VALUE-02D — Execute representative real multi-model decision

[daniele21]

Parent: #128

Goal

Close VALUE-02 with one retained REAL_ENVIRONMENT multi-model decision on representative hardware after VALUE-02A/B/C are integrated and exact-head PRE_REAL is refreshed.

Depends on

• VALUE-01D VALUE-01D — Execute and retain the representative real-device run #120 PASS and VALUE-01 DONE;
• VALUE-02A VALUE-02A — Feed configured-target discovery into campaign planning #129 integrated;
• VALUE-02B VALUE-02B — Add real multi-model campaign and browser acceptance #130 integrated;
• VALUE-02C VALUE-02C — Verify retained multi-model decision evidence #131 integrated;
• fresh exact-head PRE_REAL/Built Product PASS on the final integrated VALUE-02 software head.

Acceptance

• one Local LLM Server target exposes at least two real candidate models;
• same selected use case/device and frozen benchmark protocol across candidates;
• all candidate Runs complete and are retained immutably;
• campaign results preserve capability compatibility before recommendation;
• canonical policy produces an explainable recommendation or explicit no-rank outcome;
• built browser Campaign Results and same-case comparison match backend-owned truth;
• VALUE-02C verifier PASS;
• bounded artifacts retained; no Performance Lab-owned process/listener residue remains.

Do not infer VALUE-03 configuration optimization, VALUE-05 repeatability/thermal confidence or VALUE-06 regression claims from this run.

[daniele21]

VALUE-02D software + exact-tree PRE_REAL handoff — 2026-09-03

The VALUE-02A/B/C software is integrated on current dev@f07232aaf92ea4f90d0da88290c9b7db97f3145e, and exact-tree FULL PRE_REAL has now passed through validation-only PR #151.

Evidence identity:

• validation commit 001fcd33ab2686032f6062a8b4c2351f77e0a6bb
• shared product tree 29757436d4fa81b87f2a277cc310544be4a25f8b
• Repository Validation run 33774603131: PASS
• PRE_REAL artifact pre-real-e2e-33774603131-1, digest sha256:9ab4819deac767a81c2c491c3f3cbe7bb45056028ccc5b6f4f9ed71a830a7adf
• final manifest: PRE_REAL_E2E=PASS, ready_for_real_environment=true

#132 remains BLOCKED by VALUE-01D #120; this comment does not waive that dependency.

After #120 PASS, keep the same evidence commit/tree (unless dev moves) and execute the existing multi-model harness with 2+ models exposed by one Local LLM Server target:

uv run --extra dev --locked python tests/real_runtime/browser_value02_local_llm_server.py \
  --base-url http://127.0.0.1:1235 \
  --model <model-a> \
  --model <model-b> \
  --output-dir .performance-lab/value02-real \
  --pre-real-manifest build/pre-real-e2e/manifest.json

The harness retains runs.sqlite3, campaign.json, case-comparison.json, browser evidence and value02-manifest.json. Then verify the canonical retained evidence:

uv run --extra dev --locked python tests/real_runtime/verify_value02_evidence.py \
  --store .performance-lab/value02-real/runs.sqlite3 \
  --campaign .performance-lab/value02-real/campaign.json \
  --case-comparison .performance-lab/value02-real/case-comparison.json \
  --output .performance-lab/value02-real/value02c-verification.json

VALUE-02 becomes DONE only if the representative multi-model campaign, browser decision/same-case drill-down and VALUE-02C verifier all PASS with the canonical recommendation or explicit no-rank result.