Add frozen-model Gooo feedback pilot evidence; preserve weights and v5 core

research/feedback-path-pilot-20261001/README.md

@@ -0,0 +1,66 @@
+# Frozen-model feedback judgment pilot
+
+This appendix reuses the existing three tiny structural models. No new weights,
+optimizer steps, GPU calls or upstream Laya calls occurred. The Go research
+session explicitly asks the original frozen model after a partial batch using
+observed finite failures and actual native-compiler CI context.
+
+The native compiler's post-main CI is source `307159f041644a3aa56dfd325c345f5325aec902`,
+run `36801794319`, with six passed checks and Go-verified proof/receipt. This
+context is a judgment hint; it is not authority to change source or proof that
+the newer research runner itself passed CI.
+
+## Measured result
+
+| Observation | Result |
+| --- | --- |
+| Policies / paired comparisons | 32 / 16 |
+| Different compound intentions | 1 |
+| Same emitted Go / same finite results | 16 / 16 pairs |
+| Actual local predictions | 246, including 102 feedback predictions |
+| Feedback rounds | 17 |
+| Candidate attempts | 1,268 |
+| Finite selection cases | 208/224 across repeated policy views |
+| Separate input evaluation | 288/288 observations of 9 existing distinct inputs |
+| Native codegen calls | 32 |
+| Distinct generated-Go/input-vector executions | 2, reused by byte/input hash |
+| Pairs with fewer / more / equal candidates | 1 / 2 / 13 |
+| Model-policy median search wall time | 1.888ms baseline; 2.222ms feedback |
+
+Complete contracts satisfy 7/7 stated cases. Deliberately inconsistent contracts
+retain 6/7 and exhaust all 64 paths; reconsideration cannot manufacture a valid
+answer for the conflicting expected value. Korean QAT completion used 19→17
+candidates; Korean FP32 used 10→11 and PTQ used 17→18. No aggregate functional
+improvement occurred. The original model was not trained to interpret this
+failure summary. Extra calls and worse observations remain visible.
+
+All policies used the same compound intention, two language views, four model
+arms including disconnected controls and two contracts. Baseline runs came
+first, there was one ordering and no repetitions. Search timings exclude model
+load, native generation, Go compilation and process startup. CPU/RSS and host
+utilization changes were not measured in this pilot. These data do not establish
+general natural-language accuracy, a causal speedup, 32 new experiment ideas or
+288 independent language tasks. No-model controls reproduce their progress and
+source exactly despite the enabled control flag; they make zero model calls.
+
+## Reproduction and preserved failures
+
+The public GitHub repository is
+[gooo-neural-decision-experiments](https://github.com/kimjooyoon/gooo-neural-decision-experiments).
+Runner source is `63c9967505355ccc5c56780f901f4ce583008f17`;
+full captures are in `runs/feedback-path-pilot-fixed-20261001`.
+The audit re-interprets all 1,268 candidate bodies and checks progress, feedback,
+native and executed-Go captures, with zero new model or native calls.
+
+The earlier `runs/feedback-path-pilot-20261001` attempt is preserved. It made
+zero model predictions and one rejected native call because the runner nested
+a complete Gooo module inside another body. The compiler remained unchanged;
+the repaired runner passes the complete module directly. A separate wrong
+revision preflight was rejected before any model or native call.
+
+This directory is a separately allowlisted **evidence appendix**. The previous
+v5 core manifest, model card and nine weight files remain unchanged. Raw captures
+and implementation live on GitHub. The new appendix includes only synthetic
+intentions and observations; credentials, private repositories and host paths
+are excluded. First-shot accuracy and perfect expression are not prerequisites
+for publishing an observed partial result.

research/feedback-path-pilot-20261001/audit.json

@@ -0,0 +1,30 @@
+{
+  "actual_model_predictions": 246,
+  "candidate_attempts": 1268,
+  "candidate_count_increased_pairs": 2,
+  "candidate_count_reduced_pairs": 1,
+  "candidate_count_unchanged_pairs": 13,
+  "cpu_or_rss_measured": false,
+  "decision": "PASS",
+  "distinct_evaluation_inputs": 9,
+  "executed_disjoint_input_observations": 288,
+  "executed_disjoint_input_passed": 288,
+  "executed_finite_cases": 224,
+  "executed_finite_passed": 208,
+  "feedback_model_predictions": 102,
+  "feedback_rounds": 17,
+  "independently_interpreted_candidate_bodies": 1268,
+  "model_policy_median_base_search_ms": 1.887625,
+  "model_policy_median_feedback_search_ms": 2.222083,
+  "native_calls": 32,
+  "new_native_calls_by_audit": 0,
+  "new_predictions_by_audit": 0,
+  "pairs": 16,
+  "policies": 32,
+  "report_sha256": "050f288116b973aeb5198d35fcf9c085b178189868243eceb67a86cb0d32048e",
+  "same_emitted_go_pairs": 16,
+  "same_finite_and_evaluation_pairs": 16,
+  "schema": "gooo/feedback-path-pilot-audit/v1",
+  "scope": "Fixed one-intent pilot, one baseline-first ordering and no repetitions. Finite selection cases and existing disjoint function inputs are separate; repeated observations are not 288 independent tasks. No generic NL improvement or causal speedup claim. Audit re-interprets candidates and hashes captured native/Go outputs; it does not re-execute native or generated Go or retrain the model.",
+  "unique_generated_go_input_executions": 2
+}

research/feedback-path-pilot-20261001/preexecution.json

@@ -0,0 +1,29 @@
+{
+  "attempt_step": 8,
+  "ci_hint": {
+    "source_sha": "307159f041644a3aa56dfd325c345f5325aec902",
+    "status": "PASS"
+  },
+  "ci_hint_publication_sha256": "4710e8ca79b29596b817b5da5fc202004b341f3992de7787466630397d257870",
+  "compound_intents": 1,
+  "evaluation_inputs": 9,
+  "maximum_candidates": 64,
+  "maximum_feedback_rounds": 2,
+  "model_metadata": {
+    "fp32": "1ea3bada068f2487418f17270db4ba6785a98c3ad60e0ad7eb92359400a01bb5",
+    "ptq_ternary": "7c4eb4068d76e83620a9b8e7ce7f8d948b5e2436d44c7b79cf241da609dab894",
+    "qat_ternary": "368e37e7899cecba03874b69e933525a2f8c90ecc53d787d8e698de6aa21c087"
+  },
+  "model_weights": {
+    "fp32": "3732a103720fdb0da0ea611471ecde7472e0b9166322394ba9352ffb855b9354",
+    "ptq_ternary": "01a6574dedba327361cd3e4fda00a4d4dea34e162aab09885174bcd41ed4da47",
+    "qat_ternary": "456c686911b95bb933a7dd9275d2611974482ecf353831ab17b593c6bd3d86de"
+  },
+  "native_revision": "307159f041644a3aa56dfd325c345f5325aec902",
+  "optimizer_steps": 0,
+  "paired_comparisons": 16,
+  "policies": 32,
+  "runner_revision": "63c9967505355ccc5c56780f901f4ce583008f17",
+  "schema": "gooo/feedback-path-pilot-preexecution/v1",
+  "scope": "Same frozen compound intent, two languages, complete and inconsistent contracts, three models and offline controls; neither new independent ideas nor a feedback-trained model. CI hint is previously verified snapshot context, not current study authorization."
+}

research/feedback-path-pilot-20261001/publication-manifest.json

@@ -0,0 +1,33 @@
+{
+  "schema": "gooo/public-feedback-pilot-allowlist/v1",
+  "files": [
+    {
+      "path": "README.md",
+      "sha256": "21fe36be9e1e3018c38dee0b3093e9409683482449cf32e0cdfd75f0f0d506ff",
+      "bytes": 3709
+    },
+    {
+      "path": "audit.json",
+      "sha256": "4574d172b38fa563b45a1b8d5138c0e4007e67013764fa59ce5b9cc920187d56",
+      "bytes": 1428
+    },
+    {
+      "path": "preexecution.json",
+      "sha256": "966b93ebc6aca754b4174ebfbb931e903c1c2489e18629884208e29cf9427a03",
+      "bytes": 1425
+    },
+    {
+      "path": "report.json",
+      "sha256": "050f288116b973aeb5198d35fcf9c085b178189868243eceb67a86cb0d32048e",
+      "bytes": 4089826
+    },
+    {
+      "path": "source-input-pins.json",
+      "sha256": "b93820f4a052975ba470cb6df8cea70c28bce2af91b594f411b143ad937c7d6f",
+      "bytes": 499
+    }
+  ],
+  "unchanged_core_manifest_sha256": "a320117c9d8567cbf0f15df0fb92fc0245c84787a49e1dff9f4cb775117894e2",
+  "weights_changed": false,
+  "scope": "Separate evidence appendix to unchanged v5 weights and core files. One fixed compound intent; not model training or general language accuracy. Raw native/Go captures and source live in the public GitHub research repository."
+}

research/feedback-path-pilot-20261001/source-input-pins.json

@@ -0,0 +1,12 @@
+{
+  "binaries": {
+    "go": "132b69336a1f809932a8a20b0201dbbb980e86e3a323ae32e893639d83d71598",
+    "native": "43c1d565d1d1ebe39793dfc553c2eab02ebce57ebf5b054b68f7554d8c4f5304"
+  },
+  "fixtures": {
+    "en": "eeced46f7d08552d40785e7f1c6559698cf067ce7d3b9af6b0b65dac1aec0200",
+    "evaluation": "152b36af8d89fbf481fe3d5c441bd9b2ba5f32846a351ff5f98414d95caa2f9b",
+    "ko": "caa29e39238cf1876d671c56d06c2ad47cf8614dc054cdd80970ca6f7db2c548"
+  },
+  "schema": "gooo/feedback-path-pilot-input-pins/v1"
+}