Make adaptive compute decision-safe and runtime-parity verified

.github/workflows/runtime-quality-gates.yml

@@ -37,8 +37,9 @@ jobs:
       - name: Install focused test dependencies
         run: |
           python -m pip install --upgrade pip
-          python -m pip install pytest flask pillow
+          python -m pip install pytest flask pillow matplotlib
           python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
+          python -m pip install peft transformers
 
       - name: Compile runtime surfaces
         run: |
@@ -113,8 +114,12 @@ jobs:
           if ($seal.execution_enabled -ne $false) { throw 'Frozen shadow seal unexpectedly enables execution' }
           $campaignId = $seal.public_package.campaign_seal.seal.campaign_id
           $clusterPath = Join-Path $PWD 'build\ci_route_shadow\cluster.json'
-          @{ cluster_identifier = 'ci-shadow-cluster-1' } | ConvertTo-Json -Compress | Set-Content -LiteralPath $clusterPath -Encoding utf8
-          $commit = (& .\dist\ci_route_shadow\SupermixRouteShadow.exe commit --registry $registryPath --campaign $campaignId --seed-input $seedPath --cluster-input $clusterPath --compact) | ConvertFrom-Json
+          $clusterIdentifier = python -c "from source.route_policy_ledger import hash_session_identity; print(hash_session_identity('ci-shadow-cluster-1'))"
+          if ($LASTEXITCODE -ne 0) { throw 'CI shadow cluster hash generation failed' }
+          @{ cluster_identifier = $clusterIdentifier } | ConvertTo-Json -Compress | Set-Content -LiteralPath $clusterPath -Encoding utf8
+          $commitRaw = & .\dist\ci_route_shadow\SupermixRouteShadow.exe commit --registry $registryPath --campaign $campaignId --seed-input $seedPath --cluster-input $clusterPath --compact
+          if ($LASTEXITCODE -ne 0) { throw 'Frozen shadow commitment failed' }
+          $commit = $commitRaw | ConvertFrom-Json
           if ($commit.chosen_arm_revealed -ne $false) { throw 'Frozen shadow commitment revealed an arm early' }
           $closure = (& .\dist\ci_route_shadow\SupermixRouteShadow.exe close --registry $registryPath --campaign $campaignId --compact) | ConvertFrom-Json
           if ($closure.closure.frozen_commitment_count -ne 1) { throw 'Frozen shadow closure count mismatch' }

README.md

@@ -15,12 +15,15 @@ It is intentionally a mixed workspace, not a minimal source-only model repo.
 
 ## Current status
 
-As of July 18, 2026:
+As of July 22, 2026:
 
 - `source/` is the active Supermix Studio runtime and packaging tree
 - the curated desktop build selects `11` core model artifacts and leaves expansion to the model store
 - the route control plane includes durable lifecycle evidence, Policy Lab diagnostics,
   bounded-exposure rehearsal, and a fail-closed stateful experiment protocol preflight
+- v51 local inference supports progressive accepted-probe reuse plus a post-head,
+  allowed-label-scoped decision verifier that checks the ordered top-3 boundary before
+  an adaptive early exit
 - `runtime_python/` remains a legacy compatibility snapshot for the smaller chat runtime;
   it is not the source of truth for the multimodel Studio route control plane
 - the Windows installer contract version is `2026.07.18`
@@ -52,6 +55,8 @@ As of July 18, 2026:
 - native-image experimental checkpoints
 - training pipelines for frontier, omni, lite, and specialist model lines
 - benchmark sweeps across common text benchmarks
+- release-gated v51 adaptive compute with source/package parity and frozen-prompt
+  response-fidelity checks
 - export and publishing workflows for GitHub releases and Hugging Face model/dataset repos
 
 ## Quick start
@@ -333,18 +338,44 @@ executed envelope.
 
 ## Research and experiment notes
 
-The v51 stability benchmark supports a research-derived, shadow-only
-distribution-drift diagnostic:
+The accepted v51 decision-fidelity configuration is checkpoint- and
+workload-calibrated: ordered top-3 persistence, a `0.0005` minimum adjacent
+probability gap through the top-3/outside boundary, and exact fallback to the
+trained three-cycle reference output whenever an earlier decision is not
+certified. Legacy latent, entropy, and ACT signals cannot bypass this post-head
+verifier.
+
+The clean CPU release gate at commit `81c4dbe7` evaluated 4,096 held-out
+requests in each of two modes (isolated verifier and exact release-runtime
+defaults):
+
+- zero top-1, ordered top-3, top-3-set, or exact-output disagreements in both
+  modes, with zero per-seed accuracy deltas;
+- 3,941 certified cycle-2 exits and 155 exact cycle-3 reference fallbacks,
+  for 2.0378 mean cycles and a 32.07% cycle reduction from fixed cycle 3;
+- positive counterbalanced latency results in both modes: 4.76% weighted / 3.71%
+  median-per-seed reduction for release runtime and 7.23% / 5.57% for the
+  isolated verifier;
+- the frozen 16-prompt source/package response gate had zero response,
+  top-five-order, runtime-contract, or packaged-behavior mismatches; and
+- progressive accepted-probe auto compute matched the legacy controller exactly
+  on 256 requests while reducing forward evaluations by 31.25% and weighted
+  latency by 30.09%.
 
 ```powershell
-python source/benchmark_v51_prediction_stability.py --distribution_top_k 5
+python source/run_v51_prediction_stability_gate.py --device cpu --torch-num-threads 8 --torch-interop-threads 1 --strict-determinism --samples-per-seed 512 --enforce-gates
+python source/run_v51_chat_response_fidelity_gate.py --device cpu --torch-num-threads 8 --torch-interop-threads 1 --strict-determinism --enforce-gates
+python source/benchmark_progressive_auto_compute.py --device cpu --torch-num-threads 8 --torch-interop-threads 1 --strict-determinism --enforce-gates
 ```
 
-It reports consecutive-prefix top-k Jensen-Shannon divergence with a retained
-`other` mass bucket. The metric is telemetry only and does not authorize an
-early exit. A 480-request CPU pilot retained 96/96 agreement for the current
-patience-2/tolerance-0.005 setting with 28.8% fewer cycles; the release gate is a
-larger fresh-seed confirmation, not that pilot result.
+The stricter gate rejected the earlier `0.0001` candidate: release mode showed
+5 top-1, 18 ordered-top-3, and 10 top-3-set disagreements; isolated mode showed
+2, 6, and 4. That failure led to the common exit guard, ordered-rank state, the
+calibrated margin, bounded disagreement evidence, and reference-budget fallback.
+Top-k Jensen-Shannon divergence remains shadow telemetry only and cannot
+authorize an exit. These are deterministic results for this checkpoint,
+synthetic task, seeds, and frozen prompt matrix—not a universal reasoning or
+chat-quality guarantee.
 
 This repo is a living experiment workspace. It contains finished artifacts, release-ready packaging, and in-progress work at the same time.
 

chat_web_app.py

@@ -36,7 +36,10 @@
 # Keep the root compatibility surface behaviorally identical for callers that
 # validate CLI default inheritance after another test imported this wrapper
 # first. The canonical helper remains private in the source implementation.
-for _name in ("_runtime_compute_cli_overrides",):
+for _name in (
+    "_normalize_runtime_compute_defaults",
+    "_runtime_compute_cli_overrides",
+):
     globals()[_name] = getattr(_IMPL, _name)
 
 

launch_v51_ultra_chat_demo.ps1

@@ -59,7 +59,9 @@ try {
         "--adaptive_compute",
         "--adaptive_exit_entropy", "0.2",
         "--prediction_stability_patience", "2",
-        "--prediction_stability_tol", "0.005"
+        "--prediction_stability_tol", "0.005",
+        "--prediction_stability_margin", "0.0005",
+        "--prediction_stability_rank_depth", "3"
     ) | ForEach-Object { Quote-ProcessArg $_ }
     $process = Start-Process -FilePath "python" -ArgumentList $args -WorkingDirectory $Root -WindowStyle Hidden -PassThru -RedirectStandardOutput $Stdout -RedirectStandardError $Stderr
     Set-Content -LiteralPath $PidFile -Value $process.Id