Skip to content

Preserve rejected prefix diagnostics and isolate Git fixture signing #64

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
CompleteDotTech merged 1 commit into from
Oct 3, 2026
Merged

Preserve rejected prefix diagnostics and isolate Git fixture signing #64

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
313 changes: 313 additions & 0 deletions docs/evidence/native/prefix-gradient-diagnostics-20261002.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,313 @@
{
"schema": "layev-prefix-gradient-diagnostics/1",
"scope": "Rejected actual-weight derivative candidates and bounded localization; no native completion claim",
"inventory": {
"tensor_count": 294,
"scalar_count": 494090176,
"backbone_tensor_count": 290,
"backbone_scalar_count": 494032768
},
"gradient_tolerance": {
"atol": 2e-05,
"rtol": 2e-05,
"formula": "abs(actual-reference) <= atol + rtol*abs(reference)"
},
"frozen_request_geometry": {
"state_tokens": 25,
"suffix_tokens": [
57,
42,
135
],
"full_row_tokens": [
82,
67,
160
],
"question_groups": [
[
1,
0
],
[
2
]
],
"group_widths": [
57,
135
],
"flattened_suffix_projection_rows": [
114,
135
],
"policy": {
"max_branches": 3,
"max_padded_tokens": 2048,
"max_cache_bytes": 67108864,
"sort_by_length": true
},
"interpretation": "The cache budget produces these groups; the planner does not explicitly bucket questions by length."
},
"candidates": [
{
"label": "FP32 attention products with FP64 shared sums",
"worker_receipt_sha256": "e5a9ec7c145c1f5802da9bbe912b4646b94017df4ee1a521a93a813aac0906e5",
"source_head": "69dab1d06bdb792309ef230cc737b6d749e3f735",
"checkpoint_sha256": "5020d31211f6d374a8582e48add1426ff801e83d7b74c718fc18eb40edf23235",
"status": "complete",
"numerical_acceptance": false,
"elapsed_seconds": 143.86607869999716,
"comparisons": {
"native_backbone_vjp__hf_backbone_vjp": {
"failed_tensors": 134,
"tensor_count": 290,
"worst_ratio": 45.78206253051758
},
"native_full_task__hf_task": {
"failed_tensors": 194,
"tensor_count": 294,
"worst_ratio": 93.04010009765625
},
"native_batched_task__native_serial_task": {
"failed_tensors": 0,
"tensor_count": 294,
"worst_ratio": 0.36702612042427063
},
"native_serial_task__native_full_task": {
"failed_tensors": 38,
"tensor_count": 294,
"worst_ratio": 7.2799577713012695
},
"native_batched_task__native_full_task": {
"failed_tensors": 38,
"tensor_count": 294,
"worst_ratio": 7.2799577713012695
}
},
"reference_scope": "Fresh six-mode evaluations",
"optimizer_constructed": false
},
{
"label": "FP64 per-call projection parameter products",
"worker_receipt_sha256": "367f9eaecd675bde0a66c7c81dbf7dd54c3144bc70c836300e7b6b7a139a3590",
"source_head": "69dab1d06bdb792309ef230cc737b6d749e3f735",
"checkpoint_sha256": "5020d31211f6d374a8582e48add1426ff801e83d7b74c718fc18eb40edf23235",
"status": "complete",
"numerical_acceptance": false,
"elapsed_seconds": 124.7685586999869,
"comparisons": {
"native_backbone_vjp__hf_backbone_vjp": {
"failed_tensors": 132,
"tensor_count": 290,
"worst_ratio": 46.268577575683594
},
"native_full_task__hf_task": {
"failed_tensors": 194,
"tensor_count": 294,
"worst_ratio": 94.83277893066406
},
"native_batched_task__native_serial_task": {
"failed_tensors": 0,
"tensor_count": 294,
"worst_ratio": 0.30357491970062256
},
"native_serial_task__native_full_task": {
"failed_tensors": 21,
"tensor_count": 294,
"worst_ratio": 5.299079418182373
},
"native_batched_task__native_full_task": {
"failed_tensors": 21,
"tensor_count": 294,
"worst_ratio": 5.299413681030273
}
},
"reference_scope": "Fresh six-mode evaluations",
"optimizer_constructed": false
},
{
"label": "Branch-owned prefix VJPs with bounded parameter groups",
"worker_receipt_sha256": "175b9793ebec2b43bef1f3f7b682776ad8029b844415f92b8c512e75f78657ec",
"source_head": "69dab1d06bdb792309ef230cc737b6d749e3f735",
"checkpoint_sha256": "5020d31211f6d374a8582e48add1426ff801e83d7b74c718fc18eb40edf23235",
"status": "complete",
"numerical_acceptance": false,
"elapsed_seconds": 93.43103430001065,
"comparisons": {
"candidate_batched": {
"native_full_task": {
"failed_tensors": 27,
"worst_ratio": 4.468128204345703,
"tensor_count": 294
},
"native_batched_task": {
"failed_tensors": 19,
"worst_ratio": 5.316349506378174,
"tensor_count": 294
},
"hf_task": {
"failed_tensors": 193,
"worst_ratio": 93.84680938720703,
"tensor_count": 294
}
},
"candidate_serial": {
"native_full_task": {
"failed_tensors": 27,
"worst_ratio": 4.468128204345703,
"tensor_count": 294
},
"native_batched_task": {
"failed_tensors": 19,
"worst_ratio": 5.317352294921875,
"tensor_count": 294
},
"hf_task": {
"failed_tensors": 193,
"worst_ratio": 93.84680938720703,
"tensor_count": 294
},
"candidate_batched": {
"failed_tensors": 0,
"worst_ratio": 0.6569960117340088,
"tensor_count": 294
}
}
},
"reference_scope": "Two fresh candidate modes compared with hash-verified preserved full-row, batched and independent-HF task gradient receipts",
"optimizer_constructed": false
}
],
"localization": {
"layer_trace_summary_sha256": "7cbc6eafa164ebad3004cad13268be81d07cc8d1d4a132832c2a46c78986ea3e",
"all_24_layers_suffix_inputs_and_input_gradients_exact": true,
"layer0_prefix_input_gradient_ratio": 3.737082453179747,
"normalization_replay_sha256": "99a70847cf39da3b7b8c667415688214c7e1204f4340a069f6cd9322721322b6",
"normalization_replay_results": [
{
"norm": "0.input_layernorm",
"combined_residual_input_partition": {
"max_absolute": 0.0006080188668420305,
"max_tolerance_ratio": 3.737082453179747,
"failing_elements": 45
},
"candidate_actual_upstreams_partition": {
"max_absolute": 0.0007520939007008565,
"max_tolerance_ratio": 3.2554632715967573,
"failing_elements": 24
},
"candidate_matched_upstream_partition": {
"max_absolute": 1.3442582712741569e-05,
"max_tolerance_ratio": 0.022825267440992884,
"failing_elements": 0
},
"stock_normalization_only_partition": {
"max_absolute": 0.0007482792016162421,
"max_tolerance_ratio": 3.246066762727333,
"failing_elements": 24
},
"candidate_cached_vs_stock_normalization_only": {
"max_absolute": 3.0517578125e-05,
"max_tolerance_ratio": 0.006123797714794654,
"failing_elements": 0
}
},
{
"norm": "0.post_attention_layernorm",
"combined_residual_input_partition": {
"max_absolute": 0.00018960056513606105,
"max_tolerance_ratio": 8.580650386056718,
"failing_elements": 35
},
"candidate_actual_upstreams_partition": {
"max_absolute": 0.0001907602786559437,
"max_tolerance_ratio": 8.608627607566138,
"failing_elements": 36
},
"candidate_matched_upstream_partition": {
"max_absolute": 6.643240340054035e-06,
"max_tolerance_ratio": 0.03645106463588717,
"failing_elements": 0
},
"stock_normalization_only_partition": {
"max_absolute": 0.0001898066047942848,
"max_tolerance_ratio": 8.56558285196843,
"failing_elements": 36
},
"candidate_cached_vs_stock_normalization_only": {
"max_absolute": 1.52587890625e-05,
"max_tolerance_ratio": 0.016407333909496395,
"failing_elements": 0
}
},
{
"norm": "1.input_layernorm",
"combined_residual_input_partition": {
"max_absolute": 1.860946929355123e-05,
"max_tolerance_ratio": 0.5236419071529433,
"failing_elements": 0
},
"candidate_actual_upstreams_partition": {
"max_absolute": 1.4410558705435506e-05,
"max_tolerance_ratio": 0.5492200963343384,
"failing_elements": 0
},
"candidate_matched_upstream_partition": {
"max_absolute": 1.2946432548233133e-07,
"max_tolerance_ratio": 0.0036842560627586025,
"failing_elements": 0
},
"stock_normalization_only_partition": {
"max_absolute": 1.4439895366535893e-05,
"max_tolerance_ratio": 0.5503381849074498,
"failing_elements": 0
},
"candidate_cached_vs_stock_normalization_only": {
"max_absolute": 5.960464477539063e-08,
"max_tolerance_ratio": 0.0018730159610178924,
"failing_elements": 0
}
},
{
"norm": "1.post_attention_layernorm",
"combined_residual_input_partition": {
"max_absolute": 4.108106637090714e-06,
"max_tolerance_ratio": 0.19100081846493436,
"failing_elements": 0
},
"candidate_actual_upstreams_partition": {
"max_absolute": 3.7928492062633268e-06,
"max_tolerance_ratio": 0.1651266558372655,
"failing_elements": 0
},
"candidate_matched_upstream_partition": {
"max_absolute": 6.960863174754195e-08,
"max_tolerance_ratio": 0.0018746530652766378,
"failing_elements": 0
},
"stock_normalization_only_partition": {
"max_absolute": 3.796574497449967e-06,
"max_tolerance_ratio": 0.16528884180241957,
"failing_elements": 0
},
"candidate_cached_vs_stock_normalization_only": {
"max_absolute": 5.960464477539063e-08,
"max_tolerance_ratio": 0.001989324053746682,
"failing_elements": 0
}
}
],
"correction": "Hooks on normalization input tensors include residual-path contributions. The superseded initial replay compared those combined gradients with normalization-only derivatives. The corrected replay recomputes normalization-only VJPs separately. Local FP64 normalization alone does not repair the actual incoming prefix differences."
},
"claims": {
"production_changed": false,
"complete_native_acceptance": false,
"resume_qualified": false,
"bf16_qualified": false,
"predictive_quality_established": false,
"continuous_resource_compliance_established": false,
"issue3_complete": false
},
"limits": "Synthetic operator checks and two-layer tests do not cover the full 294-parameter gate. The branch-owned prototype preserves one prefix forward and genuine question batching, but fails full-weight equivalence. Its diagnostic execution version makes no resume compatibility claim."
}
23 changes: 23 additions & 0 deletions docs/evidence/native/prefix-gradient-diagnostics-20261002.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,23 @@
# Prefix gradient diagnostics — October 2, 2026

The [portable receipt](prefix-gradient-diagnostics-20261002.json) records three rejected actual-weight candidates. The original [six-mode failure](rank0-derivatives-failed-20261002.md) remains the baseline. Runtime source and frozen numerical criteria were preserved throughout these external diagnostics.

| Candidate | Cached/full failed tensors | Native/HF task failed tensors | Batched/serial failed tensors |
| --- | ---: | ---: | ---: |
| FP32 attention products, FP64 shared sums | 38/294 | 194/294 | 0/294 |
| FP64 per-call projection parameter products | 21/294 | 194/294 | 0/294 |
| Branch-owned prefix VJPs | 27/294 | 193/294 | 0/294 |

The first two candidates ran six fresh modes. The branch-owned candidate ran fresh batched and serial modes and compared all 294 gradients with hash-verified preserved full-row and independent-HF task receipts for the same checkpoint and production source. It is a diagnostic execution version with no resume compatibility claim. Passing small operator tests did not predict full-weight acceptance.

## Localization and correction

In a separate unmodified actual-model trace, suffix inputs and layer-input gradients were exact between serial-cached and full-row execution across all 24 layers. Shared-prefix input gradients differed; layer 0 exceeded the frozen combined tolerance by a ratio of 3.7371. This localizes the observed activation-gradient difference without proving that every parameter-gradient error has the same cause.

Normalization inputs also feed residual paths. Their tensor hooks therefore capture combined gradients, not an isolated normalization VJP. The initial replay comparison was superseded. The corrected normalization-only replay still fails with the actual incoming prefix gradients: layer 0 input normalization has a stock ratio of 3.2461 and an FP64 candidate ratio of 3.2555; post-attention normalization has ratios 8.5656 and 8.6086. Matching incoming gradients makes the local candidate comparisons pass. Improving normalization precision alone is insufficient.

The frozen cache budget groups Noul/Choice together and Score separately: widths 57/135 and flattened suffix projection rows 114/135, plus one 25-token prefix. These are budget-driven groups, not explicit length buckets. Projection parameter-sum rounding and inherited shared-prefix gradient differences were both observed; neither tested intervention passed the complete gate.

## Acceptance status

Production code, existing acceptance flags, the independent reference, and declared tolerances are unchanged. These receipts establish no completed native gate, resume/BF16 qualification, predictive-quality benefit, continuous resource compliance, or issue closure. Original checkpoint provenance, failed candidates and the corrected analysis remain preserved locally. No model artifacts or private local paths are included here.
2 changes: 2 additions & 0 deletions docs/evidence/native/rank0-derivatives-failed-20261002.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,3 +5,5 @@ The [portable receipt](rank0-derivatives-failed-20261002.json) records all six c
CPU metadata verified 294 tensors / 494,090,176 scalars, including 290 backbone tensors / 494,032,768 scalars. The earlier run stopped on an incorrect inventory assertion; that failure and original artifacts remain preserved. Correcting the assertion did not relax numerical criteria.

The checkpoint retains historical stock-optimizer birth provenance; evaluation used the reviewed 33-file runtime. This probe constructs no optimizer and establishes no CPU-state optimizer arithmetic result, quality benefit, or issue closure. Existing acceptance flags remain unchanged.

Later [prefix-gradient diagnostics](prefix-gradient-diagnostics-20261002.md) rejected three additional candidates and document the corrected normalization-only replay. They do not change this baseline or acceptance status.
4 changes: 3 additions & 1 deletion tests/test_stage3_defects.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -324,7 +324,9 @@ def test_real_git_commit_dirty_and_archive_hash(tmp_path):
code=tmp_path/'src/kev_laya';code.mkdir(parents=True);p=code/'a.py';p.write_text('x=1\n')
subprocess.run(['git','init','-q',str(tmp_path)],check=True)
def git(*args):return subprocess.run(['git','-C',str(tmp_path),*args],capture_output=True,check=True,text=True).stdout.strip()
git('add','.');git('-c','user.name=Test','-c','user.email=test@example.invalid','commit','-qm','fixture')
# This fixture tests source capture, independently of a user's signing agent.
git('add','.');git('-c','user.name=Test','-c','user.email=test@example.invalid',
'-c','commit.gpgsign=false','commit','-qm','fixture')
git('remote','add','origin','https://github.com/example/source.git')
source=capture_source(tmp_path)
assert source['commit']==git('rev-parse','HEAD') and source['git_dirty'] is False
Expand Down
Loading