Publish bounded bilingual Gooo judgment arms and negative evidence

research/bilingual-gooo-judgment-20261001/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Gooo contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

research/bilingual-gooo-judgment-20261001/README.md

@@ -0,0 +1,98 @@
+# Bilingual Gooo judgment with bounded continuation
+
+This experimental appendix keeps both trained arms, including regressions.
+First-shot accuracy is an observation; useful partial construction and finite
+continuation are the development target. These are independently initialized own
+tiny-model descendants, not fine-tuned upstream Laya weights.
+
+## Actual work and scope
+
+The source-bound generator pairs existing English/Korean Gooo views: 1,040 pairs
+reuse 640 program groups, with 800 training, 80 calibration and 160 development
+evaluation pairs. Zero new independent intentions. The entire original curriculum
+and raw captures are in the
+[public research repository](https://github.com/kimjooyoon/gooo-neural-decision-experiments).
+Gooo structure, legal options, source groups and finite soft targets remain bound
+to their original digests. Natural-language meaning and finite test agreement are
+separate labels; PROV-O vocabulary in the source is not ontology reasoning.
+
+Two arms start from the same own feedback FP32 parent and use the same seed.
+Control minimizes finite soft-target NLL; paired adds 0.25 times bilingual
+Jensen-Shannon divergence. Actual MPS training performed 224 optimizer steps
+across FP32 and QAT in both arms. PTQ reuses each FP32. The four training loops
+totaled about 1.52 seconds; the first loop includes startup and was 1.29 seconds,
+while the other loops were about 0.07–0.08 seconds. Fixed arm order and a single
+run do not establish comparative training throughput. Sampled MPS allocated
+memory peaked at 2,033,920 bytes; driver memory at 53,166,080 bytes. Python process
+lifetime peak RSS was about 504 MiB. These are not host utilization measurements.
+
+## Judgment and finite continuation
+
+Go measured 160 English/Korean pairs (320 views) in each cell. Pair agreement
+alone can be wrong, as the disconnected control illustrates.
+
+| Model | Initial agreement /160 | Same wrong intent /160 | Complete views before /320 | Added path attempts |
+|---|---:|---:|---:|---:|
+| Own parent FP32 | 91 | 4 | 243 | 77 |
+| Control FP32 | 97 | 4 | 249 | 71 |
+| Paired FP32 | 96 | 4 | 248 | 72 |
+| Own parent PTQ | 99 | 1 | 257 | 63 |
+| Control PTQ | 108 | 14 | 240 | 80 |
+| Paired PTQ | 108 | 14 | 240 | 80 |
+| Own parent QAT | 90 | 2 | 246 | 74 |
+| Control QAT | 95 | 15 | 225 | 95 |
+| Paired QAT | 95 | 15 | 225 | 95 |
+| Disconnected deterministic | 160 | 80 | 160 | 160 |
+
+The paired term did not improve this development comparison. Both new ternary
+arms increase same-wrong agreement and require more attempts than their own
+parents. This evidence is retained; no global default model promotion is made.
+
+After a separate complete contract from the authored arithmetic oracle, every
+cell reaches 320/320 complete views and 3,840/3,840 finite cases with at most one
+additional legal path per view, zero additional model calls. This is exhaustive
+two-path typed-arena evaluation, not universal intent correctness, arbitrary
+code synthesis or native execution. The Go judgment capture performed 2,880
+initial model calls. Separate full-curriculum audits in both new arms each made
+5,856 calls including numerical parity and reused parent comparisons. Six new
+exports pass Go/Python parity with maximum observed error about 1.55e-6.
+
+Runtime prediction medians were about 10.1–10.2 microseconds for FP32 and
+11.7–11.8 microseconds for ternary in this capture. FP32 tensor storage is 50,912
+bytes. Ternary packed weights are 2,759 bytes on disk, decoded int8 tensors are
+12,896 bytes resident, matrix scales 8 bytes and inference workspace 1,248 bytes.
+Five base-three digits per byte give 1.6 disk bits per matrix weight; theoretical
+log2(3) is about 1.585. Neither is total process RAM or packed compute speed.
+
+## Actual adopted-main dogfood
+
+The clean adopted compiler main `1e01c96c54f2f8dd43334b8f580af93ffaea24df`
+uses Go 1.27.1 and SDK 0.2.8. Main push CI has six canonical successful checks and
+an independently verified proof, separately preserved in this appendix.
+
+Twenty-eight actual native calls reuse one subtraction intention in English and
+Korean, six trained variants and disconnected control, sparse/full contracts.
+Twenty-four actual model predictions. Six sparse calls choose the wrong intended
+operand order while passing the one input where both paths return zero. All
+fourteen full-contract calls select the intended body, using six extra candidate
+attempts. Full contracts are separate calls with one new ranking each, not
+zero-call native continuation. Actual emitted Go execution ran the two distinct
+selected bodies on inputs 2 and 3: two processes and four function invocations.
+An independent offline audit rechecks hashes, contracts, selected function
+structure and arithmetic values without inference or subprocesses.
+
+Across those 28 native child calls, observed wall time was 5.43–13.81 ms
+(median 5.76 ms), CPU time median 4.93 ms, peak child RSS roughly 15.7–16.3 MiB.
+These whole compiler processes are distinct from model-stage latency and tensor
+memory. No causal speedup, host CPU utilization change, upstream Laya use or
+production release-readiness is claimed.
+
+## Use
+
+Both arms expose the existing Go typed-path model ABI. From the research checkout,
+use `--path-model runs/bilingual-judgment-mps-20261001/paired/models/fp32/model.json`
+with the native compiler body-codegen command and an explicit typed path plan.
+Disconnecting the model keeps the deterministic path policy. Always retain the
+source/test contracts and partial denominators; repeated calls and translations
+are not independent experiments. Model confidence is a ranking hint, not an
+authority to redefine expected values.

research/bilingual-gooo-judgment-20261001/data/manifest.json

@@ -0,0 +1,22 @@
+{
+  "dataset_sha256": "570d1d73bfea32662b869e5cfbf77e6d04df838b703191b076be58c1d2180dbf",
+  "languages_in_order": [
+    "en",
+    "ko"
+  ],
+  "new_independent_intentions": 0,
+  "pair_count": 1040,
+  "pairs_sha256": "25e090b1c4d842b34cf7a1d3a994b7b6e4be70cb59389fe33420e5d8750574ce",
+  "reused_program_groups": 640,
+  "schema": "gooo/bilingual-gooo-pairs/v1",
+  "scope": "Existing synthetic adjacent templates, same Gooo fallback structure and finite soft target. Pair equality is not correctness. Development partitions reused; no new independent experiments or untouched holdout.",
+  "source_sha256": {
+    "cmd/bilingual-pairs/main.go": "646784b3f7f82b7887aad7707acd0ddc8c79c74f83492f26e299af4886300343",
+    "internal/bilingualstudy/pairs.go": "d80edd2acab39b1e5d5bbdf1f7a9094e6ccf55c4dad739049a5d5569f43d6f7e"
+  },
+  "split_pairs": {
+    "calibration": 80,
+    "test": 160,
+    "train": 800
+  }
+}

research/bilingual-gooo-judgment-20261001/judgment/preexecution.json

@@ -0,0 +1,14 @@
+{
+  "complete_contract": "pathstudy.Inputs plus independent pathstudy.Oracle; explicit intention reference",
+  "dataset_sha256": "570d1d73bfea32662b869e5cfbf77e6d04df838b703191b076be58c1d2180dbf",
+  "development_pairs": 160,
+  "development_views": 320,
+  "model_arms": 9,
+  "schema": "gooo/bilingual-judgment-go-preexecution/v1",
+  "scope": "Sparse Gooo finite views followed by a separately bound authored full contract, two legal paths, at most one additional attempt, zero additional model calls. No native/generated-Go execution, host utilization, untouched holdout or new independent tasks.",
+  "source_hashes": {
+    "internal/bilingualstudy/pairs.go": "d80edd2acab39b1e5d5bbdf1f7a9094e6ccf55c4dad739049a5d5569f43d6f7e",
+    "internal/pathstudy/study.go": "5708818e0ee2d093fcd755573fc3c9a66eac276feccbaeff3441cdbd0aebade6",
+    "tools/audit-bilingual-judgment/main.go": "744083bfa06a7d2ad7cb1413ed3800d9b0afe147739827143e2aa213e9b81673"
+  }
+}

research/bilingual-gooo-judgment-20261001/judgment/report.json

@@ -0,0 +1,326 @@
+{
+  "actual_initial_model_calls": 2880,
+  "additional_model_calls": 0,
+  "agreement_is_correctness": false,
+  "capture_sha256": {
+    "control-fp32.json": "cc006fce86c2467e363acd6ce84b6a7378ee9eab5eb65b8c2224830a5fe96152",
+    "control-ptq_ternary.json": "92dd02f5fba9704653d099874ca92165442ea8d62e2c973bd5ddb53b41238fae",
+    "control-qat_ternary.json": "482b9428b6d31ae7e895a38af547a7531135dcd38119496a621393a26497634e",
+    "offline-deterministic.json": "58e0f3fa68a1b76770d64bd9df617a357e77dd10d6cb7133d662917ff2be5924",
+    "paired-fp32.json": "a5c957c82e403f783a19a4070084e360d7b4333132787d01670667edb9ab84fb",
+    "paired-ptq_ternary.json": "edd885f9d5a0df63108b1a3dcd0991cb6d926001de49d1965e6dd45948b06aa8",
+    "paired-qat_ternary.json": "9bee3301e9628cdabfbbdf6622615cae89174ebcf9398e88ea6534e34bcf5780",
+    "parent-fp32.json": "bbbb285bd3732864add43818d628d989360105813e78758e8ca0951fe04b42de",
+    "parent-ptq_ternary.json": "b60d7302ebbfd0c71fca327450cc329750c0b17211e1aede6887ea990f0d5452",
+    "parent-qat_ternary.json": "4fd9051b6aad4ffe27ee6f6675f9a8293bf2fc1e4d72cb1bfeb7efecc8fcb4d7"
+  },
+  "cells": {
+    "control-fp32": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 71,
+        "both_intention_correct_pairs": 93,
+        "finite_best_set_selected": 261,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 3001,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 249,
+        "initial_pair_agreement": 97,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 4,
+        "sparse_finite_passed": 261,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 0,
+      "mean_bilingual_js": 0.08167314616752759,
+      "metadata_sha256": "d19d98f6a4f2fc55524332a79db4e65ce019e5172f0e891d633e896c8f0115e6",
+      "native_calls": 0,
+      "packed_weights_bytes": 50912,
+      "prediction_median_ns": 10125,
+      "prediction_p95_ns": 10666,
+      "resident_tensor_bytes": 50912,
+      "weights_sha256": "e411e0f6bad0087b040a146d0a1030c272196d19df078aba87c38bd11e10db53",
+      "workspace_bytes": 1248
+    },
+    "control-ptq_ternary": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 80,
+        "both_intention_correct_pairs": 94,
+        "finite_best_set_selected": 254,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 2895,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 240,
+        "initial_pair_agreement": 108,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 14,
+        "sparse_finite_passed": 254,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "mean_bilingual_js": 0.021124628703467086,
+      "metadata_sha256": "88b2cdda4fd9bd9bcd12caea597eaf3ac407fb63915dd4def1621960640b8ec3",
+      "native_calls": 0,
+      "packed_weights_bytes": 2759,
+      "prediction_median_ns": 11792,
+      "prediction_p95_ns": 12250,
+      "resident_tensor_bytes": 12896,
+      "weights_sha256": "030dd5400202f0522a70217e5f16f2dcfc9cb447f899c94582d5688efae86b5e",
+      "workspace_bytes": 1248
+    },
+    "control-qat_ternary": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 95,
+        "both_intention_correct_pairs": 80,
+        "finite_best_set_selected": 244,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 2720,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 225,
+        "initial_pair_agreement": 95,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 15,
+        "sparse_finite_passed": 244,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "mean_bilingual_js": 0.05270309999302516,
+      "metadata_sha256": "91e4ac583e3e078e03d1ac88c303574c469e7683560c3dd74d736c846f6736d7",
+      "native_calls": 0,
+      "packed_weights_bytes": 2759,
+      "prediction_median_ns": 11750,
+      "prediction_p95_ns": 12042,
+      "resident_tensor_bytes": 12896,
+      "weights_sha256": "00525b0674ade3f95b85b6e620f8ea0b376c0231734239a27df1c0d5c7f710eb",
+      "workspace_bytes": 1248
+    },
+    "offline-deterministic": {
+      "counts": {
+        "added_path_attempts": 160,
+        "both_intention_correct_pairs": 80,
+        "finite_best_set_selected": 198,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 1960,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 160,
+        "initial_pair_agreement": 160,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 80,
+        "sparse_finite_passed": 198,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "mean_bilingual_js": 0,
+      "native_calls": 0
+    },
+    "paired-fp32": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 72,
+        "both_intention_correct_pairs": 92,
+        "finite_best_set_selected": 261,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 2990,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 248,
+        "initial_pair_agreement": 96,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 4,
+        "sparse_finite_passed": 261,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 0,
+      "mean_bilingual_js": 0.08171740071912106,
+      "metadata_sha256": "ccb48c6fd0f76458a231d0bd1266e88252708aa6f64093742aead6a5ede27dc3",
+      "native_calls": 0,
+      "packed_weights_bytes": 50912,
+      "prediction_median_ns": 10167,
+      "prediction_p95_ns": 10625,
+      "resident_tensor_bytes": 50912,
+      "weights_sha256": "1dad702576db50fdb08c62f07728eb06e81c6d8a3011b78ce38b0338406aacf4",
+      "workspace_bytes": 1248
+    },
+    "paired-ptq_ternary": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 80,
+        "both_intention_correct_pairs": 94,
+        "finite_best_set_selected": 254,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 2895,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 240,
+        "initial_pair_agreement": 108,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 14,
+        "sparse_finite_passed": 254,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "mean_bilingual_js": 0.021126226344055497,
+      "metadata_sha256": "5233de7a91a997358de890c6431cf0fa2fd6fcd4d3dfb10a8cd851349028b6c9",
+      "native_calls": 0,
+      "packed_weights_bytes": 2759,
+      "prediction_median_ns": 11791,
+      "prediction_p95_ns": 12125,
+      "resident_tensor_bytes": 12896,
+      "weights_sha256": "d6fb8f32a29a674b5ee904f2300f51b9a0d610f57ec0ebcbc52635f73e45d619",
+      "workspace_bytes": 1248
+    },
+    "paired-qat_ternary": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 95,
+        "both_intention_correct_pairs": 80,
+        "finite_best_set_selected": 244,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 2720,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 225,
+        "initial_pair_agreement": 95,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 15,
+        "sparse_finite_passed": 244,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "mean_bilingual_js": 0.05478020516016365,
+      "metadata_sha256": "47524ff50c41b5bde0be95a7447126ee20abfae59f2e2fdc1e3f78c42c5613d0",
+      "native_calls": 0,
+      "packed_weights_bytes": 2759,
+      "prediction_median_ns": 11750,
+      "prediction_p95_ns": 12125,
+      "resident_tensor_bytes": 12896,
+      "weights_sha256": "1673c7cf839bf8d7fe317495c5a118d67430ae3247afcc7eda2844dd9a9c11f4",
+      "workspace_bytes": 1248
+    },
+    "parent-fp32": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 77,
+        "both_intention_correct_pairs": 87,
+        "finite_best_set_selected": 261,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 2935,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 243,
+        "initial_pair_agreement": 91,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 4,
+        "sparse_finite_passed": 261,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 0,
+      "mean_bilingual_js": 0.08810483189407463,
+      "metadata_sha256": "47bd3ed2037c8ba0af31ec4cad3a47fe1182af171845406aba01c5107f4b24b7",
+      "native_calls": 0,
+      "packed_weights_bytes": 50912,
+      "prediction_median_ns": 10167,
+      "prediction_p95_ns": 10750,
+      "resident_tensor_bytes": 50912,
+      "weights_sha256": "73eedef1d60f49129190683a41cbf82229c1e3603fd44cdd17792d015f499954",
+      "workspace_bytes": 1248
+    },
+    "parent-ptq_ternary": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 63,
+        "both_intention_correct_pairs": 98,
+        "finite_best_set_selected": 273,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 3101,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 257,
+        "initial_pair_agreement": 99,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 1,
+        "sparse_finite_passed": 273,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "mean_bilingual_js": 0.022822061488925204,
+      "metadata_sha256": "a7b66676af2eb5a7e6952f3c65c812f231865ac4974a3c3c1f329b125e55124c",
+      "native_calls": 0,
+      "packed_weights_bytes": 2759,
+      "prediction_median_ns": 11791,
+      "prediction_p95_ns": 12125,
+      "resident_tensor_bytes": 12896,
+      "weights_sha256": "c0ee9fc41c16030786103cbf771afe40ba6b7866c25095214d062daa6dcc47a0",
+      "workspace_bytes": 1248
+    },
+    "parent-qat_ternary": {
+      "counts": {
+        "actual_model_calls": 320,
+        "added_path_attempts": 74,
+        "both_intention_correct_pairs": 88,
+        "finite_best_set_selected": 265,
+        "full_finite_passed_after": 3840,
+        "full_finite_passed_before": 2972,
+        "full_finite_total": 3840,
+        "full_program_views_after": 320,
+        "full_program_views_before": 246,
+        "initial_pair_agreement": 90,
+        "pairs": 160,
+        "same_wrong_intention_pairs": 2,
+        "sparse_finite_passed": 265,
+        "sparse_finite_total": 320,
+        "views": 320
+      },
+      "extra_model_calls_during_continuation": 0,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "mean_bilingual_js": 0.10470378599577415,
+      "metadata_sha256": "17dd98b2c05c7a9c203523a2e8e70ba0cbb523ae47715fdc0dc42acef81b0536",
+      "native_calls": 0,
+      "packed_weights_bytes": 2759,
+      "prediction_median_ns": 11791,
+      "prediction_p95_ns": 12083,
+      "resident_tensor_bytes": 12896,
+      "weights_sha256": "8dc9a0d86b61733e6b09ea0cf85f4b39abf7a2383a2852c1149bd30ac726ad41",
+      "workspace_bytes": 1248
+    }
+  },
+  "decision": "PASS",
+  "development_pairs": 160,
+  "first_shot_correctness_is_primary_objective": false,
+  "new_independent_intentions": 0,
+  "schema": "gooo/bilingual-judgment-go-report/v1",
+  "scope": "Sparse Gooo finite views followed by a separately bound authored full contract, two legal paths, at most one additional attempt, zero additional model calls. No native/generated-Go execution, host utilization, untouched holdout or new independent tasks."
+}

research/bilingual-gooo-judgment-20261001/main-ci/finalization.json

@@ -0,0 +1,148 @@
+{
+  "schema": "gooo/native-path-diagnosis-main-push-finalization/v1",
+  "decision": "PASS",
+  "native_main": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+  "run_id": 36848147619,
+  "run_attempt": 1,
+  "ci": {
+    "conclusion": "success",
+    "event": "push",
+    "headSha": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+    "jobs": [
+      {
+        "conclusion": "success",
+        "name": "Feedback predecessor",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "Language package execution",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "Semantic conformance",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "go test",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "gofmt",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "go test -race",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "go vet",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "COHERENCE boundary probe",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "CI policy",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "CI evidence",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "CI proof bundle",
+        "status": "completed"
+      },
+      {
+        "conclusion": "success",
+        "name": "CI failure summary",
+        "status": "completed"
+      }
+    ],
+    "status": "completed",
+    "updatedAt": "2026-10-01T10:33:16Z"
+  },
+  "canonical_checks": [
+    {
+      "id": 110323235427,
+      "name": "gofmt",
+      "status": "completed",
+      "conclusion": "success",
+      "head_sha": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+      "run_id": 36848147619,
+      "run_attempt": 1
+    },
+    {
+      "id": 110323235518,
+      "name": "go vet",
+      "status": "completed",
+      "conclusion": "success",
+      "head_sha": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+      "run_id": 36848147619,
+      "run_attempt": 1
+    },
+    {
+      "id": 110323235359,
+      "name": "go test",
+      "status": "completed",
+      "conclusion": "success",
+      "head_sha": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+      "run_id": 36848147619,
+      "run_attempt": 1
+    },
+    {
+      "id": 110323235495,
+      "name": "go test -race",
+      "status": "completed",
+      "conclusion": "success",
+      "head_sha": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+      "run_id": 36848147619,
+      "run_attempt": 1
+    },
+    {
+      "id": 110323235342,
+      "name": "Semantic conformance",
+      "status": "completed",
+      "conclusion": "success",
+      "head_sha": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+      "run_id": 36848147619,
+      "run_attempt": 1
+    },
+    {
+      "id": 110326559919,
+      "name": "CI policy",
+      "status": "completed",
+      "conclusion": "success",
+      "head_sha": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+      "run_id": 36848147619,
+      "run_attempt": 1
+    }
+  ],
+  "proof": {
+    "schema": "gooo/ci-proof/v5",
+    "decision": "PASS",
+    "head_sha": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+    "checkout_ref": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+    "event": "push",
+    "event_ref": "refs/heads/main",
+    "independent_go_verifier_exit": 0,
+    "sha256": "7b5209db2812c56f1c7ae9e48f67c5f6cc785d3e463ee5bfc7ee47a513319b37",
+    "provenance_receipt_sha256": "c424daf98d9ac1a0e33a45eec3d39d27961b8f41f5785e1dcfc7882ac9742c45",
+    "artifact_id": 11154688924,
+    "artifact_zip_sha256": "c58a2836010837cff5888376aee59b2b52851d090f629788d54f3380be94bfcb",
+    "artifact_zip_bytes": 10946
+  },
+  "prior_adoption_receipt": "publication/native-path-diagnosis-main-adoption-20261001.json",
+  "scope": "Later terminal main-push observation. The earlier frozen pending receipt and HF revision are preserved. Six canonical checks and all twelve jobs succeeded, downloaded proof independently verified, live main matches. No new inference, training, generated Go execution or release-readiness claim in this finalization."
+}

research/bilingual-gooo-judgment-20261001/models/control/fp32/model.json

@@ -0,0 +1,64 @@
+{
+  "confidence_threshold": 1.0,
+  "feature_dim": 256,
+  "feature_version": "positioned_intent_ngrams_v1",
+  "hidden_dim": 48,
+  "labels": [
+    "reference_first",
+    "reference_second",
+    "assign_first",
+    "assign_second",
+    "layout_forward",
+    "layout_reverse",
+    "schedule_forward",
+    "schedule_reverse"
+  ],
+  "max_bytes": 512,
+  "schema": "gooo/tiny-path-decision-model/v1",
+  "temperature": 1.0,
+  "tensors": [
+    {
+      "bytes": 49152,
+      "cols": 256,
+      "count": 12288,
+      "encoding": "float32_le",
+      "name": "w1",
+      "offset": 0,
+      "rows": 48,
+      "scale": 1.0
+    },
+    {
+      "bytes": 192,
+      "cols": 48,
+      "count": 48,
+      "encoding": "float32_le",
+      "name": "b1",
+      "offset": 49152,
+      "rows": 1,
+      "scale": 1.0
+    },
+    {
+      "bytes": 1536,
+      "cols": 48,
+      "count": 384,
+      "encoding": "float32_le",
+      "name": "w2",
+      "offset": 49344,
+      "rows": 8,
+      "scale": 1.0
+    },
+    {
+      "bytes": 32,
+      "cols": 8,
+      "count": 8,
+      "encoding": "float32_le",
+      "name": "b2",
+      "offset": 50880,
+      "rows": 1,
+      "scale": 1.0
+    }
+  ],
+  "variant": "fp32",
+  "weights_file": "weights.bin",
+  "weights_sha256": "e411e0f6bad0087b040a146d0a1030c272196d19df078aba87c38bd11e10db53"
+}

research/bilingual-gooo-judgment-20261001/models/control/fp32/weights.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e411e0f6bad0087b040a146d0a1030c272196d19df078aba87c38bd11e10db53
+size 50912

research/bilingual-gooo-judgment-20261001/models/control/ptq_ternary/model.json

@@ -0,0 +1,64 @@
+{
+  "confidence_threshold": 1.0,
+  "feature_dim": 256,
+  "feature_version": "positioned_intent_ngrams_v1",
+  "hidden_dim": 48,
+  "labels": [
+    "reference_first",
+    "reference_second",
+    "assign_first",
+    "assign_second",
+    "layout_forward",
+    "layout_reverse",
+    "schedule_forward",
+    "schedule_reverse"
+  ],
+  "max_bytes": 512,
+  "schema": "gooo/tiny-path-decision-model/v1",
+  "temperature": 1.0,
+  "tensors": [
+    {
+      "bytes": 2458,
+      "cols": 256,
+      "count": 12288,
+      "encoding": "ternary_base3_5",
+      "name": "w1",
+      "offset": 0,
+      "rows": 48,
+      "scale": 0.19280461966991425
+    },
+    {
+      "bytes": 192,
+      "cols": 48,
+      "count": 48,
+      "encoding": "float32_le",
+      "name": "b1",
+      "offset": 2458,
+      "rows": 1,
+      "scale": 1.0
+    },
+    {
+      "bytes": 77,
+      "cols": 48,
+      "count": 384,
+      "encoding": "ternary_base3_5",
+      "name": "w2",
+      "offset": 2650,
+      "rows": 8,
+      "scale": 0.39537540078163147
+    },
+    {
+      "bytes": 32,
+      "cols": 8,
+      "count": 8,
+      "encoding": "float32_le",
+      "name": "b2",
+      "offset": 2727,
+      "rows": 1,
+      "scale": 1.0
+    }
+  ],
+  "variant": "ptq_ternary",
+  "weights_file": "weights.bin",
+  "weights_sha256": "030dd5400202f0522a70217e5f16f2dcfc9cb447f899c94582d5688efae86b5e"
+}

research/bilingual-gooo-judgment-20261001/models/control/ptq_ternary/weights.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:030dd5400202f0522a70217e5f16f2dcfc9cb447f899c94582d5688efae86b5e
+size 2759

research/bilingual-gooo-judgment-20261001/models/control/qat_ternary/model.json

@@ -0,0 +1,64 @@
+{
+  "confidence_threshold": 1.0,
+  "feature_dim": 256,
+  "feature_version": "positioned_intent_ngrams_v1",
+  "hidden_dim": 48,
+  "labels": [
+    "reference_first",
+    "reference_second",
+    "assign_first",
+    "assign_second",
+    "layout_forward",
+    "layout_reverse",
+    "schedule_forward",
+    "schedule_reverse"
+  ],
+  "max_bytes": 512,
+  "schema": "gooo/tiny-path-decision-model/v1",
+  "temperature": 0.5,
+  "tensors": [
+    {
+      "bytes": 2458,
+      "cols": 256,
+      "count": 12288,
+      "encoding": "ternary_base3_5",
+      "name": "w1",
+      "offset": 0,
+      "rows": 48,
+      "scale": 0.1951175183057785
+    },
+    {
+      "bytes": 192,
+      "cols": 48,
+      "count": 48,
+      "encoding": "float32_le",
+      "name": "b1",
+      "offset": 2458,
+      "rows": 1,
+      "scale": 1.0
+    },
+    {
+      "bytes": 77,
+      "cols": 48,
+      "count": 384,
+      "encoding": "ternary_base3_5",
+      "name": "w2",
+      "offset": 2650,
+      "rows": 8,
+      "scale": 0.40118470788002014
+    },
+    {
+      "bytes": 32,
+      "cols": 8,
+      "count": 8,
+      "encoding": "float32_le",
+      "name": "b2",
+      "offset": 2727,
+      "rows": 1,
+      "scale": 1.0
+    }
+  ],
+  "variant": "qat_ternary",
+  "weights_file": "weights.bin",
+  "weights_sha256": "00525b0674ade3f95b85b6e620f8ea0b376c0231734239a27df1c0d5c7f710eb"
+}

research/bilingual-gooo-judgment-20261001/models/control/qat_ternary/weights.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00525b0674ade3f95b85b6e620f8ea0b376c0231734239a27df1c0d5c7f710eb
+size 2759

research/bilingual-gooo-judgment-20261001/models/paired/fp32/model.json

@@ -0,0 +1,64 @@
+{
+  "confidence_threshold": 1.0,
+  "feature_dim": 256,
+  "feature_version": "positioned_intent_ngrams_v1",
+  "hidden_dim": 48,
+  "labels": [
+    "reference_first",
+    "reference_second",
+    "assign_first",
+    "assign_second",
+    "layout_forward",
+    "layout_reverse",
+    "schedule_forward",
+    "schedule_reverse"
+  ],
+  "max_bytes": 512,
+  "schema": "gooo/tiny-path-decision-model/v1",
+  "temperature": 1.0,
+  "tensors": [
+    {
+      "bytes": 49152,
+      "cols": 256,
+      "count": 12288,
+      "encoding": "float32_le",
+      "name": "w1",
+      "offset": 0,
+      "rows": 48,
+      "scale": 1.0
+    },
+    {
+      "bytes": 192,
+      "cols": 48,
+      "count": 48,
+      "encoding": "float32_le",
+      "name": "b1",
+      "offset": 49152,
+      "rows": 1,
+      "scale": 1.0
+    },
+    {
+      "bytes": 1536,
+      "cols": 48,
+      "count": 384,
+      "encoding": "float32_le",
+      "name": "w2",
+      "offset": 49344,
+      "rows": 8,
+      "scale": 1.0
+    },
+    {
+      "bytes": 32,
+      "cols": 8,
+      "count": 8,
+      "encoding": "float32_le",
+      "name": "b2",
+      "offset": 50880,
+      "rows": 1,
+      "scale": 1.0
+    }
+  ],
+  "variant": "fp32",
+  "weights_file": "weights.bin",
+  "weights_sha256": "1dad702576db50fdb08c62f07728eb06e81c6d8a3011b78ce38b0338406aacf4"
+}

research/bilingual-gooo-judgment-20261001/models/paired/fp32/weights.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1dad702576db50fdb08c62f07728eb06e81c6d8a3011b78ce38b0338406aacf4
+size 50912

research/bilingual-gooo-judgment-20261001/models/paired/ptq_ternary/model.json

@@ -0,0 +1,64 @@
+{
+  "confidence_threshold": 1.0,
+  "feature_dim": 256,
+  "feature_version": "positioned_intent_ngrams_v1",
+  "hidden_dim": 48,
+  "labels": [
+    "reference_first",
+    "reference_second",
+    "assign_first",
+    "assign_second",
+    "layout_forward",
+    "layout_reverse",
+    "schedule_forward",
+    "schedule_reverse"
+  ],
+  "max_bytes": 512,
+  "schema": "gooo/tiny-path-decision-model/v1",
+  "temperature": 1.0,
+  "tensors": [
+    {
+      "bytes": 2458,
+      "cols": 256,
+      "count": 12288,
+      "encoding": "ternary_base3_5",
+      "name": "w1",
+      "offset": 0,
+      "rows": 48,
+      "scale": 0.1928061693906784
+    },
+    {
+      "bytes": 192,
+      "cols": 48,
+      "count": 48,
+      "encoding": "float32_le",
+      "name": "b1",
+      "offset": 2458,
+      "rows": 1,
+      "scale": 1.0
+    },
+    {
+      "bytes": 77,
+      "cols": 48,
+      "count": 384,
+      "encoding": "ternary_base3_5",
+      "name": "w2",
+      "offset": 2650,
+      "rows": 8,
+      "scale": 0.3953920304775238
+    },
+    {
+      "bytes": 32,
+      "cols": 8,
+      "count": 8,
+      "encoding": "float32_le",
+      "name": "b2",
+      "offset": 2727,
+      "rows": 1,
+      "scale": 1.0
+    }
+  ],
+  "variant": "ptq_ternary",
+  "weights_file": "weights.bin",
+  "weights_sha256": "d6fb8f32a29a674b5ee904f2300f51b9a0d610f57ec0ebcbc52635f73e45d619"
+}

research/bilingual-gooo-judgment-20261001/models/paired/ptq_ternary/weights.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d6fb8f32a29a674b5ee904f2300f51b9a0d610f57ec0ebcbc52635f73e45d619
+size 2759

research/bilingual-gooo-judgment-20261001/models/paired/qat_ternary/model.json

@@ -0,0 +1,64 @@
+{
+  "confidence_threshold": 1.0,
+  "feature_dim": 256,
+  "feature_version": "positioned_intent_ngrams_v1",
+  "hidden_dim": 48,
+  "labels": [
+    "reference_first",
+    "reference_second",
+    "assign_first",
+    "assign_second",
+    "layout_forward",
+    "layout_reverse",
+    "schedule_forward",
+    "schedule_reverse"
+  ],
+  "max_bytes": 512,
+  "schema": "gooo/tiny-path-decision-model/v1",
+  "temperature": 0.5,
+  "tensors": [
+    {
+      "bytes": 2458,
+      "cols": 256,
+      "count": 12288,
+      "encoding": "ternary_base3_5",
+      "name": "w1",
+      "offset": 0,
+      "rows": 48,
+      "scale": 0.1950712949037552
+    },
+    {
+      "bytes": 192,
+      "cols": 48,
+      "count": 48,
+      "encoding": "float32_le",
+      "name": "b1",
+      "offset": 2458,
+      "rows": 1,
+      "scale": 1.0
+    },
+    {
+      "bytes": 77,
+      "cols": 48,
+      "count": 384,
+      "encoding": "ternary_base3_5",
+      "name": "w2",
+      "offset": 2650,
+      "rows": 8,
+      "scale": 0.4011179506778717
+    },
+    {
+      "bytes": 32,
+      "cols": 8,
+      "count": 8,
+      "encoding": "float32_le",
+      "name": "b2",
+      "offset": 2727,
+      "rows": 1,
+      "scale": 1.0
+    }
+  ],
+  "variant": "qat_ternary",
+  "weights_file": "weights.bin",
+  "weights_sha256": "1673c7cf839bf8d7fe317495c5a118d67430ae3247afcc7eda2844dd9a9c11f4"
+}

research/bilingual-gooo-judgment-20261001/models/paired/qat_ternary/weights.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1673c7cf839bf8d7fe317495c5a118d67430ae3247afcc7eda2844dd9a9c11f4
+size 2759

research/bilingual-gooo-judgment-20261001/native/audit.json

@@ -0,0 +1,14 @@
+{
+  "decision": "PASS",
+  "new_go_processes": 0,
+  "new_model_calls": 0,
+  "new_native_calls": 0,
+  "preexecution_sha256": "5b974752f9e92b5a9b4cbb214aa16d2da5dd7a5b1951f32959544da9714ebeef",
+  "recorded_go_function_invocations": 4,
+  "recorded_go_processes": 2,
+  "recorded_model_predictions": 24,
+  "recorded_native_calls": 28,
+  "report_sha256": "d1679a36e8a05d7af007f1effb4ebdccda501fbe2cac9b6269a33c4130d761cf",
+  "schema": "gooo/bilingual-native-audit/v1",
+  "scope": "Offline reconciliation of captures, source/model/test identities, selected typed/native function and actual emitted-Go oracle values; ranking origin relies on pinned runner and native source evidence. No inference or subprocess replay."
+}

research/bilingual-gooo-judgment-20261001/native/execution-3999a9a5a3218e6b4c1b4ac80564351334e60f30bb71cb27c749bc264474981c.json

@@ -0,0 +1,15 @@
+{
+  "schema": "gooo/path-diagnosis-go-execution/v1",
+  "template": "",
+  "mask": 1,
+  "generated_go": "package path_diagnosis\n\n//gooo:generated:start id=\"path-diagnosis://activity/probe\" kind=\"activity\"\nfunc Probe(input int64) int64 {\n\treturn (2 - input)\n}\n\n//gooo:generated:end id=\"path-diagnosis://activity/probe\" kind=\"activity\"\n",
+  "generated_go_sha256": "3999a9a5a3218e6b4c1b4ac80564351334e60f30bb71cb27c749bc264474981c",
+  "inputs": [
+    2,
+    3
+  ],
+  "actual_go_values": [
+    0,
+    -1
+  ]
+}

research/bilingual-gooo-judgment-20261001/native/execution-e3114b2376adbc2821bb57135586eace3f125c0116c9dd5b6ca825935ea1db94.json

@@ -0,0 +1,15 @@
+{
+  "schema": "gooo/path-diagnosis-go-execution/v1",
+  "template": "",
+  "mask": 0,
+  "generated_go": "package path_diagnosis\n\n//gooo:generated:start id=\"path-diagnosis://activity/probe\" kind=\"activity\"\nfunc Probe(input int64) int64 {\n\treturn (input - 2)\n}\n\n//gooo:generated:end id=\"path-diagnosis://activity/probe\" kind=\"activity\"\n",
+  "generated_go_sha256": "e3114b2376adbc2821bb57135586eace3f125c0116c9dd5b6ca825935ea1db94",
+  "inputs": [
+    2,
+    3
+  ],
+  "actual_go_values": [
+    0,
+    1
+  ]
+}

research/bilingual-gooo-judgment-20261001/native/preexecution.json

@@ -0,0 +1,25 @@
+{
+  "go_binary_sha256": "132b69336a1f809932a8a20b0201dbbb980e86e3a323ae32e893639d83d71598",
+  "input_sha256": {
+    "fixture.gooo": "39bfb3e53a4ae2b6933b0401789b324df230e7e2ed7f539e2388d6825a88b956",
+    "plan-en-full.json": "714272167da94ee217361c48e0eace216db8207e31b4d10699f43d165540aa4b",
+    "plan-en-sparse.json": "c05593225955cf66111d29c9a188cee970008691e25d5d07c7e2ae9a7a13840e",
+    "plan-ko-full.json": "7cd81c3e7506a51a15a3811b761b45a3c92bb56aeeb794882a43cc46f76de8b3",
+    "plan-ko-sparse.json": "e4750e0315f89b76cfac0092391a64aacd43fe48bec50b3c2498c16bf9d01287"
+  },
+  "maximum_native_calls": 28,
+  "model_metadata_sha256": {
+    "control-fp32": "d19d98f6a4f2fc55524332a79db4e65ce019e5172f0e891d633e896c8f0115e6",
+    "control-ptq_ternary": "88b2cdda4fd9bd9bcd12caea597eaf3ac407fb63915dd4def1621960640b8ec3",
+    "control-qat_ternary": "91e4ac583e3e078e03d1ac88c303574c469e7683560c3dd74d736c846f6736d7",
+    "paired-fp32": "ccb48c6fd0f76458a231d0bd1266e88252708aa6f64093742aead6a5ede27dc3",
+    "paired-ptq_ternary": "5233de7a91a997358de890c6431cf0fa2fd6fcd4d3dfb10a8cd851349028b6c9",
+    "paired-qat_ternary": "47524ff50c41b5bde0be95a7447126ee20abfae59f2e2fdc1e3f78c42c5613d0"
+  },
+  "native": "1e01c96c54f2f8dd43334b8f580af93ffaea24df",
+  "native_binary_sha256": "234c3d7663afdeff882906d55f05454ce602a140e35280fbb138158b2bffc344",
+  "preregistration_sha256": "09e551b8f3820fc88313113f1befcad736540cdf8a20fbb9e612c132b24f2c31",
+  "runner": "90dbefe731cb95a65798ef6d380f22cfd811b5ca",
+  "schema": "gooo/bilingual-native-preexecution/v1",
+  "scope": "One reused subtraction intention, two languages, six trained arms and disconnected control; separate sparse and explicit full contracts. Full calls rerank once, rather than resuming the sparse call. No feedback training, upstream Laya or new independent tasks."
+}

research/bilingual-gooo-judgment-20261001/native/report.json

@@ -0,0 +1,78 @@
+{
+  "actual_generated_go_processes": 2,
+  "actual_go_function_invocations": 4,
+  "capture_sha256": {
+    "en-control-fp32-full-metrics.json": "8b47ee813e9629dae442142a42d4101e5ad2d85f8d02b1e9c220804cc3935e68",
+    "en-control-fp32-full.json": "d193c70adc43839dfc5aef879e85137c9dbbc7d1b2543564db7cb6093033346d",
+    "en-control-fp32-sparse-metrics.json": "9e3e48e8d0ce74b24bdf469cd35402b538dff2d321e54fae97f7cae0aefb821c",
+    "en-control-fp32-sparse.json": "f9ccc4a091e2dd4a3d10ed6dec66c0c2365dee25313add8a020780a1400ed72e",
+    "en-control-ptq_ternary-full-metrics.json": "9d76f29d365ae24de746160f779a3ba1b13737e1c63c389dbf74f89044ff61f7",
+    "en-control-ptq_ternary-full.json": "9978c646b83cee76c2ff5721a49c8f6fb1c46dc0ea2164916a78311adc275ec3",
+    "en-control-ptq_ternary-sparse-metrics.json": "3fc25297cec91178b7e24c021ba9379fa19bb71b5218a21218d4f60b2de96e55",
+    "en-control-ptq_ternary-sparse.json": "34c785e6557c9fe7e8d66cb5d50ac64aa14ced8f2836c0d9e0128a74e8908555",
+    "en-control-qat_ternary-full-metrics.json": "2cda3a29ce5d0b558093406c4def0db45ced7573f6fa062bfe03260f1ea7cddd",
+    "en-control-qat_ternary-full.json": "df001e70f091dfd7342c9c4fa883ab9919790279543a072c43e349d69bcf5966",
+    "en-control-qat_ternary-sparse-metrics.json": "69e07896bbb8d6b1a79ef9d7bb5639805c4437b5c1fb08beb6b0c5e5bfa931c7",
+    "en-control-qat_ternary-sparse.json": "d9440f9b9c317b007ae6b4754775b66918e98ad210359219a97d34302cda4962",
+    "en-offline-full-metrics.json": "52534b2f6a303514278f484d0ab22fdc44b20735ba1c7a3e602ab2a3ce36bdd5",
+    "en-offline-full.json": "b3ae38e491d647e3c79953ba7e3146df6ddb90376a7c0dc76bfc530acd7c7a0e",
+    "en-offline-sparse-metrics.json": "765d2a6f8cceecefa6123addada99a419aff6965a9035377ae557d9e37181839",
+    "en-offline-sparse.json": "f90a3023e0c4e5e92e3776c04d1c0be7e5f16a8ac65f24a7fe1654c24eea2766",
+    "en-paired-fp32-full-metrics.json": "302364887a46ea05cf03b1d584ea07734f370df28603a9d09bb1489338c64895",
+    "en-paired-fp32-full.json": "a0a1c724067eebc5c59163eab40497f4ab5baac3074823a5640437b68ed6b89a",
+    "en-paired-fp32-sparse-metrics.json": "a91050622aa86ec04b58ce7cd67858d210e22ffa1a7676b948f9fc31ded2c1c0",
+    "en-paired-fp32-sparse.json": "8a49198231b332af05812a402461ce9482ae9925838619919daedbe376b663c2",
+    "en-paired-ptq_ternary-full-metrics.json": "64f3a1ca14c09ab530fd2a96dc77197c8b96d030e315694882803685dd2914a6",
+    "en-paired-ptq_ternary-full.json": "dc72167f9d351047af395c1adc51979723d71615df82ccc1c4f66cac68e42142",
+    "en-paired-ptq_ternary-sparse-metrics.json": "4d6141b6c1dbc5b804033f8d3f16384a5ca5919521eb5d1750ec1bf6493cac03",
+    "en-paired-ptq_ternary-sparse.json": "be3fc0bcdc57eb45d5d82900469b5006288cb4f3657e4e48be8e2fccf9a728bb",
+    "en-paired-qat_ternary-full-metrics.json": "39e06e4468a3858fe96ed15820ae2b423cbeb6747cf117ed3bc68e1cf2cdd377",
+    "en-paired-qat_ternary-full.json": "f8c9e03878fbdfcf935ced71f466833f23ea3cdd14c18562afaa516438602fd1",
+    "en-paired-qat_ternary-sparse-metrics.json": "25f6de31a78c1c41f1d838b57bbec8afd46cd2e9b5174ad23b3b787fdb834b72",
+    "en-paired-qat_ternary-sparse.json": "647a1bdb948c9494996bbd3a7179182a58e7c90d16384bd8ed4a22f3a6b81e9e",
+    "ko-control-fp32-full-metrics.json": "83b5119a782efccfb62a83c7e303795c825ad1ad4ed56d5b9b223692150ec5f5",
+    "ko-control-fp32-full.json": "cff8fa48a8012872547ee6f4715262145c6269ef6ebc23099db79372eaa1a2c2",
+    "ko-control-fp32-sparse-metrics.json": "5ca8e9094acc4f7316417b8e2c31197553374b8f2a2a64adc53eb928d6bfb74d",
+    "ko-control-fp32-sparse.json": "34120b9492f121516da9bebed4be2ba5117af3fd78424135fb19d4e1258bd5b6",
+    "ko-control-ptq_ternary-full-metrics.json": "3953489a36a831d56a8bfa8ecfe5756a24cc69c44ae84fc569127ae687396833",
+    "ko-control-ptq_ternary-full.json": "80ce654dcbdf897af894eb6a3b0c11f4bfa9abe1c783cbd1ec40898e3da1c2bb",
+    "ko-control-ptq_ternary-sparse-metrics.json": "fd08be527a26b363b5c37fd48a5b601c16dd5eafc4dd503d1fa9f83434caff05",
+    "ko-control-ptq_ternary-sparse.json": "3ba3412b4c00629d184abaaa89ca62eb308b60e2a68961ee373c1037f9ce26a2",
+    "ko-control-qat_ternary-full-metrics.json": "2f673664339b2c4189d15b5c86c62ff2c9d6bed7b7133ff2945b764129aeab25",
+    "ko-control-qat_ternary-full.json": "e514751dfa02e05f26f345a89131e1a9f46ac200d260bf88aaefe7c1961b0e4a",
+    "ko-control-qat_ternary-sparse-metrics.json": "c85aa1c081a3b3eb883e9dfc495611132830a4612c29173b4c35e86c45c90145",
+    "ko-control-qat_ternary-sparse.json": "745d1537726c6fa8ff6c79cdbe3f6da7b093f67554d06de6c650ac828aa58082",
+    "ko-offline-full-metrics.json": "ca019924f9b9e0c4001ae57f50684731f2e476cb2e31fc46cd69f2647c2a0f06",
+    "ko-offline-full.json": "cd4240a54fdc85d44f7e6fe59ae177e5169ea9f918faff210cada39d4628c4c6",
+    "ko-offline-sparse-metrics.json": "f69d4b4e73a13d6792fa96eb94e26ae9329cd76ac1528d8a0a087dd30cc9abc3",
+    "ko-offline-sparse.json": "c467369281241fb03e130053f4e52c9d7676a4943852f1047b630a3a73bae684",
+    "ko-paired-fp32-full-metrics.json": "b27cad4df1d126c73dfd51d6d483f82e7c88fdf1b997cde522e40f0dbdd829f2",
+    "ko-paired-fp32-full.json": "a4c2a73478f7a896520055ae9551183eab7ccbb866992cb37816b2b2be02c801",
+    "ko-paired-fp32-sparse-metrics.json": "e7506018f7c35c040447a7288466f016f0852d46319649b4f48f257b7d27ac9f",
+    "ko-paired-fp32-sparse.json": "e0d6dfa17e6897184540cc6f78ebbdad42dfcced195142fbf08f2bea39b340bc",
+    "ko-paired-ptq_ternary-full-metrics.json": "ea8ce3c09f9cb998aeddcda156a3cf73343f1c7eb2925512f2e6ded7c948d4bb",
+    "ko-paired-ptq_ternary-full.json": "ed635ba04c60e12214c69aaf057da00d620ee6f60427fa30d3bf9a374fd45c38",
+    "ko-paired-ptq_ternary-sparse-metrics.json": "90b20a6b847d0a66fb633a5728ed14f9aa47e9ad34709e186ee3568d956dc913",
+    "ko-paired-ptq_ternary-sparse.json": "bc2936e47bbbd3b811bdb4c1a35bb192c752d4894c7bab8e6f53b5e04ca44664",
+    "ko-paired-qat_ternary-full-metrics.json": "eb703cb129f8d6cf8ad3baa3fafee98caf0f1a827bf7a820912a3e663fd1e5b8",
+    "ko-paired-qat_ternary-full.json": "c9a7761d9242a8c678f66e1d6f0117ca883033968c289fe9922b13d2d11749c1",
+    "ko-paired-qat_ternary-sparse-metrics.json": "33729d4252e414d22d3cbf4ca3d1e22ed964c09f1fbe5d18ec2e11c2039174d1",
+    "ko-paired-qat_ternary-sparse.json": "ce554299dc84f7735087db98c684ae1a2d81eb3bfa302e7fc3244845eec55177"
+  },
+  "counts": {
+    "full_contract_complete": 14,
+    "full_extra_candidate_attempts": 6,
+    "model_predictions": 24,
+    "native_calls": 28,
+    "sparse_selected_wrong_intention": 6
+  },
+  "decision": "PASS",
+  "execution_sha256": {
+    "execution-3999a9a5a3218e6b4c1b4ac80564351334e60f30bb71cb27c749bc264474981c.json": "bad884a152914c8af63473ffac5bb26b3fc4799e05c3cbf920ebd289102f3c19",
+    "execution-e3114b2376adbc2821bb57135586eace3f125c0116c9dd5b6ca825935ea1db94.json": "447b9212899990a6e7696150f99e0ee83ae0afe4342922dc96927db9e5eea742"
+  },
+  "limitation": "Fixed arm order and 28 child calls are integration evidence, not a causal latency or host utilization experiment. Full contracts are new native calls, not zero-call continuation.",
+  "new_independent_intentions": 0,
+  "schema": "gooo/bilingual-native-smoke/v1",
+  "scope": "One reused subtraction intention, two languages, six trained arms and disconnected control; separate sparse and explicit full contracts. Full calls rerank once, rather than resuming the sparse call. No feedback training, upstream Laya or new independent tasks."
+}

research/bilingual-gooo-judgment-20261001/preregistration.md

@@ -0,0 +1,39 @@
+# Bounded bilingual Gooo judgment, preregistered comparison
+
+Commit this protocol, Go pair generator and trainer before training or inference.
+Use the existing 6,240-view feedback curriculum. Pair only its Gooo views by
+program and adjacent English/Korean template, with identical fallback body,
+finite cases, eligible labels and soft target. Reuse 640 program groups in 1,040
+pairs: train 800, calibration 80, development test 160. There are zero new
+independent intentions; prior development evaluations have already used these
+groups. Do not call this an untouched language benchmark.
+
+Fine tune the fixed own feedback FP32 parent in two arms with identical seed and
+optimizer settings: finite soft-target NLL alone and the same NLL plus 0.25 times
+bilingual Jensen-Shannon divergence. Each arm runs 8 FP32 and 8 QAT epochs in
+128-pair batches, exactly 224 total optimizer steps across both arms. PTQ exports
+reuse each arm FP32. Select checkpoints on calibration loss only. Temperatures
+use calibration finite NLL; positive scaling does not change pair argmax.
+
+Report pair agreement, same-wrong-intention agreement, both-correct intention,
+finite best-set selection, probability divergence, finite completeness before
+and after a separate explicit complete test contract, added path attempts,
+model calls, latency and resident tensor bytes. An English/Korean agreement can
+be wrong. Sparse finite passing behavior is not a proof of instruction meaning.
+The complete contract must come from the independent authored arithmetic oracle;
+model confidence and a diagnostic witness cannot supply expected values.
+
+After training and numerical parity, use the adopted clean main compiler for
+28 actual smoke calls: two languages times six trained variants plus disconnected
+control times sparse/full contracts. Reuse the previous subtraction fixture.
+Sparse uses input 2 with expected 0 and budget 1; full additionally defines input
+3 with expected 1 and budget 2. Full calls rerank once; they are separate native
+calls, not a zero-call continuation of the sparse process. Execute each distinct
+selected emitted Go body on inputs 2 and 3 and retain actual values, including
+wrong sparse choices. This is one reused intention and zero new independent tasks.
+
+Go performs pairing, runtime, evaluation and codegen orchestration. Python is
+offline MPS training/export only. Retain both arms and all negative results.
+Publish weights as an experimental appendix, not as a silently promoted default.
+Bind source, data, pair references, parent, new weights, tests and outputs by
+digests. Keep PROV-O roles distinct. No new upstream Laya calls are implied.

research/bilingual-gooo-judgment-20261001/publication-manifest.json

@@ -0,0 +1,176 @@
+{
+  "schema": "gooo/bilingual-judgment-publication/v1",
+  "repository": "asketeddy/gooo-feedback-path-tiny-v1",
+  "prefix": "research/bilingual-gooo-judgment-20261001/",
+  "model_metadata_sha256": {
+    "control/fp32": "d19d98f6a4f2fc55524332a79db4e65ce019e5172f0e891d633e896c8f0115e6",
+    "control/ptq_ternary": "88b2cdda4fd9bd9bcd12caea597eaf3ac407fb63915dd4def1621960640b8ec3",
+    "control/qat_ternary": "91e4ac583e3e078e03d1ac88c303574c469e7683560c3dd74d736c846f6736d7",
+    "paired/fp32": "ccb48c6fd0f76458a231d0bd1266e88252708aa6f64093742aead6a5ede27dc3",
+    "paired/ptq_ternary": "5233de7a91a997358de890c6431cf0fa2fd6fcd4d3dfb10a8cd851349028b6c9",
+    "paired/qat_ternary": "47524ff50c41b5bde0be95a7447126ee20abfae59f2e2fdc1e3f78c42c5613d0"
+  },
+  "files": [
+    {
+      "path": "LICENSE",
+      "sha256": "3ad2cd8fe84a937a0005a2934e377432f2f86fe10ff86c4242cc48f49ebf4947",
+      "bytes": 1074
+    },
+    {
+      "path": "README.md",
+      "sha256": "6c0755154d4dc97e1a5887e4b9184bff92728a7630a93526a014a02cc5ae9d9c",
+      "bytes": 5673
+    },
+    {
+      "path": "data/manifest.json",
+      "sha256": "62a63812d9318031178b52823a211b77fa6e0c1f14c78f4e32c9d31c8fe016b5",
+      "bytes": 901
+    },
+    {
+      "path": "data/pairs.jsonl",
+      "sha256": "25e090b1c4d842b34cf7a1d3a994b7b6e4be70cb59389fe33420e5d8750574ce",
+      "bytes": 517568
+    },
+    {
+      "path": "judgment/preexecution.json",
+      "sha256": "3bf89a79e7942503826cba1e7a859fc6afa5fc2730a5777bc01b0fc6ea7a7972",
+      "bytes": 949
+    },
+    {
+      "path": "judgment/report.json",
+      "sha256": "89a07a45360b472b1e4343acbffb2912efbad5a56c2cfdd1c0d82ac808807775",
+      "bytes": 12783
+    },
+    {
+      "path": "main-ci/finalization.json",
+      "sha256": "73fe6b5ff14f0d49af85314187689ddda86a69c48242daa5990eb9f7daab5c4c",
+      "bytes": 4319
+    },
+    {
+      "path": "models/control/fp32/model.json",
+      "sha256": "d19d98f6a4f2fc55524332a79db4e65ce019e5172f0e891d633e896c8f0115e6",
+      "bytes": 1303
+    },
+    {
+      "path": "models/control/fp32/weights.bin",
+      "sha256": "e411e0f6bad0087b040a146d0a1030c272196d19df078aba87c38bd11e10db53",
+      "bytes": 50912
+    },
+    {
+      "path": "models/control/ptq_ternary/model.json",
+      "sha256": "88b2cdda4fd9bd9bcd12caea597eaf3ac407fb63915dd4def1621960640b8ec3",
+      "bytes": 1346
+    },
+    {
+      "path": "models/control/ptq_ternary/weights.bin",
+      "sha256": "030dd5400202f0522a70217e5f16f2dcfc9cb447f899c94582d5688efae86b5e",
+      "bytes": 2759
+    },
+    {
+      "path": "models/control/qat_ternary/model.json",
+      "sha256": "91e4ac583e3e078e03d1ac88c303574c469e7683560c3dd74d736c846f6736d7",
+      "bytes": 1345
+    },
+    {
+      "path": "models/control/qat_ternary/weights.bin",
+      "sha256": "00525b0674ade3f95b85b6e620f8ea0b376c0231734239a27df1c0d5c7f710eb",
+      "bytes": 2759
+    },
+    {
+      "path": "models/paired/fp32/model.json",
+      "sha256": "ccb48c6fd0f76458a231d0bd1266e88252708aa6f64093742aead6a5ede27dc3",
+      "bytes": 1303
+    },
+    {
+      "path": "models/paired/fp32/weights.bin",
+      "sha256": "1dad702576db50fdb08c62f07728eb06e81c6d8a3011b78ce38b0338406aacf4",
+      "bytes": 50912
+    },
+    {
+      "path": "models/paired/ptq_ternary/model.json",
+      "sha256": "5233de7a91a997358de890c6431cf0fa2fd6fcd4d3dfb10a8cd851349028b6c9",
+      "bytes": 1344
+    },
+    {
+      "path": "models/paired/ptq_ternary/weights.bin",
+      "sha256": "d6fb8f32a29a674b5ee904f2300f51b9a0d610f57ec0ebcbc52635f73e45d619",
+      "bytes": 2759
+    },
+    {
+      "path": "models/paired/qat_ternary/model.json",
+      "sha256": "47524ff50c41b5bde0be95a7447126ee20abfae59f2e2fdc1e3f78c42c5613d0",
+      "bytes": 1344
+    },
+    {
+      "path": "models/paired/qat_ternary/weights.bin",
+      "sha256": "1673c7cf839bf8d7fe317495c5a118d67430ae3247afcc7eda2844dd9a9c11f4",
+      "bytes": 2759
+    },
+    {
+      "path": "native/audit.json",
+      "sha256": "ca48a4b670ae115480908b75cc630ab2efa9208b1b60901d2d6a0867e2f53dd5",
+      "bytes": 713
+    },
+    {
+      "path": "native/execution-3999a9a5a3218e6b4c1b4ac80564351334e60f30bb71cb27c749bc264474981c.json",
+      "sha256": "bad884a152914c8af63473ffac5bb26b3fc4799e05c3cbf920ebd289102f3c19",
+      "bytes": 521
+    },
+    {
+      "path": "native/execution-e3114b2376adbc2821bb57135586eace3f125c0116c9dd5b6ca825935ea1db94.json",
+      "sha256": "447b9212899990a6e7696150f99e0ee83ae0afe4342922dc96927db9e5eea742",
+      "bytes": 520
+    },
+    {
+      "path": "native/preexecution.json",
+      "sha256": "5b974752f9e92b5a9b4cbb214aa16d2da5dd7a5b1951f32959544da9714ebeef",
+      "bytes": 1827
+    },
+    {
+      "path": "native/report.json",
+      "sha256": "d1679a36e8a05d7af007f1effb4ebdccda501fbe2cac9b6269a33c4130d761cf",
+      "bytes": 7303
+    },
+    {
+      "path": "preregistration.md",
+      "sha256": "09e551b8f3820fc88313113f1befcad736540cdf8a20fbb9e612c132b24f2c31",
+      "bytes": 2647
+    },
+    {
+      "path": "training/attempts.json",
+      "sha256": "b64b89746b3c21c33538a4ccae8baa5a54b9675d44cb09beb6a8d8a75f9708ae",
+      "bytes": 730
+    },
+    {
+      "path": "training/control/go-audit.json",
+      "sha256": "17be6916de9c109364b2705224e99162a4c5014d98782ece27ab5c9f80884f74",
+      "bytes": 7564
+    },
+    {
+      "path": "training/control/go-parity.json",
+      "sha256": "63436a2918e2f926e2557aa8c68b4a4c705f9fc37525ab497fc854f0292a5292",
+      "bytes": 756836
+    },
+    {
+      "path": "training/paired/go-audit.json",
+      "sha256": "9b35598066ac9adba53da7d0f5f8ed5c4b31cd535ecfe6e5f9b56e9545047f01",
+      "bytes": 7562
+    },
+    {
+      "path": "training/paired/go-parity.json",
+      "sha256": "0ec3c26f75454b5afe43a6bf0e8a566464561e72be7b1dd9baf11a4ee68968c6",
+      "bytes": 756862
+    },
+    {
+      "path": "training/preexecution.json",
+      "sha256": "535d154552df0780b9c9600b272920878f9de405c9a3710785c511989b2f3c02",
+      "bytes": 1424
+    },
+    {
+      "path": "training/report.json",
+      "sha256": "fc3b1ff3bee93306a66f5cd1cf80e379f93cae40cca3bf9e9f9569ac678994c8",
+      "bytes": 11734
+    }
+  ],
+  "scope": "Experimental own-model control and bilingual consistency weights, reused synthetic Gooo pairs, negative comparisons, actual adopted-main compiler dogfood and explicit full-contract continuation. Full raw prediction/native captures are on GitHub. Existing core weights/card untouched. No private logs, paths, credentials, upstream Laya weights or default checkpoint promotion."
+}

research/bilingual-gooo-judgment-20261001/training/attempts.json

@@ -0,0 +1,9 @@
+{
+  "schema": "gooo/bilingual-training-attempts/v1",
+  "source_revision": "d4016f9708a931fe1095d7509a4c4787fab21dff",
+  "attempts": [
+    {"ordinal": 1, "stage": "source_preflight", "outcome": "REJECTED", "reason": "supplied revision did not match committed HEAD", "output_created": false, "optimizer_steps": 0},
+    {"ordinal": 2, "stage": "bounded_mps_training", "outcome": "TRAINED_AND_EXPORTED", "optimizer_steps": 224, "report": "runs/bilingual-judgment-mps-20261001/report.json"}
+  ],
+  "scope": "The failed source-pin invocation is preserved; it performed no training. Corrected invocation used the actual committed source, same preregistered parameters and fresh output. No result-based rerun or hyperparameter search."
+}

research/bilingual-gooo-judgment-20261001/training/control/go-audit.json

@@ -0,0 +1,190 @@
+{
+  "cells": {
+    "feedback-fp32": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 0,
+      "metadata_sha256": "d19d98f6a4f2fc55524332a79db4e65ce019e5172f0e891d633e896c8f0115e6",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 50912,
+      "resident_tensor_bytes": 50912,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 763,
+        "original_intention_label_agreement": 751,
+        "finite_cases_passed": 3802,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 8828987,
+        "paired_soft_target_nll": 0.43563684393612945,
+        "mean_probability_on_best_set": 0.7615854227206328
+      },
+      "temperature": 1,
+      "weights_sha256": "e411e0f6bad0087b040a146d0a1030c272196d19df078aba87c38bd11e10db53",
+      "workspace_bytes": 1248
+    },
+    "feedback-ptq_ternary": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "metadata_sha256": "88b2cdda4fd9bd9bcd12caea597eaf3ac407fb63915dd4def1621960640b8ec3",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 2759,
+      "resident_tensor_bytes": 12896,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 745,
+        "original_intention_label_agreement": 731,
+        "finite_cases_passed": 3723,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 10327248,
+        "paired_soft_target_nll": 0.5499252980681922,
+        "mean_probability_on_best_set": 0.6488880047959411
+      },
+      "temperature": 1,
+      "weights_sha256": "030dd5400202f0522a70217e5f16f2dcfc9cb447f899c94582d5688efae86b5e",
+      "workspace_bytes": 1248
+    },
+    "feedback-qat_ternary": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "metadata_sha256": "91e4ac583e3e078e03d1ac88c303574c469e7683560c3dd74d736c846f6736d7",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 2759,
+      "resident_tensor_bytes": 12896,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 696,
+        "original_intention_label_agreement": 677,
+        "finite_cases_passed": 3446,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 10239754,
+        "paired_soft_target_nll": 0.5322039571345915,
+        "mean_probability_on_best_set": 0.6933053385131884
+      },
+      "temperature": 0.5,
+      "weights_sha256": "00525b0674ade3f95b85b6e620f8ea0b376c0231734239a27df1c0d5c7f710eb",
+      "workspace_bytes": 1248
+    },
+    "offline": {
+      "generated_go_processes": 0,
+      "native_calls": 0,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 518,
+        "original_intention_label_agreement": 480,
+        "finite_cases_passed": 2514,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 0,
+        "actual_local_predictions": 0,
+        "summed_observed_prediction_ns": 0,
+        "paired_soft_target_nll": 0.6931471805599301,
+        "mean_probability_on_best_set": 0.5395833333333333
+      }
+    },
+    "parent-fp32": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 0,
+      "metadata_sha256": "1ea3bada068f2487418f17270db4ba6785a98c3ad60e0ad7eb92359400a01bb5",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 50912,
+      "resident_tensor_bytes": 50912,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 814,
+        "original_intention_label_agreement": 806,
+        "finite_cases_passed": 4072,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 9218169,
+        "paired_soft_target_nll": 0.4243594405789759,
+        "mean_probability_on_best_set": 0.7569873561085224
+      },
+      "temperature": 1.3202666,
+      "weights_sha256": "3732a103720fdb0da0ea611471ecde7472e0b9166322394ba9352ffb855b9354",
+      "workspace_bytes": 1248
+    },
+    "parent-ptq_ternary": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "metadata_sha256": "7c4eb4068d76e83620a9b8e7ce7f8d948b5e2436d44c7b79cf241da609dab894",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 2759,
+      "resident_tensor_bytes": 12896,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 808,
+        "original_intention_label_agreement": 798,
+        "finite_cases_passed": 4038,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 10603733,
+        "paired_soft_target_nll": 0.47582618713480274,
+        "mean_probability_on_best_set": 0.6987255461003449
+      },
+      "temperature": 0.5938498,
+      "weights_sha256": "01a6574dedba327361cd3e4fda00a4d4dea34e162aab09885174bcd41ed4da47",
+      "workspace_bytes": 1248
+    },
+    "parent-qat_ternary": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "metadata_sha256": "368e37e7899cecba03874b69e933525a2f8c90ecc53d787d8e698de6aa21c087",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 2759,
+      "resident_tensor_bytes": 12896,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 817,
+        "original_intention_label_agreement": 799,
+        "finite_cases_passed": 4076,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 10512801,
+        "paired_soft_target_nll": 0.4499589790650594,
+        "mean_probability_on_best_set": 0.7650202976724335
+      },
+      "temperature": 1.3202666,
+      "weights_sha256": "456c686911b95bb933a7dd9275d2611974482ecf353831ab17b593c6bd3d86de",
+      "workspace_bytes": 1248
+    }
+  },
+  "dataset_sha256": "570d1d73bfea32662b869e5cfbf77e6d04df838b703191b076be58c1d2180dbf",
+  "decision": "PASS",
+  "development_test_views": 960,
+  "existing_original_test_instructions": 320,
+  "model_scoring_actual_predictions": 5760,
+  "new_generated_go_processes": 0,
+  "new_native_calls": 0,
+  "parity": {
+    "actual_local_predictions": 96,
+    "decision": "PASS",
+    "fixture_sha256": "63436a2918e2f926e2557aa8c68b4a4c705f9fc37525ab497fc854f0292a5292",
+    "maximum_logit_or_probability_absolute_error": 0.0000015497207641601562,
+    "rows": 96,
+    "tolerance": 0.0001
+  },
+  "parity_actual_predictions": 96,
+  "schema": "gooo/feedback-path-model-audit/v1",
+  "scope": "Go predictions on six model arms and deterministic offline control, selected typed-body finite outcomes and numeric parity. Reused development split, closed eligible pairs and 76 finite ties are explicit. Pair scores are ranking hints, not general intention correctness, native codegen or process-wide memory/utilization measurements.",
+  "total_actual_predictions": 5856
+}

research/bilingual-gooo-judgment-20261001/training/paired/go-audit.json

@@ -0,0 +1,190 @@
+{
+  "cells": {
+    "feedback-fp32": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 0,
+      "metadata_sha256": "ccb48c6fd0f76458a231d0bd1266e88252708aa6f64093742aead6a5ede27dc3",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 50912,
+      "resident_tensor_bytes": 50912,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 763,
+        "original_intention_label_agreement": 750,
+        "finite_cases_passed": 3802,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 9938749,
+        "paired_soft_target_nll": 0.435348147756339,
+        "mean_probability_on_best_set": 0.7616412710946964
+      },
+      "temperature": 1,
+      "weights_sha256": "1dad702576db50fdb08c62f07728eb06e81c6d8a3011b78ce38b0338406aacf4",
+      "workspace_bytes": 1248
+    },
+    "feedback-ptq_ternary": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "metadata_sha256": "5233de7a91a997358de890c6431cf0fa2fd6fcd4d3dfb10a8cd851349028b6c9",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 2759,
+      "resident_tensor_bytes": 12896,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 745,
+        "original_intention_label_agreement": 731,
+        "finite_cases_passed": 3723,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 11202884,
+        "paired_soft_target_nll": 0.5499090188010417,
+        "mean_probability_on_best_set": 0.6489005811188344
+      },
+      "temperature": 1,
+      "weights_sha256": "d6fb8f32a29a674b5ee904f2300f51b9a0d610f57ec0ebcbc52635f73e45d619",
+      "workspace_bytes": 1248
+    },
+    "feedback-qat_ternary": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "metadata_sha256": "47524ff50c41b5bde0be95a7447126ee20abfae59f2e2fdc1e3f78c42c5613d0",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 2759,
+      "resident_tensor_bytes": 12896,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 699,
+        "original_intention_label_agreement": 680,
+        "finite_cases_passed": 3467,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 10984238,
+        "paired_soft_target_nll": 0.5320869778256322,
+        "mean_probability_on_best_set": 0.6931064019052796
+      },
+      "temperature": 0.5,
+      "weights_sha256": "1673c7cf839bf8d7fe317495c5a118d67430ae3247afcc7eda2844dd9a9c11f4",
+      "workspace_bytes": 1248
+    },
+    "offline": {
+      "generated_go_processes": 0,
+      "native_calls": 0,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 518,
+        "original_intention_label_agreement": 480,
+        "finite_cases_passed": 2514,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 0,
+        "actual_local_predictions": 0,
+        "summed_observed_prediction_ns": 0,
+        "paired_soft_target_nll": 0.6931471805599301,
+        "mean_probability_on_best_set": 0.5395833333333333
+      }
+    },
+    "parent-fp32": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 0,
+      "metadata_sha256": "1ea3bada068f2487418f17270db4ba6785a98c3ad60e0ad7eb92359400a01bb5",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 50912,
+      "resident_tensor_bytes": 50912,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 814,
+        "original_intention_label_agreement": 806,
+        "finite_cases_passed": 4072,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 9850328,
+        "paired_soft_target_nll": 0.4243594405789759,
+        "mean_probability_on_best_set": 0.7569873561085224
+      },
+      "temperature": 1.3202666,
+      "weights_sha256": "3732a103720fdb0da0ea611471ecde7472e0b9166322394ba9352ffb855b9354",
+      "workspace_bytes": 1248
+    },
+    "parent-ptq_ternary": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "metadata_sha256": "7c4eb4068d76e83620a9b8e7ce7f8d948b5e2436d44c7b79cf241da609dab894",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 2759,
+      "resident_tensor_bytes": 12896,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 808,
+        "original_intention_label_agreement": 798,
+        "finite_cases_passed": 4038,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 11413299,
+        "paired_soft_target_nll": 0.47582618713480274,
+        "mean_probability_on_best_set": 0.6987255461003449
+      },
+      "temperature": 0.5938498,
+      "weights_sha256": "01a6574dedba327361cd3e4fda00a4d4dea34e162aab09885174bcd41ed4da47",
+      "workspace_bytes": 1248
+    },
+    "parent-qat_ternary": {
+      "confidence_threshold": 1,
+      "generated_go_processes": 0,
+      "matrix_scale_bytes": 8,
+      "metadata_sha256": "368e37e7899cecba03874b69e933525a2f8c90ecc53d787d8e698de6aa21c087",
+      "native_calls": 0,
+      "packed_weights_file_bytes": 2759,
+      "resident_tensor_bytes": 12896,
+      "scores": {
+        "observed_views": 960,
+        "finite_best_set_selected": 817,
+        "original_intention_label_agreement": 799,
+        "finite_cases_passed": 4076,
+        "finite_cases_total": 5120,
+        "multiple_best_finite_labels": 76,
+        "global_abstentions": 960,
+        "actual_local_predictions": 960,
+        "summed_observed_prediction_ns": 11467270,
+        "paired_soft_target_nll": 0.4499589790650594,
+        "mean_probability_on_best_set": 0.7650202976724335
+      },
+      "temperature": 1.3202666,
+      "weights_sha256": "456c686911b95bb933a7dd9275d2611974482ecf353831ab17b593c6bd3d86de",
+      "workspace_bytes": 1248
+    }
+  },
+  "dataset_sha256": "570d1d73bfea32662b869e5cfbf77e6d04df838b703191b076be58c1d2180dbf",
+  "decision": "PASS",
+  "development_test_views": 960,
+  "existing_original_test_instructions": 320,
+  "model_scoring_actual_predictions": 5760,
+  "new_generated_go_processes": 0,
+  "new_native_calls": 0,
+  "parity": {
+    "actual_local_predictions": 96,
+    "decision": "PASS",
+    "fixture_sha256": "0ec3c26f75454b5afe43a6bf0e8a566464561e72be7b1dd9baf11a4ee68968c6",
+    "maximum_logit_or_probability_absolute_error": 0.0000015497207641601562,
+    "rows": 96,
+    "tolerance": 0.0001
+  },
+  "parity_actual_predictions": 96,
+  "schema": "gooo/feedback-path-model-audit/v1",
+  "scope": "Go predictions on six model arms and deterministic offline control, selected typed-body finite outcomes and numeric parity. Reused development split, closed eligible pairs and 76 finite ties are explicit. Pair scores are ranking hints, not general intention correctness, native codegen or process-wide memory/utilization measurements.",
+  "total_actual_predictions": 5856
+}

research/bilingual-gooo-judgment-20261001/training/preexecution.json

@@ -0,0 +1,28 @@
+{
+  "arms": {
+    "control": 0.0,
+    "paired": 0.25
+  },
+  "dataset_sha256": "570d1d73bfea32662b869e5cfbf77e6d04df838b703191b076be58c1d2180dbf",
+  "device": "mps",
+  "epochs_per_trainable_variant": 8,
+  "manifest_sha256": "62a63812d9318031178b52823a211b77fa6e0c1f14c78f4e32c9d31c8fe016b5",
+  "maximum_total_optimizer_steps": 224,
+  "numpy": "2.4.6",
+  "pairs_sha256": "25e090b1c4d842b34cf7a1d3a994b7b6e4be70cb59389fe33420e5d8750574ce",
+  "parent": {
+    "metadata_sha256": "47bd3ed2037c8ba0af31ec4cad3a47fe1182af171845406aba01c5107f4b24b7",
+    "weights_sha256": "73eedef1d60f49129190683a41cbf82229c1e3603fd44cdd17792d015f499954"
+  },
+  "schema": "gooo/bilingual-judgment-training-preexecution/v1",
+  "scope": "Reused closed bilingual Gooo judgment pairs, finite-set supervision; agreement can be wrong. No new intentions, untouched holdout, online learning or Laya weight tuning.",
+  "seed": 20261008,
+  "source_revision": "d4016f9708a931fe1095d7509a4c4787fab21dff",
+  "torch": "2.14.0",
+  "training_sources_sha256": {
+    "train_bilingual_judgment_v1.py": "5f7ae5a45248a91bea3f9945317bd7629ad06b6d06a317034913d8b5e1dd87fc",
+    "train_feedback_path_v1.py": "3c5bb7a83be4098d91023ff85bbad0ce7251fcc4a929e7ab1bda6437fe2f007a",
+    "train_pilot_v2.py": "4ddfe02e007cb6eb84fa696c5f965eaebbfebc681b2a723f2895e1f402fd72d4",
+    "train_typed_path_v1.py": "8858b61bdb135e5b871238f5451adacd2a9670d1c6087673e9ec5772eb00f8ef"
+  }
+}

research/bilingual-gooo-judgment-20261001/training/report.json

@@ -0,0 +1,310 @@
+{
+  "evaluation": "Separate Go parity, bilingual judgment and finite continuation reports required",
+  "exports": {
+    "control": {
+      "fp32": {
+        "metadata_sha256": "d19d98f6a4f2fc55524332a79db4e65ce019e5172f0e891d633e896c8f0115e6",
+        "packed_weights_bytes": 50912,
+        "temperature": 1.0,
+        "weights_sha256": "e411e0f6bad0087b040a146d0a1030c272196d19df078aba87c38bd11e10db53"
+      },
+      "ptq_ternary": {
+        "metadata_sha256": "88b2cdda4fd9bd9bcd12caea597eaf3ac407fb63915dd4def1621960640b8ec3",
+        "packed_weights_bytes": 2759,
+        "temperature": 1.0,
+        "weights_sha256": "030dd5400202f0522a70217e5f16f2dcfc9cb447f899c94582d5688efae86b5e"
+      },
+      "qat_ternary": {
+        "metadata_sha256": "91e4ac583e3e078e03d1ac88c303574c469e7683560c3dd74d736c846f6736d7",
+        "packed_weights_bytes": 2759,
+        "temperature": 0.5,
+        "weights_sha256": "00525b0674ade3f95b85b6e620f8ea0b376c0231734239a27df1c0d5c7f710eb"
+      }
+    },
+    "paired": {
+      "fp32": {
+        "metadata_sha256": "ccb48c6fd0f76458a231d0bd1266e88252708aa6f64093742aead6a5ede27dc3",
+        "packed_weights_bytes": 50912,
+        "temperature": 1.0,
+        "weights_sha256": "1dad702576db50fdb08c62f07728eb06e81c6d8a3011b78ce38b0338406aacf4"
+      },
+      "ptq_ternary": {
+        "metadata_sha256": "5233de7a91a997358de890c6431cf0fa2fd6fcd4d3dfb10a8cd851349028b6c9",
+        "packed_weights_bytes": 2759,
+        "temperature": 1.0,
+        "weights_sha256": "d6fb8f32a29a674b5ee904f2300f51b9a0d610f57ec0ebcbc52635f73e45d619"
+      },
+      "qat_ternary": {
+        "metadata_sha256": "47524ff50c41b5bde0be95a7447126ee20abfae59f2e2fdc1e3f78c42c5613d0",
+        "packed_weights_bytes": 2759,
+        "temperature": 0.5,
+        "weights_sha256": "1673c7cf839bf8d7fe317495c5a118d67430ae3247afcc7eda2844dd9a9c11f4"
+      }
+    }
+  },
+  "limitations": [
+    "Control and paired arms share seed/data/initial FP32; QAT starts from each arm FP32",
+    "Agreement can select the same wrong path; finite tie targets do not prove linguistic intent",
+    "Existing development split reused, zero new independent intention groups",
+    "MPS memory samples and process lifetime RSS are not host GPU/CPU utilization",
+    "Ternary disk packing decodes to int8; weights are experimental without default promotion"
+  ],
+  "optimizer_steps": 224,
+  "preexecution_sha256": "535d154552df0780b9c9600b272920878f9de405c9a3710785c511989b2f3c02",
+  "schema": "gooo/bilingual-judgment-training-report/v1",
+  "status": "TRAINED_AND_EXPORTED",
+  "training": {
+    "control": {
+      "fp32": {
+        "consistency_weight": 0.0,
+        "epochs": 8,
+        "history": [
+          {
+            "calibration_bilingual_js": 0.08627698570489883,
+            "calibration_finite_nll": 0.41395625472068787,
+            "epoch": 1,
+            "train_objective": 0.32499743359429495
+          },
+          {
+            "calibration_bilingual_js": 0.08874155580997467,
+            "calibration_finite_nll": 0.43212684988975525,
+            "epoch": 2,
+            "train_objective": 0.2924360249723707
+          },
+          {
+            "calibration_bilingual_js": 0.08281097561120987,
+            "calibration_finite_nll": 0.40457218885421753,
+            "epoch": 3,
+            "train_objective": 0.3055798347507204
+          },
+          {
+            "calibration_bilingual_js": 0.07940874248743057,
+            "calibration_finite_nll": 0.39476338028907776,
+            "epoch": 4,
+            "train_objective": 0.29483249357768465
+          },
+          {
+            "calibration_bilingual_js": 0.08098433911800385,
+            "calibration_finite_nll": 0.40144309401512146,
+            "epoch": 5,
+            "train_objective": 0.27147663278239115
+          },
+          {
+            "calibration_bilingual_js": 0.08445011079311371,
+            "calibration_finite_nll": 0.41266027092933655,
+            "epoch": 6,
+            "train_objective": 0.27507355383464266
+          },
+          {
+            "calibration_bilingual_js": 0.08766905963420868,
+            "calibration_finite_nll": 0.41783779859542847,
+            "epoch": 7,
+            "train_objective": 0.27138627426964895
+          },
+          {
+            "calibration_bilingual_js": 0.08710484206676483,
+            "calibration_finite_nll": 0.4101777970790863,
+            "epoch": 8,
+            "train_objective": 0.27402038659368244
+          }
+        ],
+        "loop_wall_seconds": 1.288649250054732,
+        "mps_allocated_peak_sampled_bytes": 2033920,
+        "mps_driver_peak_sampled_bytes": 53166080,
+        "optimizer_steps": 56,
+        "process_lifetime_peak_rss_bytes": 521945088,
+        "selection": "minimum calibration finite NLL plus weighted bilingual JS; development test excluded",
+        "training_pairs": 800,
+        "training_views": 1600
+      },
+      "qat_ternary": {
+        "consistency_weight": 0.0,
+        "epochs": 8,
+        "history": [
+          {
+            "calibration_bilingual_js": 0.018491018563508987,
+            "calibration_finite_nll": 0.5630286931991577,
+            "epoch": 1,
+            "train_objective": 0.47267789925847736
+          },
+          {
+            "calibration_bilingual_js": 0.019454553723335266,
+            "calibration_finite_nll": 0.5532973408699036,
+            "epoch": 2,
+            "train_objective": 0.45271050930023193
+          },
+          {
+            "calibration_bilingual_js": 0.01863386668264866,
+            "calibration_finite_nll": 0.546625554561615,
+            "epoch": 3,
+            "train_objective": 0.4514418712684086
+          },
+          {
+            "calibration_bilingual_js": 0.0178647693246603,
+            "calibration_finite_nll": 0.544744610786438,
+            "epoch": 4,
+            "train_objective": 0.4563551204545157
+          },
+          {
+            "calibration_bilingual_js": 0.01690916158258915,
+            "calibration_finite_nll": 0.5444742441177368,
+            "epoch": 5,
+            "train_objective": 0.441873094865254
+          },
+          {
+            "calibration_bilingual_js": 0.015987079590559006,
+            "calibration_finite_nll": 0.5429291725158691,
+            "epoch": 6,
+            "train_objective": 0.43860537239483427
+          },
+          {
+            "calibration_bilingual_js": 0.015393257141113281,
+            "calibration_finite_nll": 0.5442538261413574,
+            "epoch": 7,
+            "train_objective": 0.4459982897554125
+          },
+          {
+            "calibration_bilingual_js": 0.015749486163258553,
+            "calibration_finite_nll": 0.5421851277351379,
+            "epoch": 8,
+            "train_objective": 0.4413776695728302
+          }
+        ],
+        "loop_wall_seconds": 0.07522045820951462,
+        "mps_allocated_peak_sampled_bytes": 2033920,
+        "mps_driver_peak_sampled_bytes": 53166080,
+        "optimizer_steps": 56,
+        "process_lifetime_peak_rss_bytes": 522682368,
+        "selection": "minimum calibration finite NLL plus weighted bilingual JS; development test excluded",
+        "training_pairs": 800,
+        "training_views": 1600
+      }
+    },
+    "paired": {
+      "fp32": {
+        "consistency_weight": 0.25,
+        "epochs": 8,
+        "history": [
+          {
+            "calibration_bilingual_js": 0.08620867133140564,
+            "calibration_finite_nll": 0.41367703676223755,
+            "epoch": 1,
+            "train_objective": 0.3263478917734964
+          },
+          {
+            "calibration_bilingual_js": 0.08859829604625702,
+            "calibration_finite_nll": 0.4316110610961914,
+            "epoch": 2,
+            "train_objective": 0.29387152194976807
+          },
+          {
+            "calibration_bilingual_js": 0.08270813524723053,
+            "calibration_finite_nll": 0.40424594283103943,
+            "epoch": 3,
+            "train_objective": 0.307049532021795
+          },
+          {
+            "calibration_bilingual_js": 0.07947338372468948,
+            "calibration_finite_nll": 0.39479437470436096,
+            "epoch": 4,
+            "train_objective": 0.29635951135839733
+          },
+          {
+            "calibration_bilingual_js": 0.08124524354934692,
+            "calibration_finite_nll": 0.40207332372665405,
+            "epoch": 5,
+            "train_objective": 0.2729123703071049
+          },
+          {
+            "calibration_bilingual_js": 0.08452852070331573,
+            "calibration_finite_nll": 0.412418931722641,
+            "epoch": 6,
+            "train_objective": 0.27659926457064493
+          },
+          {
+            "calibration_bilingual_js": 0.08743865042924881,
+            "calibration_finite_nll": 0.41656702756881714,
+            "epoch": 7,
+            "train_objective": 0.2729770711490086
+          },
+          {
+            "calibration_bilingual_js": 0.08687029778957367,
+            "calibration_finite_nll": 0.40897130966186523,
+            "epoch": 8,
+            "train_objective": 0.27547467393534525
+          }
+        ],
+        "loop_wall_seconds": 0.08172204112634063,
+        "mps_allocated_peak_sampled_bytes": 2033920,
+        "mps_driver_peak_sampled_bytes": 53166080,
+        "optimizer_steps": 56,
+        "process_lifetime_peak_rss_bytes": 528695296,
+        "selection": "minimum calibration finite NLL plus weighted bilingual JS; development test excluded",
+        "training_pairs": 800,
+        "training_views": 1600
+      },
+      "qat_ternary": {
+        "consistency_weight": 0.25,
+        "epochs": 8,
+        "history": [
+          {
+            "calibration_bilingual_js": 0.01868617720901966,
+            "calibration_finite_nll": 0.5630553960800171,
+            "epoch": 1,
+            "train_objective": 0.47338799067905973
+          },
+          {
+            "calibration_bilingual_js": 0.019224733114242554,
+            "calibration_finite_nll": 0.5534323453903198,
+            "epoch": 2,
+            "train_objective": 0.45437736596379963
+          },
+          {
+            "calibration_bilingual_js": 0.018178831785917282,
+            "calibration_finite_nll": 0.5529493689537048,
+            "epoch": 3,
+            "train_objective": 0.45243516990116667
+          },
+          {
+            "calibration_bilingual_js": 0.017919739708304405,
+            "calibration_finite_nll": 0.5446000099182129,
+            "epoch": 4,
+            "train_objective": 0.46067777276039124
+          },
+          {
+            "calibration_bilingual_js": 0.016833722591400146,
+            "calibration_finite_nll": 0.5442108511924744,
+            "epoch": 5,
+            "train_objective": 0.44320770672389437
+          },
+          {
+            "calibration_bilingual_js": 0.01646900363266468,
+            "calibration_finite_nll": 0.5432742238044739,
+            "epoch": 6,
+            "train_objective": 0.44047230056353975
+          },
+          {
+            "calibration_bilingual_js": 0.01569296605885029,
+            "calibration_finite_nll": 0.5453439354896545,
+            "epoch": 7,
+            "train_objective": 0.4464301126343863
+          },
+          {
+            "calibration_bilingual_js": 0.015689831227064133,
+            "calibration_finite_nll": 0.541408360004425,
+            "epoch": 8,
+            "train_objective": 0.44318767956324984
+          }
+        ],
+        "loop_wall_seconds": 0.0721320838201791,
+        "mps_allocated_peak_sampled_bytes": 2033920,
+        "mps_driver_peak_sampled_bytes": 53166080,
+        "optimizer_steps": 56,
+        "process_lifetime_peak_rss_bytes": 528809984,
+        "selection": "minimum calibration finite NLL plus weighted bilingual JS; development test excluded",
+        "training_pairs": 800,
+        "training_views": 1600
+      }
+    }
+  }
+}