diff --git a/AGENTS.md b/AGENTS.md index e55f1c642..a405f9324 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -10,3 +10,5 @@ - Current E1 scope follows the standing continue-work authorization relayed in Slack `1794651141.035699` and Claude caps `1794651169.320839`: bounded CPU-only label-free preparation and synthetic checks may proceed after the count checkpoint, without another user approval loop. Total 16 CPU-hours / 25 GiB, at most 8 processes, per-job 30 min / 1.25 GiB tree RSS, at most 2000 registration pairs; the 10-part stage additionally stays sequential within 3 h / 5 GiB. Never read labels, fit models, use GPU, run pilot/full matrix, or exceed caps in this scope. See `docs/s2-e1-budget-checkpoint-2026-11-14.md`. - Post-witness update (2026-11-14 21:12) supersedes treating the preceding E1-only scope as the entire remaining project scope. Follow the latest section of `docs/collaboration-stage-review-policy.md` and Claude Slack `1794661944.308269`: custody join, label-free copy calibration, k=3 per-route candidate count, bounded registration (up to 4000 pairs / 100 worker CPU-h / 16 workers), band review, component freeze, then coordinated S2 continuation. No repeated approval gates or reruns of completed review; preserve concrete execution limits and distinguish future stages from completed work. User preference: tens of CPU-hours alone are not a reason to block progress. No additional scientific task is implied by this preference. + +- S2 execution update (2026-11-15): components/splits and the fixed ten-part measurement are complete; ten-part 10/10 PASS is in `docs/s2-ten-part-results-2026-11-15.md`. The 499 non-test preparation queue has started. Read `docs/s2-pilot-live-handoff-2026-11-15.md` and existing receipts before any launch; never restart from chat idle. Claude accepted implementation fixes at `1794683016.134929`, accepted the frozen ID-order dispatch at `1794683049.057259`, and raised the first-50 continuation ruling to 80 worker CPU-h / 100 GiB at `1794682712.505709`. The running original supervisor still enforces its pinned 60-hour cap; any required continuation may contain only never-attempted parts and must preserve cumulative accounting. After all 499 terminal attempts and required evidence, the authorized single-GPU pilot includes an E256 timing probe and four runs within 2 GPU-hours total, per-node train/validation extrapolation and a frozen common epoch limit. Keep 125 test labels sealed; no repeated user-go or main-review round is required for these accepted fixes. This does not authorize the full training matrix. diff --git a/artifacts/s2-pilot-inputs-2026-11-15/batch-started.json b/artifacts/s2-pilot-inputs-2026-11-15/batch-started.json new file mode 100644 index 000000000..10be6a9b2 --- /dev/null +++ b/artifacts/s2-pilot-inputs-2026-11-15/batch-started.json @@ -0,0 +1,5 @@ +{ + "started_unix": 1794682970.6834989, + "pid": 2, + "contract_sha256": "f464b1ed962c880794b658cd81dd9ab4cc3ebb7df255bd603f463b883ffad8b7" +} diff --git a/artifacts/s2-pilot-inputs-2026-11-15/public/dispatch-ruling-accepted.json b/artifacts/s2-pilot-inputs-2026-11-15/public/dispatch-ruling-accepted.json new file mode 100644 index 000000000..31c37e13f --- /dev/null +++ b/artifacts/s2-pilot-inputs-2026-11-15/public/dispatch-ruling-accepted.json @@ -0,0 +1,35 @@ +{ + "status": "ID_ORDER_ACCEPTED_KEEP_UNCHANGED", + "ruling": "1794682712.505709", + "ruling_received_after_launch": true, + "dispatch": "fixed ascending case-ID lexical order; not node-size sorted; not claimed stratified", + "first50_nodes_quantiles": { + "min": 6564.0, + "median": 25204.5, + "p90": 52040.50000000001, + "max": 145152.0 + }, + "all_nodes_quantiles": { + "min": 4607.0, + "median": 26829.0, + "p90": 50997.99999999999, + "max": 187345.0 + }, + "first50_tertile_counts": [ + 19, + 17, + 14 + ], + "first50_nodes": 1512476, + "first50_mean": 30249.52, + "all_mean": 30448.725450901802, + "original_active_cap_worker_cpu_hours": 60, + "new_ruling_worker_cpu_hours": 80, + "original_contract_unchanged": true, + "original_code_pins_unchanged": true, + "restarts": 0, + "dispatch_acceptance": "1794683049.057259", + "implementation_spotcheck": "1794683016.134929", + "first50_node_share": 0.09954485723691736, + "continuation_rule": "If original 60h supervisor stops at first50 projection in (60,80]h, freeze a continuation for never-attempted parts only; preserve total accounting; no retries; >80h or >100GiB stops" +} diff --git a/artifacts/s2-pilot-inputs-2026-11-15/public/dispatch-ruling-received.json b/artifacts/s2-pilot-inputs-2026-11-15/public/dispatch-ruling-received.json new file mode 100644 index 000000000..6d63374d5 --- /dev/null +++ b/artifacts/s2-pilot-inputs-2026-11-15/public/dispatch-ruling-received.json @@ -0,0 +1,31 @@ +{ + "status": "RUNNING_ORIGINAL_FROZEN_CONTRACT_NEW_RULING_REPORTED", + "ruling": "1794682712.505709", + "ruling_received_after_launch": true, + "dispatch": "fixed ascending case-ID lexical order; not node-size sorted; not claimed stratified", + "first50_nodes_quantiles": { + "min": 6564.0, + "median": 25204.5, + "p90": 52040.50000000001, + "max": 145152.0 + }, + "all_nodes_quantiles": { + "min": 4607.0, + "median": 26829.0, + "p90": 50997.99999999999, + "max": 187345.0 + }, + "first50_tertile_counts": [ + 19, + 17, + 14 + ], + "first50_nodes": 1512476, + "first50_mean": 30249.52, + "all_mean": 30448.725450901802, + "original_active_cap_worker_cpu_hours": 60, + "new_ruling_worker_cpu_hours": 80, + "original_contract_unchanged": true, + "original_code_pins_unchanged": true, + "restarts": 0 +} diff --git a/artifacts/s2-pilot-inputs-2026-11-15/public/first50-barrier-observed.json b/artifacts/s2-pilot-inputs-2026-11-15/public/first50-barrier-observed.json new file mode 100644 index 000000000..1a390280a --- /dev/null +++ b/artifacts/s2-pilot-inputs-2026-11-15/public/first50-barrier-observed.json @@ -0,0 +1,25 @@ +{ + "observed_unix": 1794683565.68047, + "progress": { + "completed": 43, + "active": [ + 6, + 7, + 31, + 34, + 47, + 48, + 49 + ], + "next_index": 50, + "first50_complete": false, + "worker_cpu_seconds": 5315.484242, + "live_guard_and_worker_cpu_seconds": 2758.26, + "active_tree_rss_bytes": 2200363008, + "disk_bytes": 543756288, + "elapsed_seconds": 594.4620338910026, + "stop_reason": null + }, + "first51st_job_directory_exists": false, + "observation": "All initial 50 dispatched; remaining work waits for their terminal barrier" +} diff --git a/artifacts/s2-pilot-inputs-2026-11-15/public/gpu-readonly-inventory.json b/artifacts/s2-pilot-inputs-2026-11-15/public/gpu-readonly-inventory.json new file mode 100644 index 000000000..a6e37b93f --- /dev/null +++ b/artifacts/s2-pilot-inputs-2026-11-15/public/gpu-readonly-inventory.json @@ -0,0 +1,35 @@ +{ + "checked_unix": 1794683192.7040098, + "purpose": "Read-only device inventory; recheck idle before actual pilot", + "gpu_computation_started": false, + "devices": [ + { + "index": 0, + "uuid": "GPU-e8853ae7-a302-6a17-bcfa-aa767e9ed20c", + "name": "Tesla V100-SXM2-16GB", + "memory_used_MiB": 5, + "utilization_percent": 0 + }, + { + "index": 1, + "uuid": "GPU-f1c9fd5a-e917-56e0-21fb-24035ba44a36", + "name": "Tesla V100-SXM2-16GB", + "memory_used_MiB": 5, + "utilization_percent": 0 + }, + { + "index": 2, + "uuid": "GPU-b8c11c01-c91f-aa0b-abf6-dad35085cc76", + "name": "Tesla V100-SXM2-16GB", + "memory_used_MiB": 5, + "utilization_percent": 0 + }, + { + "index": 3, + "uuid": "GPU-8b71b032-1b75-2891-935c-ce106a53034a", + "name": "Tesla V100-SXM2-16GB", + "memory_used_MiB": 5, + "utilization_percent": 0 + } + ] +} diff --git a/docs/s2-pilot-live-handoff-2026-11-15.md b/docs/s2-pilot-live-handoff-2026-11-15.md new file mode 100644 index 000000000..a70602880 --- /dev/null +++ b/docs/s2-pilot-live-handoff-2026-11-15.md @@ -0,0 +1,17 @@ +# S2 当前执行接续记录 + +十件原队列已完整结束:10/10 PASS,正式交付见s2-ten-part-results-2026-11-15.md。代码/证据固定提交cdb52988c3bc915257478ff49eb131d4f49032c4已推送。Claude 1794683016.134929抽查认可修正;不另设主审/用户审批轮。 + +499非测试准备已于2026-11-15 03:02:50(Asia/Shanghai)启动;exec session16682。根目录artifacts/s2-pilot-inputs-2026-11-15。只通过现存progress.json、jobs回执、batch-complete.json及锁核状态,不因聊天idle或其他PID命名空间看不到进程而重启。已启动标记和契约不可覆盖。 + +固定外折0、seed0,387学习+112验证,125测试封存。原运行源码/契约pin继续按60worker CPU-h、16workers、100GiB,每件30min/1.25GiB、首50终态屏障执行;原始契约SHA f464b1ed962c880794b658cd81dd9ab4cc3ebb7df255bd603f463b883ffad8b7。 + +Claude 1794682712.505709补充将检查点继续上限设为80worker CPU-h;该补充在队列启动后才读取。1794683049.057259已明确接受现有启动前冻结的case-ID字典序,不调整。首50节点份额和相对全池分布已追加在public/dispatch-ruling-accepted.json。 + +若首50投影≤60h,原队列自主继续;若(60,80]h,原监督会停在屏障,再冻结只含从未尝试件的续队列,合并全部历史CPU账本,仍≤80h/100GiB;>80h或磁盘超限停止报告。不得修改正在运行的pin,不重试已尝试/失败件、不替补。继续上限变更只记录追加裁量,不能假称原监督已改成80。 + +待499件全部终态,且无未尝试项/全局停止,先固定准备结果和失败清单;再用prepare_s2_pilot_training.py生成成功清单和pilot-plan。GPU UUID要在实际启动前再次检查空闲;不把read-only库存快照当启动许可或当前空闲保证。 + +pilot仅通过run_s2_pilot.py启动:E256 20件学习几何合成目标计时probe≤5min;训练和FP32推理各自按节点总量外推,原学习11,863,762节点/验证3,330,152节点;共同E_max=min(200,floor(1500/epoch秒))且cosine一致。probe和设置计入单卡四run总7200s,各run≤1800s。probe/任意run失败停止、不重试。四个run全部成功才选一次全局宽。 + +运行入口需要可见V100的主机环境;普通受限环境无法可靠读取GPU。用户已授权原计算范围与可控Git/Slack分享,不再要求重复授权。主分工保持:Claude整合/裁量,MeltAstra执行证据;用户休息勿扰。 diff --git a/tools/summarize_s2_pilot_checkpoint.py b/tools/summarize_s2_pilot_checkpoint.py new file mode 100644 index 000000000..e99967299 --- /dev/null +++ b/tools/summarize_s2_pilot_checkpoint.py @@ -0,0 +1,45 @@ +"""Publish the fixed first-50 evidence without changing the running queue.""" +import argparse +import json +import math +from pathlib import Path +import numpy as np +from freeze_g1_components import sha +from run_s2_pilot_preparation import projection + + +def summarize(root): + c=json.loads((root/'contract.json').read_text());checkpoint=json.loads((root/'first-50-checkpoint.json').read_text()) + assert checkpoint['completed']==50 + jobs=[] + for i in range(50): + p=root/'jobs'/f'job-{i:04d}'/'guard-result.json';r=json.loads(p.read_text()) + jobs.append(dict(index=i,case=c['cases'][i]['case'],receipt_sha256=sha(p),**r)) + projected=projection(c,jobs) + assert math.isclose(projected['projected_worker_cpu_seconds'],checkpoint['projected_worker_cpu_seconds'],rel_tol=1e-12,abs_tol=1e-8) + nodes=np.array([s['tet_nodes'] for s in c['cases']]);cuts=np.quantile(nodes,[1/3,2/3]) + def distribution(a): + return dict(parts=len(a),nodes=int(a.sum()),mean=float(a.mean()), + quantiles=dict(zip(['min','p10','median','p90','max'],np.quantile(a,[0,.1,.5,.9,1]).tolist())), + full_pool_tertile_counts=np.bincount(np.searchsorted(cuts,a),minlength=3).tolist()) + hours=projected['projected_worker_cpu_seconds']/3600;disk=checkpoint['projected_disk_bytes']/2**30 + if hours>80 or disk>100:decision='STOP_AND_REPORT_OVER_NEW_CAP' + elif hours>60:decision='ORIGINAL_QUEUE_STOPS_FREEZE_NEVER_ATTEMPTED_CONTINUATION_WITHIN_80H' + else:decision='ORIGINAL_QUEUE_CONTINUES_WITHIN_60H' + report=dict(status='FIRST_50_TERMINAL_REVIEW',dispatch='frozen ascending case-ID order; unchanged', + contract_sha256=sha(root/'contract.json'),checkpoint_sha256=sha(root/'first-50-checkpoint.json'), + fixed_first50=distribution(nodes[:50]),full_pool=distribution(nodes),first50_node_share=float(nodes[:50].sum()/nodes.sum()), + tertile_cut_nodes=cuts.tolist(),jobs=jobs,failed=sum(r['worker']['status']!='PASS' for r in jobs), + worker_cpu_seconds=sum(r['worker']['cpu_seconds'] for r in jobs),guard_cpu_seconds=sum(r['supervisor_cpu_seconds'] for r in jobs), + peak_individual_worker_tree_rss_bytes=max(r['worker']['peak_tree_rss_bytes'] for r in jobs), + projection=projected,projected_worker_cpu_hours=hours,projected_gib=disk, + original_supervisor_cap_cpu_hours=60,new_ruling_cap_cpu_hours=80,new_ruling='1794682712.505709', + unchanged_order_and_continuation_ruling='1794683049.057259',decision=decision, + running_contract_unchanged=True,retry=False,replacement=False,test_labels_read=False,GPU_training_started=False) + with (root/'public/first50-review.json').open('x') as f:json.dump(report,f,indent=2);f.write('\n') + print(json.dumps({k:v for k,v in report.items() if k!='jobs'})) + return report + + +if __name__=='__main__': + p=argparse.ArgumentParser();p.add_argument('root',type=Path);a=p.parse_args();summarize(a.root)