VALUE-02 — Real model decision

[daniele21]

Goal

Prove the first real multi-model deployment decision on one representative device: select 2+ models from the same Local LLM Server target, execute the same use-case benchmark protocol, preserve compatibility semantics, and return either an explainable policy-backed recommendation or an explicit no-rank outcome.

Dependency / state

Preparation and isolated development may proceed now, but VALUE-02 remains BLOCKED for integration/acceptance until VALUE-01D #120 is PASS and VALUE-01 is marked DONE. Do not move dev or invalidate the retained VALUE-01 PRE_REAL evidence before then.

Existing baseline

Already implemented on dev:

• multi-candidate campaign contracts and execution;
• strict-quality-dominance@1.0.0 policy;
• compatibility before recommendation;
• explicit no-rank when no candidate dominates;
• separate Quality / Performance / Resources evidence;
• same-case candidate drill-down;
• browser J0 exercising two candidates with policy-backed results.

Execution slices

VALUE-02A — Real target candidate discovery

Make successful discovery for a configured target feed the canonical campaign-planning candidate inventory, so Find best setup can see all /v1/models candidates on the selected Local LLM Server target. Owner: application planning + thin UI API adapter. No frontend-owned model truth.

VALUE-02B — Real multi-model campaign/browser harness

Add a target-environment operator path that selects at least two discovered models on one target, launches the frozen campaign, waits for terminal state, and exercises Campaign Results + same-case comparison through the built browser. Owner: tests/real_runtime/; inference remains Local LLM Server-owned.

VALUE-02C — Multi-model evidence verifier

Verify that retained candidate Runs share the intended use case/device/protocol, compatibility is evaluated before ranking, the persisted decision policy is strict-quality-dominance@1.0.0, recommendation/no-rank is justified from canonical results, and evidence artifacts remain inspectable/private-safe.

VALUE-02D — Representative execution

After #120 and exact-head PRE_REAL on the integrated VALUE-02 software head, execute the real 2+ model decision on representative hardware and retain evidence.

Acceptance

• 2+ real models discovered from one Local LLM Server target;
• same use case/device and frozen benchmark protocol;
• immutable Run per candidate;
• capability compatibility established before recommendation;
• policy-backed recommendation or explicit no-rank/trade-off outcome;
• no hidden weighted/universal score;
• separate Quality / Performance / Resources evidence remains truthful;
• at least one same-case cross-candidate drill-down is retained;
• real browser Results surface matches canonical backend decision;
• no project-owned listener/process residue;
• hosted CI remains deterministic/software evidence only; final VALUE-02D is REAL_ENVIRONMENT.

Non-goals

• parameter sweeps/configuration optimization (VALUE-03);
• repeated-load/thermal confidence (VALUE-05);
• regression policy outcomes (VALUE-06);
• LLS evaluation cutover (VALUE-07).

[daniele21]

VALUE-02 software preparation complete — integration still gated by #120

All three pre-integration software lanes are implemented on branches cut from the still-current dev@796dc110ca13ce84d9ed726ed85cac8ba6aeda63 and have exact-head deterministic CI PASS across Repository Health, Repository Validation, Browser Acceptance and Built Product.

• VALUE-02A / VALUE-02A — Feed configured-target discovery into campaign planning #129 / PR VALUE-02A: make configured targets multi-model safe #138 — acf66a560fe1481db74e2d08a53ea2dc7a83f89e: configured-target /v1/models discovery feeds canonical Campaign candidates; per-candidate first-party LLS identity/telemetry model attribution is rebound correctly; unhealthy probes preserve the last successful session inventory.
• VALUE-02B / VALUE-02B — Add real multi-model campaign and browser acceptance #130 / PR VALUE-02B: add real multi-model campaign acceptance #136 — 0a282ba17336c2359b4f1fbfc36ed01c1cc16737: target-environment multi-model Campaign + built-browser Results/same-case harness.
• VALUE-02C / VALUE-02C — Verify retained multi-model decision evidence #131 / PR VALUE-02C: verify retained multi-model decision evidence #137 — 75127587e0e814ef5c9597d8317c1197566a4235: bounded verifier over canonical Campaign/Run/same-case evidence and strict-quality-dominance@1.0.0 truth.

These are software-ready branches, not integrated VALUE-02. All three PRs intentionally remain draft and must not merge while VALUE-01D #120 is open, because moving dev would invalidate the retained exact-head VALUE-01 PRE_REAL handoff.

After #120 PASS: refresh A/B/C against current dev, rerun exact-head deterministic gates, integrate A before B (C may follow), run fresh PRE_REAL on the fully integrated VALUE-02 head, then execute VALUE-02D #132 as REAL_ENVIRONMENT with 2+ real models. No multi-model/device evidence is claimed yet.

[daniele21]

VALUE-02 combined integration preflight — PASS on current dev

The three prepared software lanes have now been validated together, not only independently.

Validation-only draft PR: #141

• base: dev@98a32f4f123df8355589768174da81757f592839
• combined head: 10570e555d306f620fb84c8c87e7895668cfa842
• diff vs current dev: exactly 10 VALUE-02 files (A=4, B=4, C=2)
• Repository Health: PASS
• Repository Validation: PASS
• Browser Acceptance: PASS
• Built Product: PASS

The integration branch preserves the exact already-validated B/C blobs and the A backend owner change; no extra recommendation/comparability policy was introduced while composing the lanes.

A README-only movement of dev was detected during this validation. It was classified as base drift, reconciled into #141, and VALUE-01D PRE_REAL was separately regenerated through validation-only PR #142. #120 now contains the fresh exact-tree handoff.

Current state

• VALUE-02A VALUE-02A — Feed configured-target discovery into campaign planning #129 / PR VALUE-02A: make configured targets multi-model safe #138: software-ready
• VALUE-02B VALUE-02B — Add real multi-model campaign and browser acceptance #130 / PR VALUE-02B: add real multi-model campaign acceptance #136: software-ready
• VALUE-02C VALUE-02C — Verify retained multi-model decision evidence #131 / PR VALUE-02C: verify retained multi-model decision evidence #137: software-ready
• combined A+B+C / PR VALUE-02: integrate the multi-model decision loop #141: deterministic integration PASS
• VALUE-02D VALUE-02D — Execute representative real multi-model decision #132: still gated by VALUE-01D VALUE-01D — Execute and retain the representative real-device run #120 and then fresh integrated PRE_REAL

No VALUE-02 change is merged to dev while #120 remains open. Hosted CI establishes software integration only; it does not satisfy the final 2+ real-model/device decision evidence.

[daniele21]

VALUE-02 integration path consolidated

The software preparation is now represented by one coherent draft integration PR: #141 (10570e555d306f620fb84c8c87e7895668cfa842).

• VALUE-02A/B/C content is all present in VALUE-02: integrate the multi-model decision loop #141.
• Combined exact-head CI on VALUE-02: integrate the multi-model decision loop #141: Repository Health PASS, Repository Validation PASS, Browser Acceptance PASS, Built Product PASS.
• Former lane PRs VALUE-02B: add real multi-model campaign acceptance #136/VALUE-02C: verify retained multi-model decision evidence #137/VALUE-02A: make configured targets multi-model safe #138 are closed as superseded integration paths; their branches/commits remain provenance only.
• dev remains 98a32f4f123df8355589768174da81757f592839 so the fresh VALUE-01D PRE_REAL handoff is still valid.
• VALUE-02: integrate the multi-model decision loop #141 must remain draft/unmerged until VALUE-01D — Execute and retain the representative real-device run #120 PASS.

After #120, #141 is the single software integration candidate to refresh/preflight/merge before fresh VALUE-02 PRE_REAL and VALUE-02D #132.

[daniele21]

VALUE-02 integration candidate refreshed on repo-template-sw 0.9.1 — 2026-09-03

Canonical PR #141 has been rebuilt directly on current dev@aad1c4d66cb12b299dacb47c82e373cb96f4227c.

• head: 51221a736c22421bec78545eaee4fcfd056d8f02
• relationship to dev: exactly one commit ahead, zero behind
• diff: exactly the 10 VALUE-02A/B/C files; no baseline/workflow churn
• Repository Health run 33742864571: PASS
• Repository Validation run 33742864470: PASS
• selector: ITERATION + STRONG
• required gates: repository guards + Python validation + frontend validation
• Python 3.12: PASS
• Python 3.13: PASS
• Frontend Node 24 check/test/build: PASS

Under the adopted 0.9.1 model, browser/product/Built Product are correctly skipped at the draft iteration stage; they belong to the integration preflight after #120 allows #141 to leave draft. No evidence reuse was attempted because the target baseline changed materially.

#141 remains DRAFT and MUST NOT merge while VALUE-01D #120 is open. Once #120 passes, mark #141 ready, let the 0.9.1 INTEGRATION selector run the required exact-head cone, then integrate only if that final evidence is green.

[daniele21]

VALUE-02 software integrated for manual product testing — 2026-09-03

PR #141 is merged to dev at 0b9c376f2dbcbb167f8df8488374c1b0c5d0ace2.

Integrated software scope:

• VALUE-02A target/model discovery;
• VALUE-02B bounded real multi-model campaign/browser harness;
• VALUE-02C retained decision evidence verifier.

Pre-merge 0.9.1 INTEGRATION / STRONG validation passed on the exact candidate tree, including repository guards, Python 3.12/3.13, Built Product, PRE_REAL browser flow, packaging and reusable validation evidence publication.

The merge was intentional to enable product-owner manual testing on one coherent dev baseline. This does not mark VALUE-02 acceptance complete: representative 2+ model REAL_ENVIRONMENT evidence / VALUE-02D remains pending, and the canonical outcome must still be a policy-backed recommendation or explicit no-rank result.