Harden private A100 SFT smoke preflight

[Raidriar7170]

边界

• PR merge base（当前 origin/main）：4c99e8d6cf1fe2782b2760e36437ff5fae164dff
• 本轮 Changes Requested base：41c53b8eceaafef339e8984e861f16f749e85f38
• Head：3a4bf38a574ed81534eaea1103b491b137da3049
• 仅实现 Qwen2.5-7B-Instruct LoRA SFT 的 1-step/1–2 rows smoke 前置与证据门；不授权完整训练、DPO、GRPO、prediction、evaluation、lockbox 或模型改进声明。

Must Fix 实现位置

1. Public result 严格 allowlist

   • src/voice2task/training.py:359 _public_sft_preflight_result
   • src/voice2task/training.py:460 public_training_result
   • src/voice2task/cli/train.py:100 CLI 只输出重建后的单个 JSON。
   • 顶层、preflight section/key/value 均 fail closed；adapter 仅公开 filename/size/SHA-256。

2. Descriptor-relative output claim

   • src/voice2task/training.py:860 _open_bound_output_parent
   • src/voice2task/training.py:1036 _claim_sft_output_directory
   • exact parent 必须预存在；绑定 root/parent dev/ino/uid/gid/mode；使用 O_DIRECTORY/O_NOFOLLOW 与 dir_fd leaf mkdir；前后 identity check；不做 uncertainty pathname cleanup。
   • 用户授权的 threat model：防御所有 checkpoint-observable drift/symlink/inode exchange/concurrent leaf；不声称对同 UID 在最终 checkpoint→mkdirat syscall 窗口内的恶意 rename 具备内核原子性，部署必须保证该 namespace 前提。

3. Shared GPU idle/free-memory gate

   • src/voice2task/training.py:1240 _probe_sft_gpu
   • 记录 free memory、minimum free memory、aggregate compute-process count、idle_verified。
   • 新 blocker：GPU_FREE_MEMORY_INSUFFICIENT、GPU_BUSY、GPU_OCCUPANCY_PROBE_FAILED。
   • src/voice2task/training.py:4552 real runner 入口和权重加载紧前均复检。

4. Assistant-only objective

   • src/voice2task/training.py:3944 assistant indices 独立从 tokenizer offsets/span 构造。
   • src/voice2task/training.py:3998 验证 prompt 全部 -100、assistant 目标非空且连续、labels/input_ids 对齐、三张量等长。

5. Adapter 真实更新证据

   • src/voice2task/training.py:2341 before/after adapter tensor snapshot。
   • src/voice2task/training.py:2379 digest、changed tensor count、finite 证据。
   • src/voice2task/training.py:2401 smoke postconditions 和 adapter 文件 SHA-256。

新增重点回归

• test_public_training_result_is_exact_allowlist_for_arbitrary_absolute_paths
• test_public_training_result_rebuilds_preflight_without_unknown_nested_private_fields
• test_public_training_result_rejects_private_paths_in_every_allowed_string_container
• test_output_claim_blocks_root_same_path_inode_exchange_without_external_mutation
• test_output_claim_blocks_parent_exchange_to_external_symlink_with_zero_external_mutation
• test_output_claim_checks_parent_path_identity_after_open_before_leaf_mkdir
• test_output_claim_rejects_concurrent_final_leaf_creation
• test_gpu_probe_reports_free_memory_and_idle_aggregate
• test_gpu_probe_blocks_insufficient_free_memory_or_busy_device
• test_gpu_probe_fails_closed_when_occupancy_probe_fails
• test_sft_preflight_rejects_one_prompt_label_leak_even_with_other_masked_tokens
• test_real_sft_rechecks_gpu_idle_state_before_model_weight_load
• test_adapter_state_evidence_requires_real_finite_change
• test_smoke_postconditions_require_adapter_update_evidence

OpenSpec

• 主 openspec/specs/supervised-contract-tuning/spec.md 已同步。
• Archived：openspec/changes/archive/2026-07-14-enable-private-a100-sft-smoke-v1/
• openspec list --json：active changes = 0。
• 未增加 active-change failure allowlist。

验证

• PYTHONPATH=src pytest -q → 1413 passed
• focused training suites → 203 passed
• PYTHONPATH=src ruff check . → PASS
• OPENSPEC_TELEMETRY=0 openspec validate --all --strict → 15 passed, 0 failed
• PYTHONPATH=src python scripts/check_current_truth_surface.py → exit 0
• git diff --check → PASS
• targeted mypy 命令仅报告两个未修改文件中的既有 imported-module 诊断：
  • slot_error_analysis.py:574
  • copy_backed_shadow_interface.py:357
    本次触及的 training.py / cli/train.py 无诊断、无新增错误。

Stop hook

本机 ~/.codex/hooks.json 已移除会因缺少 tree-sitter 而 exit 1 的 graph-on-stop.js 子 hook，保留 Hivemind stop.js；使用合法 hook JSON payload 直接验证 stop_hook_exit=0。这是本机配置修复，不属于仓库提交。

Committed-HEAD preflight

Exit code：1。stdout 为单个脱敏 JSON，stderr 无第二个结果：

{
  "blockers": [
    "CONFIG_FILE_MISSING"
  ],
  "config": {},
  "dataset": {},
  "git": {},
  "gpu": {},
  "model": {},
  "objective": {},
  "output": {},
  "ready": false,
  "runtime": {},
  "schema_version": "voice2task-sft-preflight-v1",
  "status": "blocked"
}

• Private config：不存在；精确建议路径受 .gitignore 保护，未提交。
• 真实 A100：未检测（preflight 在 missing config 处 fail closed，未进入 GPU probe）。
• 本地 Qwen2.5-7B-Instruct：未检测（无合法 private model path）。
• 真实 1-step smoke：未运行。
• 唯一真实 blocker：CONFIG_FILE_MISSING。
• 未使用 0.5B、mock、fixture 或 dry-run 冒充真实 7B smoke。

最终状态

• git status --short：空
• git diff --check：PASS
• clean-evaluation truth surface 保持原值，execution_readiness=false。
• 没有执行完整训练、DPO、GRPO、prediction、evaluation，也没有作出模型改进、生产就绪或 clean-evaluation ready 声明。

[Raidriar7170]

A100 SFT smoke runtime follow-up (2026-07-14)

Status: PREFLIGHT_BLOCKED. This is not SMOKE_COMPLETED.

• PR base SHA: 4c99e8d6cf1fe2782b2760e36437ff5fae164dff
• Runtime-fix base SHA: 3a4bf38a574ed81534eaea1103b491b137da3049
• Head SHA: f214b9be551e364742c2b4a07c341f5062dca552
• Branch: codex/enable-private-a100-sft-smoke-v1
• Draft PR remains open; no merge was performed.

Must Fix implementation

• src/voice2task/training.py: _sample_sft_gpu_compute_process_count, _run_sft_gpu_fact_helper, _validated_sft_gpu_helper_result, _probe_sft_gpu
• src/voice2task/_sft_gpu_probe_helper.py: isolated typed CUDA fact protocol and pre-import socket-denying audit hook
• tests/test_sft_gpu_occupancy_helper.py: parent/helper ordering, typed blockers, strict protocol, minimal environment, real isolated child, real socket API denial, cross-field validation
• tests/test_a100_sft_review_hardening.py and tests/test_private_a100_sft_preflight.py: shared probe regression coverage
• openspec/specs/supervised-contract-tuning/spec.md: synced container PID namespace contract

Archived OpenSpec:
openspec/changes/archive/2026-07-14-fix-container-pid-aware-sft-gpu-occupancy-v1

Validation

• Focused GPU/preflight suite: 241 passed
• PYTHONPATH=src pytest -q: 1451 passed in 25.76s
• PYTHONPATH=src ruff check .: PASS
• OPENSPEC_TELEMETRY=0 openspec validate --all --strict: 15 passed, 0 failed
• PYTHONPATH=src python scripts/check_current_truth_surface.py: exit 0
• git diff --check: PASS
• Stop hook with direct {} payload: exit 0
• Targeted mypy: exit 1 only for two pre-existing diagnostics in untouched imported modules (slot_error_analysis.py:574, copy_backed_shadow_interface.py:357); no diagnostic in the touched GPU helper/probe code.

Runtime detection

• Real A100 detected: yes (NVIDIA A100-SXM4-80GB, CC 8.0, BF16 supported, 79.151 GiB total, 78.734 GiB free, two sampled compute-process counts both zero)
• Local Qwen2.5-7B-Instruct identity verified: yes (15,231,271,888-byte four-shard inventory; public identity and geometry matched)
• Dedicated private venv: dependency imports pass; pip check reports no broken requirements
• Private config: present, ignored, untracked, non-symlink
• Real 1-step smoke executed: no
• Unique real blocker: MAX_SEQUENCE_LENGTH_EXCEEDED
• Preflight exit: 1; ready=false; stderr bytes: 0; final run directory was not created
• Stage 2 development screen: not started

Complete sanitized preflight JSON

{"blockers":["MAX_SEQUENCE_LENGTH_EXCEEDED"],"config":{"allow_heavy_training":true,"base_model_public_id":"Qwen/Qwen2.5-7B-Instruct","bf16":true,"config_sha256":"2b0631a1ae11c3bff56313356c633bd844e4182a527bc125d876ff4e6ef3a8a6","dataset_split":"train","dtype_is_bfloat16":true,"fp16":false,"gradient_accumulation_steps":1,"gradient_checkpointing":true,"local_files_only":true,"logging_steps":1,"lora_policy_valid":true,"low_cpu_mem_usage":true,"max_seq_length":1024,"max_steps":1,"max_train_rows":2,"minimum_free_disk_gib":20.0,"per_device_train_batch_size":1,"private_file":{"git_ignored":true,"git_tracked":false,"nonsymlink":true,"under_private_runtime":true},"report_to_empty":true,"save_strategy":"no","seed":42,"tf32":true,"torch_dtype_is_bfloat16":true,"trust_remote_code":false,"use_cache":false},"dataset":{"manifest_file":"manifest_public_sample.json","manifest_id":"public-sample-20260619T090925Z","manifest_sha256":"f866c173795e97953b1dec85611b405867d0a29497910282f99d399f109cda95","non_train_rows_selected":0,"selected_row_count":2,"selected_row_ids_sha256":"0fdcbfe62867b87ededaf2c99a122438cb4eb89ef214c4e02fca703805738409","selected_split":"train","sft_file":"sft_public_sample.jsonl","sft_sha256":"4b677420f766555c04199f15f69f41f3b3ad36ad3cd5c33d2b40b0e3f8573587"},"git":{"commit_sha":"f214b9be551e364742c2b4a07c341f5062dca552","tracked_worktree_clean":true},"gpu":{"bf16_supported":true,"compute_capability":"8.0","compute_process_count":0,"cuda_version":"12.9","explicit_selection":true,"free_memory_gib":78.734,"idle_verified":true,"minimum_free_memory_gib":35.0,"minimum_memory_gib":35.0,"name":"NVIDIA A100-SXM4-80GB","total_memory_gib":79.151,"visible_device_count":1},"model":{"geometry_matches_qwen2_5_7b":true,"local_files_only":true,"minimum_weight_bytes":12884901888,"public_id":"Qwen/Qwen2.5-7B-Instruct","snapshot_revision_sha256":"2a1aaaf30e79bc9817f385e41260f822b4533f570a62f7253afc536e8f724363","stable_fingerprints":{"config.json":"7463bb0ea78315365e6c6b74de4e73bbcc8359dfb0c5a737584e077d42c0b03c","generation_config.json":"3a8f9087e486054c8a4a08dae2e5a3ba62e23da212b5b8c08bc42cb983c3459f","merges.txt":"599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3","model.safetensors.index.json":"624bf7c47cd12468fdc16e38a47cf4f19e0415b859a223ba3c027eed2f0e1028","tokenizer.json":"c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539","tokenizer_config.json":"5b5d4f65d0acd3b2d56a35b56d374a36cbc1c8fa5cf3b3febbbfabf22f359583","vocab.json":"ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"},"total_weight_bytes":15231271888,"weight_inventory":[{"name":"model-00001-of-00004.safetensors","size":3945441440},{"name":"model-00002-of-00004.safetensors","size":3864726352},{"name":"model-00003-of-00004.safetensors","size":3864726424},{"name":"model-00004-of-00004.safetensors","size":3556377672}]},"objective":{"assistant_target_present":false,"max_sequence_length":1024,"prompt_labels_masked":false,"records_checked":0},"output":{"blockers":[],"free_disk_gib":10142733.748,"minimum_free_disk_gib":20.0,"output_path_sha256":"6d5c0189521d027b29cdb81e4a139a419519312c4cd9a1ae6ac32630fde3b368","ready":true,"root_path_sha256":"c26d3afb706ca1995065c67df889a9c850ddf18072679c265574ec6bcfca84da","writable":true},"ready":false,"runtime":{"missing":[],"pip_check":"ok","python":"3.11.4","python_requirement":">=3.10","versions":{"accelerate":"1.13.0","datasets":"5.0.0","peft":"0.18.1","torch":"2.10.0+cu129","transformers":"5.5.0","trl":"1.5.1"}},"schema_version":"voice2task-sft-preflight-v1","status":"blocked"}

Repository status after push is clean and synchronized with the remote branch. No full training, DPO, GRPO, prediction, evaluation, lockbox access, or model-improvement claim was performed.

[Raidriar7170]

Real 1-step runtime evidence — postcondition failure (2026-07-15)

Authoritative external status: RUNTIME_POSTCONDITION_FAILED. Do not describe this run as SMOKE_COMPLETED.

The earlier MAX_SEQUENCE_LENGTH_EXCEEDED root cause was measured with the exact local tokenizer and shared assistant-only builder:

• selected row 1: 1077 total tokens (1021 prompt, 56 assistant)
• selected row 2: 1076 total tokens (1020 prompt, 56 assistant)
• ignored private config was changed only from max_seq_length=1024 to the minimum safe value 1077

A fresh shared preflight returned exit 0, ready=true, blockers=[], checked both rows, and reported maximum_observed_tokens=1077.

Exactly one real sft --run-training process was then launched on the same explicitly selected GPU. It returned exit 0 and produced real adapter-update evidence:

{"adapter_files":[{"name":"README.md","sha256":"088fc697ae52cb14f1e808e5c967c8be13cb0256da184ad882dd41ba8e95c617","size":5262},{"name":"adapter_config.json","sha256":"5212eea727806033f85531cabb70ca59f6c93152aa3f783c543e169314644760","size":1029},{"name":"adapter_model.safetensors","sha256":"13fce8410c1c21d86c1dc628c9f20e9477294c707d1fe7235e4a111834a8fab7","size":20214760}],"clean_evaluation":{"acquisition_source_status":"UNAVAILABLE","authoritatively_bound_binding_count":0,"clean_population_status":"NOT_MATERIALIZED","execution_readiness":false,"freeze_authorized":false,"human_acceptance_status":"NOT_RECORDED","protocol_freeze_status":"NOT_FROZEN"},"observed_optimizer_steps":1,"smoke_status":"SMOKE_COMPLETED","stage":"sft","train_result_metrics":{"adapter_state_digest_after":"a2395bb971e32f30cf760f2da2fc9b06e727c82a2ffac6b753ac7b0b21f7105d","adapter_state_digest_before":"ba19fe10ce78816dbb3e3944dda2ae6d3e354277ad0ce1cad58279ec1ed68a52","adapter_tensor_count":224,"all_adapter_tensors_finite":true,"changed_adapter_tensor_count":112,"epoch":0.5,"train_loss":0.0034358606208115816,"train_runtime":0.9902,"train_samples_per_second":1.01,"train_steps_per_second":1.01,"trainable_parameter_count":5046272},"training_budget":{"configured_max_steps":1,"effective_batch_size":1,"gradient_accumulation_steps":1,"num_train_epochs":1.0,"observed_optimizer_steps":1,"per_device_train_batch_size":1,"scheduler_max_steps":1,"step_matched_not_token_matched":true,"step_matching_unit":"optimizer_steps","target_tokens_per_single_pass":112,"target_tokens_seen_estimate":56,"target_tokens_seen_status":"estimated_from_label_tokens_and_step_budget","theoretical_examples_seen":1,"train_row_count":2},"training_rows_used":2,"training_status":"training_completed"}

Additional verified postconditions:

• adapter metadata exists and records the same update evidence
• before/after adapter digests differ
• 112 adapter tensors changed; all adapter tensors are finite
• finite loss; no NaN/Inf metric
• no full base-model weight file was written
• selected GPU compute-process count returned to zero
• one launch marker exists; no second run was attempted
• tracked worktree remains clean; private config, logs, cache, metadata and adapter remain ignored/untracked
• offline/local-only environment was used; no prediction, evaluation, DPO, GRPO or full training was run

Blocking postcondition

CLI stdout was not one JSON document:

• stdout: 7054 bytes / 197 lines
• final result JSON prefix: 298 non-JSON bytes
• prefix consists of two Trainer Python mapping repr lines (step metrics and train summary)
• one and only one JSON object contains training_status
• stderr contains ordinary logs and no second result JSON

Therefore the embedded metadata value smoke_status=SMOKE_COMPLETED is not accepted as the external status. Independent review verdict: Changes Requested / Runtime Postcondition Failed.

Sanitized ready preflight JSON

{"blockers":[],"config":{"allow_heavy_training":true,"base_model_public_id":"Qwen/Qwen2.5-7B-Instruct","bf16":true,"config_sha256":"1a3932231d9c6df01167024bbfaf81ad549bf20ac672c15c5ee5b4a7b889f734","dataset_split":"train","dtype_is_bfloat16":true,"fp16":false,"gradient_accumulation_steps":1,"gradient_checkpointing":true,"local_files_only":true,"logging_steps":1,"lora_policy_valid":true,"low_cpu_mem_usage":true,"max_seq_length":1077,"max_steps":1,"max_train_rows":2,"minimum_free_disk_gib":20.0,"per_device_train_batch_size":1,"private_file":{"git_ignored":true,"git_tracked":false,"nonsymlink":true,"under_private_runtime":true},"report_to_empty":true,"save_strategy":"no","seed":42,"tf32":true,"torch_dtype_is_bfloat16":true,"trust_remote_code":false,"use_cache":false},"dataset":{"manifest_file":"manifest_public_sample.json","manifest_id":"public-sample-20260619T090925Z","manifest_sha256":"f866c173795e97953b1dec85611b405867d0a29497910282f99d399f109cda95","non_train_rows_selected":0,"selected_row_count":2,"selected_row_ids_sha256":"0fdcbfe62867b87ededaf2c99a122438cb4eb89ef214c4e02fca703805738409","selected_split":"train","sft_file":"sft_public_sample.jsonl","sft_sha256":"4b677420f766555c04199f15f69f41f3b3ad36ad3cd5c33d2b40b0e3f8573587"},"git":{"commit_sha":"f214b9be551e364742c2b4a07c341f5062dca552","tracked_worktree_clean":true},"gpu":{"bf16_supported":true,"compute_capability":"8.0","compute_process_count":0,"cuda_version":"12.9","explicit_selection":true,"free_memory_gib":78.734,"idle_verified":true,"minimum_free_memory_gib":35.0,"minimum_memory_gib":35.0,"name":"NVIDIA A100-SXM4-80GB","total_memory_gib":79.151,"visible_device_count":1},"model":{"geometry_matches_qwen2_5_7b":true,"local_files_only":true,"minimum_weight_bytes":12884901888,"public_id":"Qwen/Qwen2.5-7B-Instruct","snapshot_revision_sha256":"2a1aaaf30e79bc9817f385e41260f822b4533f570a62f7253afc536e8f724363","stable_fingerprints":{"config.json":"7463bb0ea78315365e6c6b74de4e73bbcc8359dfb0c5a737584e077d42c0b03c","generation_config.json":"3a8f9087e486054c8a4a08dae2e5a3ba62e23da212b5b8c08bc42cb983c3459f","merges.txt":"599bab54075088774b1733fde865d5bd747cbcc7a547c5bc12610e874e26f5e3","model.safetensors.index.json":"624bf7c47cd12468fdc16e38a47cf4f19e0415b859a223ba3c027eed2f0e1028","tokenizer.json":"c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539","tokenizer_config.json":"5b5d4f65d0acd3b2d56a35b56d374a36cbc1c8fa5cf3b3febbbfabf22f359583","vocab.json":"ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910"},"total_weight_bytes":15231271888,"weight_inventory":[{"name":"model-00001-of-00004.safetensors","size":3945441440},{"name":"model-00002-of-00004.safetensors","size":3864726352},{"name":"model-00003-of-00004.safetensors","size":3864726424},{"name":"model-00004-of-00004.safetensors","size":3556377672}]},"objective":{"assistant_target_present":true,"max_sequence_length":1077,"maximum_observed_tokens":1077,"prompt_labels_masked":true,"records_checked":2},"output":{"blockers":[],"free_disk_gib":10142325.843,"minimum_free_disk_gib":20.0,"output_path_sha256":"3016709988fdcbd111576e8e41a21782e357575052879c9897de01cd98ad541b","ready":true,"root_path_sha256":"c26d3afb706ca1995065c67df889a9c850ddf18072679c265574ec6bcfca84da","writable":true},"ready":true,"runtime":{"missing":[],"pip_check":"ok","python":"3.11.4","python_requirement":">=3.10","versions":{"accelerate":"1.13.0","datasets":"5.0.0","peft":"0.18.1","torch":"2.10.0+cu129","transformers":"5.5.0","trl":"1.5.1"}},"schema_version":"voice2task-sft-preflight-v1","status":"ready"}

Stage 2 remains blocked and was not started. Fixing stdout and authorizing any second real smoke are separate future decisions.

[Raidriar7170]

Bounded stdout-contract fix closeout

• Fix base: f214b9be551e364742c2b4a07c341f5062dca552
• New head: df25648d1d837d7aa4af7757b9aea68783d2a5b6
• Branch: codex/enable-private-a100-sft-smoke-v1
• PR remains Draft; no merge was performed.

Must Fix implementation

• Line-aware non-JSON stderr diagnostic stream: src/voice2task/cli/train.py:23-44
• Unified backend-dispatch stdout isolation with final success/failure JSON outside the redirect: src/voice2task/cli/train.py:87-146
• Real observed Trainer mapping, noisy exception, partial JSON, and mock-only DPO regressions: tests/test_a100_sft_review_hardening.py:122-207
• Synced one-result contract: openspec/specs/supervised-contract-tuning/spec.md:154-171
• Human Brief: docs/human-briefs/2026-07-15-fix-real-sft-cli-single-json-v1.html
• Archived OpenSpec: openspec/changes/archive/2026-07-15-fix-real-sft-cli-single-json-v1/
• Active OpenSpec changes: 0

TDD and review

• RED: 3/3 new regressions failed on the unmodified CLI because backend output contaminated stdout.
• GREEN focused: 3 passed.
• Combined A100 review-hardening + private-preflight regression: 154 passed.
• Independent read-only review: Pass; Must Fix 0; Should Fix 0; Re-plan not needed.
• A real tqdm(..., file=sys.stdout) compatibility path was exercised without training.

Final validation

• PYTHONPATH=src pytest -q: 1454 passed
• PYTHONPATH=src ruff check .: PASS
• targeted Mypy: no diagnostics in training.py or cli/train.py; only the two recorded imported-module baseline diagnostics remain
• OPENSPEC_TELEMETRY=0 openspec validate --all --strict: 15/15 PASS
• PYTHONPATH=src python scripts/check_current_truth_surface.py: exit 0
• formal public validation: 696 SFT rows / 2100 DPO pairs, zero failures
• DPO pair check: 2100
• dataset/schema/DPO/evaluator focused suite: 88 passed
• public leak scan: 0 findings
• Human Brief links: 5/5
• private-config guards: ignored, untracked, scoped status clean
• direct Stop-hook probes: valid input exit 0; fatal-parse catch exit 0
• git diff --check: PASS
• git status --short: empty after commit

Runtime truth boundary

No second real smoke was run in this fix. The prior authorized one-step process still remains Runtime Postcondition Failed because its complete stdout was not a single JSON document. Although the prior run produced genuine one-step adapter-update evidence, this PR does not describe it as SMOKE_COMPLETED; the repaired protocol has not been revalidated by a separately authorized second real run.

No new GPU/model probe or preflight was run. The prior sanitized preflight and one-run evidence remain in the earlier audit comment: #3 (comment)

Clean-evaluation truth remains unchanged: source unavailable, authoritative binding count 0, human acceptance not recorded, protocol not frozen, clean population not materialized, freeze_authorized=false, and execution_readiness=false.

No full training, DPO, GRPO, prediction, evaluation, lockbox access, or model-improvement claim was performed.

[Raidriar7170]

Real 1-step A100 SFT smoke rerun after CLI stdout fix (2026-07-15)

Authoritative status: SMOKE_COMPLETED.

• Executable code commit: df25648
• Sanitized evidence commit: d86475b
• Branch: codex/enable-private-a100-sft-smoke-v1
• PR remains Draft/Open; no merge or archive was performed.
• OpenSpec change remains active: rerun-real-a100-sft-smoke-after-cli-fix-v1

Fresh gates before the only launch

• Remote tracked checkout clean at the executable commit; CLI stdout isolation present.
• Private config: ignored, untracked, non-symlink.
• Local-only Qwen/Qwen2.5-7B-Instruct identity and inventory verified: 15,231,271,888 total weight bytes; snapshot revision SHA-256 2a1aaaf30e79bc9817f385e41260f822b4533f570a62f7253afc536e8f724363.
• Exactly one A100-SXM4-80GB selected explicitly: 0 compute processes; 78.734 GiB free against a 35 GiB threshold.
• Fresh output identity gate passed.
• Current formal manifest public-sample-20260619T090925Z, train split, 2 rows, max_steps=1.
• Shared preflight: exit 0; stdout one JSON document; stderr empty; ready=true; blockers=[].

Only authorized launch

• Launch count: 1; no retry.
• Process exit: 0.
• Stdout: 6,758 bytes and exactly one complete JSON document.
• Stderr: 1,573 bytes; 0 whole-document JSON results and 0 JSON lines; no competing result.
• training_status=training_completed.
• smoke_status=SMOKE_COMPLETED.
• observed_optimizer_steps=1; training_rows_used=2.
• changed_adapter_tensor_count=112 of 224; all_adapter_tensors_finite=true; before/after adapter digests differ.
• Adapter config: 1,029 bytes; adapter weights: 20,214,760 bytes.
• Complete run tree: 4 files / 20,234,578 bytes; 0 base-shard matches and 0 suspicious full-weight files.

Independent review and validation

• Read-only Reviewer: Pass; Must Fix 0; Should Fix 0.
• Focused CLI/preflight/smoke tests: 206 passed.
• No-training dataset/schema/DPO contract tests: 30 passed.
• OpenSpec strict validation: 16/16 passed.
• Public leak scan: 0 findings.
• git diff --check: passed.

No prediction, evaluation, DPO, GRPO, public test, lockbox, clean-evaluation population access, extra seed, or metric experiment was run. This is bounded infrastructure-smoke evidence only; it is not a quality, held-out, generalization, safety, or production-readiness result.