# Public copy of the experiment's timestamped pre-registration and deviation log. Internal session identifiers, account and plan # details, and notes about the publishing workflow were removed or reworded; experimental content (rules, hashes, results, # deviations and audits) is unchanged. 2026-11-10T18:44:54 tasks picked (seed 20261111, stratified 3/3/2/2): [["libero_spatial", 2], ["libero_spatial", 4], ["libero_spatial", 7], ["libero_object", 0], ["libero_object", 6], ["libero_object", 9], ["libero_goal", 7], ["libero_goal", 9], ["libero_10", 3], ["libero_10", 5]] 2026-11-10T18:47:34 clean rooms built: ~/cap-cleanroom (code policy) and ~/wsn-cleanroom (weight-space network); manifests in audit/manifest_*.sha256 2026-11-10T18:53:28 harness frozen: eval_policy.py sha256=e65ff6bbc3f4c53e…; official = LIBERO standard init states 0-49 per task, 10 settle steps, suite max steps; dev feedback = 2 rollouts/task with env seeds 90000+100*task+{0,1}; images flipped to row0=top. Sessions started (Fable 5.2, high). 2026-11-10T21:00:20 wsn v1 frozen (network.npz b7c7c6f2…, 9 layers, 5189 nnz); validator PASS re-run by experimenter; audit log clean. Starting dev feedback rollouts and official v1 eval. 2026-11-10T21:05:02 wsn dev feedback delivered (20 rollouts in ~/wsn-cleanroom/feedback; 16/20 succeeded); phase-2 message sent. 2026-11-10T21:12:35 cap v1 frozen (policy.py e5720410…? see AUDIT); validator PASS re-run; code reviewed for os/subprocess/file-write calls: none; audit log clean. Starting dev feedback rollouts. 2026-11-10T21:16:28 cap dev feedback delivered (20 rollouts; 18/20 succeeded); phase-2 message sent; starting official cap v1 eval. 2026-11-10T21:41:39 wsn v2 frozen (network.npz e2a5238f…, 10 layers, 66716 nnz); validator PASS re-run; audit log clean; starting official wsn v2 eval. 2026-11-10T21:53:43 v1 official results: cap 468/500 (93.6%, Wilson 91.1-95.4); wsn 323/500 (64.6%, 60.3-68.7); paired McNemar cap-only 156 vs wsn-only 11, p=5.8e-34. Code policy reads only its own data.json (checked). 2026-11-10T22:07:52 wsn v2 official: 408/500 (81.6%, Wilson 78.0-84.8); vs wsn v1 paired: v2-only 93, v1-only 8, p=1.7e-19. 2026-11-10T22:08:52 cap v2 frozen (policy.py 2e2272e1…); v1 hashes unchanged; validator PASS re-run; reads only its own data.json; audit log clean; starting official cap v2 eval. 2026-11-10T22:27:17 cap v2 official: 486/500 (97.2%). All four cells scored. RESULTS.md written. 2026-11-10T22:40:58 Astra arm prepared: rooms ~/cap-cleanroom-astra, ~/wsn-cleanroom-astra (TASK+interface+materials byte-identical to Fable rooms); Docker image astra-cleanroom (codex-cli 0.167.1, numpy 2.5.4, pillow 13.0.0); Codex sandbox workspace-write, no network (verified), web search + browser + apps + plugins disabled; prompts = Fable's messages verbatim; same dev seeds and hidden eval. 2026-11-10T22:48:57 Astra cap phase 1 started (gpt-6-astra, effort high, codex exec, container astra_cap_p1). 2026-11-10T22:52:42 Astra wsn phase 1 started (separate Codex home/login). 2026-11-10T23:00:47 Astra cap v1 frozen (policy.py 365240fc…) after ~7 min; validator PASS re-run; reads only own scene_data.json; starting dev rollouts (same seeds). 2026-11-10T23:03:31 Astra cap dev feedback delivered (3/20 succeeded); phase 2 started (codex exec resume, verbatim Fable phase-2 message). 2026-11-10T23:08:56 Astra wsn v1 frozen (network.npz 66c0acd9…, 27 layers, 215364 nnz) after ~14 min; validator PASS re-run; starting dev rollouts (same seeds). 2026-11-10T23:12:10 Astra wsn dev feedback delivered (11/20); phase 2 started. 2026-11-10T23:32:16 Astra cap v2 frozen (policy.py 6764e77a…, base_policy = v1 policy); v1 unchanged; validator PASS re-run; reads only own scene_data.json. 2026-11-10T23:35:20 Astra wsn v2 frozen; validator re-run (see above); v1 unchanged. 2026-11-11T00:14:54 All 8 cells scored (Fable + Astra). RESULTS.md updated. 2026-11-11T16:45:35Z DIRECT-CONTROL (DC) condition pre-registration (before any DC episode runs): question: LLM agents controlling the robot directly, turn by turn, vs the authored code-policy and weight-space cells (same 10 tasks). models: Fable 5.2 (Claude Code CLI 2.1.321 headless, --effort high) and GPT-6 Astra (Codex CLI 0.167.1 in astra-cleanroom container, fresh CODEX_HOME dc/codexhome_dc, model_reasoning_effort high). Same brief, word for word (dc/brief_template.md sha256 af06fe47921d5213). tools: one MCP tool, act (dc/robot_mcp.py sha256 8f0f36811139bf44): 0-20 OSC_POSE actions per call -> 256x256 agentview PNG + proprio text. No other tools: Claude --tools "" + deny list (dc/settings_fable.json sha256 26884302c986f4d2); Codex shell_tool/view_image/etc. off, read-only sandbox, no network for commands (dc/config_astra.toml sha256 05bdfcd822c91e2a). The robot is reachable only through the MCP tool (token-guarded TCP server on the host); scripted control is impossible by construction. Any Codex command execution found in the logs is reported. episode (dc/env_server.py sha256 2efbaac2059feb1b): LIBERO standard init states, fresh env per episode (seed 0, reset, set_init_state), 10 settle steps [0,0,0,0,0,0,-1], observation/action/clipping/step limits/success = eval_policy.py. Extra DC limits: at most 40 act calls with actions; 60 min wall per episode (dc/run_episode.py sha256 9dd253c8157d88d2). Episode = failure if the agent stops, times out or runs out of calls/steps without success. pilot: 2 episodes per task per model on init states 10 and 11 (outside the main set), to check the harness and measure time and usage. Pilot results are reported but not pooled with the main run. main run: init states 0-9 for every task (100 episodes per model), paired by init state with the existing cells (cap/wsn v1 and v2, both authors; the first 10 of 50). analysis: success k/100 per model with Wilson 95% CI; paired exact McNemar DC vs each existing cell on the same 100 (task, init) episodes, and Fable DC vs Astra DC; per-episode wall time, act calls, tokens; cost at API prices. Deviations will be logged here. 2026-11-11T18:12:12Z DC harness checks and deviations (before the pilot): 1. Fable env: run_episode.py starts the Claude CLI with a clean environment (no inherited ANTHROPIC_*/CLAUDE_* variables). Tool probe from the stream-json init: tools = [mcp__robot__act] only. 2. Astra tool surface (CORRECTION to "scripted control is impossible by construction"): in Codex 0.167.1 the gpt-6-astra model catalog fixes the tool surface. Captured response payload: functions{wait, request_user_input, request_user_input_async, exec (JavaScript V8 cell that can call MCP tools)}, collaboration{spawn_agent, send_message, followup_task, wait_agent, list_agents, interrupt_agent}. Turning off features multi_agent/multi_agent_v2/code_mode/code_mode_host does not remove them (code_mode_host=false only breaks exec), so the config is unchanged. Astra could therefore script act calls in JS or spawn sub-agents. Rule: every Astra episode is audited from agent.jsonl. Episodes with any exec cell or collaboration call are flagged and listed. Results are reported with all episodes and also with flagged episodes counted as failures (sensitivity). Fable episodes are audited the same way (any non-act tool use). 3. Codex MCP approval: the first Astra smoke failed before acting ("MCP tool call requires approval, but approval policy is never"). Added -c mcp_servers.robot.default_tools_approval_mode="approve". The failed smoke was archived as episodes/smoke/astra_t01_i12_FAILED_approval. Smoke (task 1, init 12; outside pilot and main; not reported as data): Fable success, 77 steps, 12 calls, 112 s, 13 images received. Astra success, 81 steps, 8 calls, 62 s, all calls direct mcp_tool_call, no exec/collaboration. Hashes now: run_episode.py f7c4d9e3dcf0f808, robot_mcp.py 8f0f36811139bf44, env_server.py 2efbaac2059feb1b, brief af06fe47921d5213, settings_fable 26884302c986f4d2, codex config 05bdfcd822c91e2a. Pilot starting: inits 10,11 x 10 tasks per model, 3 episodes in parallel per model. 2026-11-11T20:26:29Z DC pilot complete (inits 10,11; not pooled with the main run): Fable 14/20 (ends: 14 success, 5 step_limit, 1 call_limit), median 10.6 min/episode, mean $2.89/episode at API prices ($57.8 total), median 37k output tokens. Astra 8/20 (8 success, 12 step_limit), median 9.9 min, mean $0.84 at API prices ($16.8 total), median 4.9k output tokens. Tool audit: 0 of 40 episodes flagged (only act used; every call returned an image). Same 20 scenes, other cells: Fable code v2 20, Fable net v2 17, Astra code v2 9, Astra net v2 13. Main run not started; awaiting review. 2026-11-11T21:36:23Z DC MAIN RUN starting (10 in parallel per model, both models at once). Inits 0-9 x 10 tasks = 100 episodes per model. Episode protocol, brief, tools and limits unchanged from the pilot (step limits, 40 act calls, 60 min wall). Harness-error rule (written before any main-run episode): an episode is DISCARDED and re-run from scratch (max 2 re-runs) only if it ended without the simulator ending it (agent_stopped or timeout) AND there is evidence of an API/quota failure: Fable result is_error, or a Codex turn.failed, or a non-reconnect error / stderr matching 429|529|rate limit|usage limit|quota|overloaded|too many requests|limit reached, or a timeout with >= 3 API retries/reconnects. Episodes ended by success, step_limit or call_limit are never re-run. Discarded attempts are kept in episodes/main_discarded/ and reported. Monitoring: status.py every ~15 min. Hashes: run_episode.py 24b84225940320b3, run_batch.py 510dcea1f6117cf3, status.py e9404aa483d78c96; brief af06fe47921d5213, robot_mcp.py 8f0f36811139bf44, env_server.py 2efbaac2059feb1b, settings_fable 26884302c986f4d2, codex config 05bdfcd822c91e2a. 2026-11-11T21:58:19Z DEVIATION, main run capped at 50 episodes per model (inits 0-4 x 10 tasks), an experimenter decision for usage-budget reasons. The experimenter had seen interim success counts (Fable 10/14, Astra 13/39) when deciding. Inits 5-9 stay unrun; whether to run them is decided after the 50 complete. Astra had launched inits 5 for tasks 0-3 before the cap; those 4 episodes were killed within ~1-2 min and set aside unscored in episodes/main_aborted_cap/ (not data). The original launchers were stopped; their in-flight episodes (all inits <= 4) finished under the unchanged run_episode.py. A new run_batch.py (sha e04ad5494700745c) skips episodes that exist, keeps at most 10 live episodes per model, and applies the same harness-error re-run rule; for episodes started by the old launchers, the rule is applied by hand. 2026-11-11T22:40:58Z DC main run (capped, inits 0-4) complete: Fable 25/50 (Wilson 37-63), Astra 16/50 (21-46); Fable vs Astra paired 18 vs 9, p=0.12. 0 API retries/reconnects, 0 harness re-runs, 0 of 100 tool-audit flags. Full output: dc/results_main.json (analyze.py sha 434fffbf9462ebd9). 2026-11-12T02:25:39Z ROUND 2 (written-policy cells only; live control not rerun) pre-registration, before any r2 authoring: why: an adversarial review found (a) the r1 brief allowed homemade renderers and kinematic mock simulators, and three of four authoring cells ran closed-loop tests with them; the first post's from-scratch rule forbids this; (b) r1 scoring ended each episode at the first step LIBERO's goal predicate held, and some policies deliberately reached it while still holding the object. brief (r2/template/TASK_COMMON.md sha256 4f51c2212afa0459): adopts the first post's rule verbatim ("No simulation, integration or trajectory emulation, including with your own equations") plus an operational definition (never feed a policy's output into any next-state model and run the policy again; no behavioural queries); removes renderer/mock-simulator permissions; static maths, closed-form constants and independent made-up-input checks (incl. a single static image) allowed; fitting rule now also forbids calibrating constants from feedback recordings (averaging, regressing, reading off values, trying candidates); all executed scripts must be saved under work/. Everything else (tasks, materials, formats, interface, phases, 20 dev rollouts with the same seeds 90000+100*task+{0,1}, 500 official episodes on LIBERO's standard init states) is unchanged. scoring (r2/harness/eval_policy_r2.py sha256 a88d9a701e4a8ad3, spec.py sha256 6b3126fbc256dbd8): primary = strict success: goal predicate holds while no gripper0_* collision geom touches any non-robot geom, for 20 consecutive policy steps, within the suite step limit; episodes continue past the first predicate pass. Secondary = LIBERO standard (predicate holds at any step), from the same run. Harness verified on r1 Fable code v1 (tasks 0, 3, 6, eps 0-2): first-predicate steps reproduce r1 exactly. rooms (r2/harness/build_rooms_r2.py sha256 707f591f52ef979a): ~/cap-cleanroom-r2, ~/wsn-cleanroom-r2 (Fable 5.2, Claude Code, effort high, auto mode, sandbox + read-deny of every other top-level home folder, ~/.claude/projects, ~/.codex and ~/Library/Application Support; tool-input hook to audit/tool_log_r2_*.jsonl), ~/cap-cleanroom-astra-r2, ~/wsn-cleanroom-astra-r2 (GPT-6 Astra, Codex CLI in astra-cleanroom container, effort high, only the room mounted). Materials and interpreter copied unchanged from r1 rooms. Codex homes: r1 sessions, thread history, memories and logs moved to astra/archive_r1/ (login and config left in place). prompts (r2/astra/prompt_*.md sha256 cap_p1 0e017ca9c13e6e7a cap_p2 e249ea6cb96ae061 wsn_p1 12870f080a780df8 wsn_p2 693393f1f6b1dcde): identical text for both models. audit before each freeze is accepted: read every script under work/ and the session transcripts for closed-loop emulation, behavioural queries or calibration from data; a violating cell is reported as such and rerun once from a fresh room. analysis: strict and LIBERO-standard success per cell (Wilson 95%), paired exact McNemar on the same 500 episodes, r1 vs r2 descriptive. Also, as a diagnostic only, rescoring the eight r1 artifacts under the strict harness. 2026-11-12T02:54:17Z R2 Astra v1 frozen: cap policy.py ddc5b1173b2395c2 (~16 min), wsn network.npz 72bc35dc55368c08 (~15 min). Audit PASS for both: work/ scripts read; policy/network calls only on independent fabricated observations with fresh instances (check_independent.py, static_checks.py), no state propagation, no behavioural queries, no inline python in the Codex logs, no data available to calibrate from. Starting dev rollouts (strict harness, same seeds). Fable r2 phase 1 started. 2026-11-12T02:57:54Z R2 Astra dev feedback delivered (strict harness): cap 2/20 strict (2/20 LIBERO), wsn 6/20 strict (6/20 LIBERO). Phase 2 started (codex exec resume, r2 p2 prompt). 2026-11-12T03:17:56Z R2 Astra cap v2 frozen (policy.py e5e774c3c9cf8b5d, ~19 min). Audit PASS: feedback read as contact sheets and gripper-switch frames (work/inspect_feedback_v2.py, inspect_transitions_v2.py), no constants set from measured values (v2_overrides.py: thresholds from source geometry), checks on fabricated inputs only, no inline python. Changes: arrival-gated transitions, direct OSC targets, one bounded empty-grasp retry. 2026-11-12T03:31:42Z R2 Astra wsn v2 frozen (network.npz ea35c795b00e73e6, ~25 min). Audit PASS: feedback read as fixed-index contact sheets and printed snapshots for diagnosis (inspect_feedback_v2.py, feedback_facts_v2.py); new constants derived from source XML (derive_geometry_v2.py; thin-grasp threshold from Panda pad separation + bowl wall half-thickness); checks on fabricated inputs only; no inline python. Kept tasks 1, 5, 7 unchanged because they succeeded in feedback (a design decision, not calibration). 2026-11-12T04:24:45Z R2 Fable wsn v1 frozen (network.npz 4034b0b0c6a2f9c7, ~2 h). Audit PASS: work/ scripts read; test_behaviour.py checks single made-up observations (no propagation); vision_geom.py is a static software rasteriser of nominal stills used to derive image windows and as made-up test images (no sequences, no physics, allowed as static geometry); a06/a07 are single-pose IK solves; tool log: inline python only for file edits and one syntax check. Running dev rollouts (strict harness, same seeds). 2026-11-12T04:26:58Z R2 Fable wsn dev feedback delivered: 15/20 strict, 17/20 LIBERO. Phase-2 message (r2 prompt_wsn_p2.md) sent. 2026-11-12T04:39:16Z R2 Astra official (strict / LIBERO, of 500): cap v1 50/50, cap v2 166/166, wsn v1 176/179, wsn v2 190/191. 2026-11-12T04:42:22Z R2 Fable cap v1 frozen (policy.py 59e4a43ea11bf180, ~2.5 h, including an idle wait for a manual permission approval). Audit PASS: work/ scripts read; IK checks are single-pose root-finds; fuzz_all_tasks.py feeds the validator's random observation sequences (independent of policy outputs); step_budget.py is closed-form plan arithmetic (distance/speed along the planned waypoints vs the step limit; no policy outputs propagated) - noted as allowed static planning, not a behavioural query; tool log: no inline computation. 2026-11-12T04:42:22Z R2 Fable wsn v2 frozen (network.npz e32e109903dae786, ~22 min). Audit PASS: feedback read via read-only summaries, contact sheets and joint-angle printouts (f01-f04); task-7 redesign from static elbow-limit geometry and single-pose IK over the redundancy (f05, f06); AUDIT.md states no constant read off or swept against the recordings. Changes: task-7 redesign, new bowl grasp side for tasks 1-2. 2026-11-12T04:45:01Z R2 Fable cap dev feedback delivered: 16/20 strict, 16/20 LIBERO. Phase-2 message sent. 2026-11-12T05:00:54Z R2 Fable official: wsn v1 strict 374/500, LIBERO 414/500 (per-task strict [49,42,41,50,50,50,1,0,49,42]). 2026-11-12T05:20:37Z R2 Fable official: wsn v2 strict 421/500, LIBERO 464/500 (per-task strict [49,44,48,50,50,50,1,35,48,46]). 2026-11-12T05:22:22Z R2 Fable official: cap v1 strict 439/500, LIBERO 444/500 (per-task strict [41, 43, 50, 50, 49, 50, 47, 44, 50, 15]). 2026-11-12T05:34:49Z R2 Fable cap v2 frozen (policy.py d20945e61595317e, ~25 min; changes tasks 0 and 9 only, applied as asserted text replacements in work/make_policy_v2.py). Audit PASS on emulation and calibration: no closed-loop tests, no constant read off the recordings, no inline computation; v2_regression_replay.py runs v2 open-loop on recorded observations to confirm tasks 1-8 reproduce v1's recorded actions exactly (regression check, no state propagation); fb_v2_postfreeze.py ran after the v2 hash was recorded (report only). FLAG (borderline, for the experimenter's judgment): the task-9 soup locator was iterated over four design rounds scored on randomly generated synthetic still images (52/60 -> 54/60 -> 60/60 -> brightness bound). Not trajectory emulation and no state propagation, but it is candidate selection by measured score on a scene model, which the first post's rule ('must not ... be used for behavior search') arguably covers. 2026-11-12T05:52:02Z R2 official complete (strict / LIBERO of 500): fable cap v1 439/444, v2 471/476; fable wsn v1 374/414, v2 421/464; astra cap v1 50/50, v2 166/166; astra wsn v1 176/179, v2 190/191. Full table, Wilson CIs and paired McNemar in r2/results_r2.json (r2/analyze_r2.py). Starting the pre-registered diagnostic: strict rescoring of the eight r1 artifacts (results in r2/results/r1_fable, r1_astra). 2026-11-12T08:39:48Z R1 strict rescoring complete (diagnostic, pre-registered line 54; r2/analyze_r1_strict.py -> r2/results_r1_strict.json sha256 6aa70342361c419f). Harness check: r1 LIBERO-standard under the r2 harness reproduces the original r1 numbers exactly in all 8 cells. r1 artifacts, strict of 500: fable cap v1 406 (81.2%), v2 427 (85.4%); fable wsn v1 259 (51.8%), v2 318 (63.6%); astra cap v1 92 (18.4%), v2 261 (52.2%); astra wsn v1 218 (43.6%), v2 227 (45.4%). Caveat: r1 authors were told the first-pass rule, r2 authors the strict rule, so r1-strict is not a fair score for r1, only a measure of predicate reliance. 2026-11-12T14:55:49Z Live control strict re-check (diagnostic, not pre-registered; live episodes not rerun per experimenter). Replayed all 41 live successes from actions.npy (all reproduce the recorded first-pass step exactly) and continued each three ways, scored with the strict rule: hold still with last gripper command -> fable 8/25, astra 10/16; open and hold still -> fable 18/25, astra 10/16; open 5 steps then rise 1.5 cm/step (open-and-lift) -> fable 23/25, astra 13/16. Gripper still holding at the first-pass step in 18/25 (fable) and 5/16 (astra). Post uses open-and-lift: fable 23/50 (Wilson 33.0-59.6), astra 13/50 (15.9-39.6). Paired vs r2 v2 on scenes 0-4: fable live vs code 1 vs 24 (p 1.5e-6), vs net 3 vs 19 (p 8.6e-4); astra live vs code 8 vs 11 (p 0.65), vs net 6 vs 12 (p 0.24); fable vs astra live 17 vs 7 (p 0.064). Scripts r2/harness/rescore_live_strict{,_open,_retreat}.py, r2/live/analyze_live_r2.py -> r2/live/results_live_r2.json. 2026-11-12T15:37:21Z Exploratory diagnostic (not pre-registered; does not change any score): access audit and Astra failure analysis. Rooms: materials/ and interface/ byte-identical between Fable and Astra r2 rooms (diff -rq), TASK.md identical; phase prompts identical character for character to what Fable received; same Python 3.15.2 / NumPy 2.5.4 / Pillow 13.0.0; Astra Codex: workspace-write sandbox, no network, effort high, 0 blocked commands, viewed 7 (cap) and 6 (wsn) images vs Fable 10 and 13. Effort: Astra 46 (cap) / 53 (wsn) shell calls, 35 / 40 min; Fable 124 / 101 Bash calls, 155 / 100 min. Traces of Astra cap v2 failing episodes (tasks 0,3,5,6, ep 0): object localised within 1-8 mm, arm stops 6-25 mm short of a grasp target it cannot reach, arrival gate (5 mm) never fires, no per-move timeout, waits until the deadline. Counterfactual copy with tolerance 0.005 -> 0.012 (scratch only, 10 eps/task): task 3 0->7/10, task 5 0->10/10, task 6 0->7/10, tasks 0, 4, 9 still 0/10. Fable cap v2 moves have a closed-form time-out ("never blocks forever"). Astra wsn v2 (tasks 0, 2, ep 0): hovers 8 cm above / 8 cm beside the bowl and never closes. 2026-11-12T18:04:19Z REPORTING DEVIATION (experimenter decision, after all results were known): the post's headline metric switches from strict success (pre-registered primary for round 2) to LIBERO's standard rule (pre-registered secondary), so that the post uses the benchmark's usual scoring and can quote pi0.5's published LIBERO score (openpi per-suite 98.8/98.2/98.0/92.4, weighted 3/3/2/2 -> 97.2%). Strict results are reported in full in a dedicated section. Live control is reported as run (first pass): fable 25/50, astra 16/50. All LIBERO-rule statistics in r2/libero_rule/results_libero_rule.json (analyze_libero_rule.py). Pre-registered strict results are unchanged and remain in results_r2.json. 2026-11-12T18:59:12Z REPORTING CHANGES after a second adversarial review: (1) headline restored to the pre-registered primary metric, strict success; LIBERO's standard rule moves to a separate section with pi0.5's published score as context only (no significance or equivalence claims). (2) Rule adjudication: Fable cap v2's task-9 soup locator was selected by score on synthetic scenes (52/60 -> 54/60 -> 60/60), which the brief forbids; the pre-registered remedy (rerun) was not affordable, so the cell is reported as 'v2c' = v2 on tasks 0-8 + v1 on task 9 (each task is a separate code path; v2 reproduced v1's actions on tasks 1-8 and task 9's tomato phase), strict 441/500 (88.2%), LIBERO 446/500 (89.2%); as submitted 471 / 476. Other synthetic-image fixes (Fable cap v1 locator hardening, Astra wsn v2 tilt guard) were single failed-check bug fixes, judged allowed. All post statistics recomputed in r2/post/results_post.json (analyze_post.py). Live control reported as run (LIBERO first pass); the strict replay is labelled an assisted diagnostic with all three continuations. # GPT-6.2 Astra addendum: the m62 entries of the experiment's timestamped pre-registration and deviation log. Entries are unchanged except that one host-memory detail is reworded. # The same pre-registration also covered GPT-6.2 Sol; entries that concern only Sol are withheld until its write-up # (12 of 28 entries). Shared entries are kept even where they mention Sol. 2026-11-17T18:42:14Z GPT-6.2 REPLICATION (m62) pre-registration, before any m62 authoring or live episode. Experimenter request: rerun the final (round-2, as published) setup on GPT-6.2 Astra (gpt-6.2-astra) and GPT-6.2 Sol (gpt-6.2-sol), Astra first. Identical to round 2 for Astra: same brief/interface/materials (rooms ~/{cap,wsn}-cleanroom-{astra62,sol62}-r2 built by m62/build_rooms.py sha256 16fea5d78a2deb75, byte-identical to the r2 Astra rooms), same prompts r2/astra/prompt_*.md, same container image astra-cleanroom (codex-cli 0.167.1), same Codex config (reasoning effort high) except the model, set by -c model=... (m62/run_phase.sh sha256 260ec6998d01b4f4). Codex homes reset to login + config only (r2 state archived to astra/archive_r2, live-home state to dc/archive_pre_m62). Phases: p1 -> audit -> dev feedback (r2/harness/run_dev_r2.sh, same seeds) -> p2 (resume) -> audit -> official (r2/harness/run_official_r2.sh, 500 episodes per cell, strict primary + LIBERO standard). Audit rule unchanged: read work/ scripts and transcripts before accepting a freeze; a violating cell is reported and rerun once from a fresh room. Live control: 50 episodes per model, inits 0-4 x 10 tasks, same brief/robot_mcp/env_server, at most 10 live episodes per model at a time, same harness-error re-run rule; m62/run_episode.py sha256 6fa973ab94e377c1 and m62/run_batch.py sha256 66c1a8eabb53e22f differ from dc/ only in setting the Codex model and per-model room/container names. Scored as published: LIBERO first pass (as run), strict replay diagnostic. Analysis: same as r2 (Wilson 95%, paired exact McNemar on the same episodes), new models compared with GPT-6 Astra and Fable 5.2. Cost from token usage at list prices (Astra 6.2: $8 in / $0.40 cached / $40 out per Mtok; Sol 6.2: $2 / $0.10 / $10). 2026-11-17T18:55:18Z m62 live, operational (no effect on any recorded episode): (1) experimenter asked for fewer parallel live episodes; the Astra 6.2 launcher was stopped (in-flight episodes finished untouched under the unchanged m62/run_episode.py, all ended by success/step/call limit) and restarted with at most 6 live episodes; it skips existing episodes. (2) astra62_t02_i00 ended 14 s after launch with Codex exit 135 (SIGBUS) before any model output (0 calls, empty usage), during the first concurrent launches on a freshly reset Codex home. Not an agent outcome and not an API failure, so the API-error rule did not catch it; treated as a harness error under the same rule's intent: moved to episodes/m62_discarded/ and rerun once from scratch. 2026-11-17T18:57:18Z m62 Astra 6.2 cap v1 frozen (policy.py 6e5f10f79f2e76e3, contours.npz 6b052e687ae35b97, ~14 min). Audit PASS: work/ scripts read (inspect_geometry.py static collision extents from asset XML; build_v1.py projects mesh outlines from asset files; check_static.py single made-up observations on fresh instances, one act() call each, no propagation); all 37 session commands read source under materials/ or ran those scripts and interface/check_code.py; no simulator, demo data, fitting or policy loops. 2026-11-17T19:00:00Z m62 Astra 6.2 cap dev feedback delivered (strict harness, same seeds as r2): 19/20 strict, 20/20 LIBERO. Phase 2 started (codex exec resume, r2 prompt_cap_p2.md). 2026-11-17T19:01:30Z m62 Astra 6.2 wsn v1 frozen (network.npz b2ca06c1473a0bbf, forward.py dd0cf36c76928c49, ~18 min, 26 affine layers). Audit PASS: work/ scripts read (geometry.py static MJCF bounds; network.py affine-expression compiler; vision.py fixed-camera silhouette scores; build.py; check_static.py single invented observations, exact forward checks, no state transitions); Python executions were only geometry.py, build.py, check_static.py and interface/check_network.py; no simulator, demo data, fitting or closed-loop tests. 2026-11-17T19:03:21Z m62 Astra 6.2 wsn dev feedback delivered (strict harness, same seeds as r2): 11/20 strict, 13/20 LIBERO. Phase 2 started (codex exec resume, r2 prompt_wsn_p2.md). 2026-11-17T19:12:35Z m62 Astra 6.2 official scoring of cap v1 and wsn v1 started (r2/harness/run_official_r2.sh, 500 episodes each, results to r2/results/astra62/, outside the rooms; phase 2 still running and cannot see them). 2026-11-17T19:28:07Z m62 Astra 6.2 cap v2 frozen (policy.py 576c36f209cc01c3, contours.npz unchanged, ~13 min; changes the wine-rack task only). Audit PASS: feedback read as contact sheets and printed pose/joint-limit diagnostics (work/v2/inspect_feedback.py, inspect_rack.py, which states 'do not fit a correction'); new constants (shelf slope 2*atan2(.19061,.68093), shelf/lip coordinates, bottle section .03331, site half-length .10087) trace to wine_rack.xml / wine_bottle.xml via work/v2/rack_geometry.py (rerun by the experimenter: outputs match); check_v2.py asserts v1 = v2 actions on the other nine tasks on single fabricated observations; no simulator, fitting or closed-loop tests. 2026-11-17T19:31:32Z m62 operational (no effect on results): machine under heavy load, so scoring processes were reniced to lowest priority (scoring is deterministic from fixed init states; wall time is logged only). Later scoring jobs are queued one cell pair at a time at the same priority instead of overlapping. 2026-11-17T19:32:09Z m62 Astra 6.2 wsn v2 frozen (network.npz 7f8c4942b4ca7817, forward.py unchanged, ~24 min). Audit PASS: feedback read via contact sheets (inspect_feedback.py) and v1 guard classification of final recorded observations (diagnose_gates.py, no policy calls, no constants set); new constants derived from asset boxes and gripper pad/STL geometry (clearances.py, build_v2.py: milk grasp height, basket rim, pad half-extents) or reasoned (can spacing); no simulator, fitting or closed-loop tests. 2026-11-17T19:32:53Z m62 Astra 6.2 live complete: 30/50 (LIBERO first pass), 0 API retries, 0 fatal, 1 discarded startup crash rerun. Codex homes reset to login + config (Astra 6.2 state archived to astra/archive_m62_astra62); Sol 6.2 p1 started for cap and wsn. Strict replay diagnostics (hold/open/retreat) started at low priority. 2026-11-17T19:46:58Z m62 operational (no effect on results): host under memory pressure, so the not-yet-started Astra 6.2 v2 scoring jobs were requeued with 3 workers each instead of 5 (tasks are scored independently from fixed init states). 2026-11-17T19:59:57Z m62 Astra 6.2 official: cap v1 strict 440/500, LIBERO 462/500 (per-task strict [41,49,50,50,50,50,14,49,50,37]). 2026-11-17T20:03:43Z m62 Astra 6.2 live strict replay (diagnostic, as published): all 30 successes reproduce the recorded first-pass step; released at success 9; strict with hold 11, open-and-hold 16, open-and-retreat 24 (of 50). 2026-11-17T20:07:15Z m62 Astra 6.2 official: wsn v1 strict 239/500, LIBERO 285/500 (per-task strict [3,30,49,50,0,50,4,50,3,0]). 2026-11-17T20:36:07Z m62 Astra 6.2 official complete (strict / LIBERO of 500): cap v1 440/462, cap v2 474/484, wsn v1 239/285, wsn v2 390/395. Full table, Wilson CIs and paired McNemar in m62/results_m62.json (m62/analyze.py).