only when the serving runtime declares an explicit mutable lifecycle.

Vertical slices

VALUE-03A — Canonical declared parameter domains

Introduce a typed backend-owned contract for runtime/endpoint-declared bounded parameter domains. Preserve source/provenance and distinguish request-level mutable parameters from observational/runtime-load configuration. Empty/unknown domains remain unavailable, never inferred.

VALUE-03B — Frozen configuration expansion

Expand a reviewed search strategy into a bounded deterministic set of configurations per candidate, include every configuration in the plan digest/estimate, and create one immutable Run per candidate + configuration. No hidden/random search.

VALUE-03C — Optimization UX

Render backend-declared domains through progressive disclosure, provide Standard as the normal default only when bounded domains exist, keep Fixed as fallback, and show exact planned configurations/request count before launch.

VALUE-03D — Representative real configuration decision

After the software convergence on dev, on one representative device/LLS target execute 2+ supported configurations for at least one real candidate, retain configuration fingerprints/evidence, and produce the canonical recommendation/no-rank result without inventing unsupported controls.

Acceptance

• no parameter is searchable unless support and a bounded domain are declared by a canonical owner;
• searched request-level values are preserved in GenerationConfig and therefore execution fingerprints;
• runtime/model-load parameters are not mutated unless an explicit serving-runtime lifecycle contract exists;
• quantization is never treated as a sweep knob;
• plan digest, estimates, campaign rows and retained Runs agree on the exact configuration matrix;
• unsupported/unknown values remain typed unavailable;
• real acceptance is REAL_ENVIRONMENT; hosted CI proves software semantics only.

Non-goals

• cross-device trade-off policy (VALUE-04);
• repeatability/thermal confidence (VALUE-05);
• LLS evaluation cutover (VALUE-07).

[daniele21]

VALUE-03A preparation status — SOFTWARE READY, integration blocked

Producer contract is integrated in Local LLM Server:

• daniele21/local-llm-server#205 merged
• LLS merge SHA e7e186c96b66f4ddb3a01e926b5a43b69aceeca6

Performance Lab consumer preparation:

• draft PR VALUE-03A: consume declared generation domains #149
• exact head caed430641d40228e9778155f5b8157daf5238a5
• base dev@0b9c376f2dbcbb167f8df8488374c1b0c5d0ace2
• Repository Health 33766112408: PASS
• Repository Validation 33766112384: PASS (Python 3.12/3.13 + frontend selected gates)

VALUE-03A now consumes only explicitly declared, validated, model-scoped request-generation domains and keeps runtime-load configuration separate. Missing/invalid domain metadata remains unavailable/fail-closed. It does not enable sweep strategies and does not yet translate Korgis max_tokens into Performance Lab max_output_tokens; deterministic expansion/fingerprint mapping remains VALUE-03B ownership.

Do not mark VALUE-03 ACTIVE or integrate #149 while VALUE-02D #132 remains open. #120 also remains without representative REAL_ENVIRONMENT PASS. Hosted validation above proves software semantics only.

[daniele21]

Software convergence update: VALUE-03A plus the first VALUE-03B frozen-configuration identity slice are integrated on dev (through dc2629acc678f1714fdd368bd72fe2478d96f188, now included in current dev).

Current software gap remains deliberate: declared parameter domains are consumed, but search strategies beyond Fixed are not yet expanded into deterministic configuration matrices. Before implementing the remaining VALUE-03B/C behavior, the exact sampling/cardinality semantics for Quick/Standard/Thorough/Custom must come from a canonical product/contract decision rather than being invented in code.

VALUE-03D remains REAL_ENVIRONMENT and is deferred until final software convergence.