Failure-Selected Counterfactual Restart Curriculum

Results at a glance 1

Overview

Fresh local negative on 2026-07-14; aggregate sealed.

This experiment tests whether selecting the stronger parent's fresh procedural failures and teaching clean verified restarts can beat exactly exposure-matched replay without conditioning on the parent's failed trajectory.

Research Program

• Program: agentic_breadth_installation.
• Program question: can one Qwen3.5-4B checkpoint acquire execution, induction, verification, repair, state, and commitment behavior without trading away any held-out benchmark family?
• Prior anchors: C50's emission-seam installation result, C58's partial success from context-removal recomputation, and the terminal negative qwen35_4b_universal_on_policy_prefix_repair_token_match.

Question

Does selecting tasks the deployed parent fails, then supervising a concise verified solution from the original prompt, install broader competence than either continued replay or conditioning the correction on the parent's long wrong prefix?

Hypothesis

The predecessor solved on-policy data availability but put 47,123 wrong parent tokens in the candidate's conditioning context and gave it 33,421 fewer supervised target tokens than replay. A counterfactual restart moves the intervention before the error: selection remains on-policy at the task level, while every trainable row begins at the original prompt and contains a full truth-audited solution and answer. Exact matching of forward tokens, loss-bearing target tokens, and absolute loss mass will isolate that mechanism from extra compute or extra supervision.

Setup

• Model: only Qwen/Qwen3.5-4B, revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
• Parent: the published explicit replay_after_close composite, weight SHA-256 7ab4c419...36e2e; runtime LoRA is forbidden.
• Collection substrate: 624 fresh procedural tasks, 48 per each of the 13 universal skills, construction seed 77,114. Truth source SHA-256 is 81edc9ea...de304; oracle-free runner input SHA-256 is 25382689...0f5b.
• Collection: one vLLM event, natural thinking, greedy n=1, seed 66,114, 1,024-token cap, 4,096 context, and identical pinned runner geometry.
• Failure rule: cap contact, missing answer, wrong answer, or more than 128 thinking tokens. Hard correctness/cap failures rank before budget-only failures.
• Selection: exactly four failures per skill, 52 total, deterministic seed 55,114. An undersupplied skill ends the experiment before training.
• Frozen control: same-parent replay continuation. Both arms have 320 rows, 200 byte-identical aligned replay rows, 40 effective-batch-eight updates, 297,731 forward tokens, 126,796 loss-bearing targets, and absolute loss mass 27,632.8.
• Training warm start: the published replay adapter, weights/config SHA-256 bb59d3bd...5154d / 0dfd9bda...120f. Each arm continues independently from that same adapter.
• Train seed: 48. Fresh local seed: 88,010. Conditional aggregate seed: 78,140.
• Hidden-label boundary: no benchmark item, transcript, source, or detailed result is read. Aggregate access remains sealed until strict local promotion.

Run

The CPU smoke path reauthenticates task construction, collection, selection, source tokenization, materialized streams, independent token validation, and the second adversarial review:

PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -B \
  experiments/qwen35_4b_universal_failure_selected_restart_target_match/scripts/run.py --smoke

The full staged sequence is complete. The smoke path reauthenticates every tracked receipt/log, external adapter, full composite tree, and frozen local design. Result artifacts are immutable and the harness refuses overwrite; no further model or aggregate event is authorized in this directory.

Results

The one preregistered parent event completed from pushed-green commit 1744e753: 624/624 completions, 304,013 sampled tokens, 879.9 tok/s, and 394.96 seconds of wrapper wall time. Rollout/metadata/log/receipt SHA-256 values are 4bf15134...1099f, b43b3a0...1206d, 668e9b70...369ff, and 1d35c63a...2b381. The receipt records a clean main preflight, the authenticated merged replay parent, no recovery or generation rerun, benchmark_data_read=false, and a sealed aggregate seed.

This is collection evidence, not a capability result. Failure composition and quota availability are now frozen by the separately checkpointed mining stage:

• 602/624 rows were eligible and 228 were hard correctness/cap failures.
• Every skill cleared the four-row quota; availability ranged from 40 to 48.
• The selected 52 rows are exactly four per skill: 40 hard failures and 12 correct but over-budget cases. Hard-failure availability was below four for abstain, count, route, and select, so their remaining slots prospectively used budget-only rows.
• Selected reasons total 29 cap contacts, 26 missing answers, 13 wrong answers, and 51 over-budget flags; reasons overlap by row.
• Inventory/restart/selection/summary hashes are c19d3de7...66240, 022b1ea4...d951f, 567d6b02...b662, and 2e8a2192...e28ddf.
• All 52 rows restart from the original prompt and zero contain a parent prefix. At this selection checkpoint training was unauthorized; benchmark/aggregate gates remain sealed.

This is still construction evidence, not a capability result.

Exact exposure is feasible without modifying any target, duplicating any row, or truncating any sequence. A deterministic integral solver froze disjoint 68-row candidate-filler and 120-row control blocks around the inherited 200-row shared core. Independent encoding of the final files confirmed:

• 320/320 encoded rows and zero skips in each arm;
• exact equality at 297,731 forward tokens, 126,796 nonzero target tokens, and absolute loss mass 27,632.8;
• exactly 200 byte-identical rows at the same stream positions;
• zero parent-prefix tokens and four clean restarts per each of 13 skills;
• source-token/manifest/control/candidate/final-receipt hashes ac9b9c8a...0bd6, 7ba55045...91de1, 7a8d4566...b5078, 28deb20e...3190, and 52a761ef...170.

The candidate has 16,414 more total thinking-span tokens and 16,414 fewer masked context tokens because of differing zero-weight forced-close composition, but equal answer tokens, close tokens, actual loss-bearing tokens, and weighted loss mass. The second review records this residual sequence-composition difference and authorizes only replay-control training after publication. This remains construction evidence, not a capability result.

From pushed-green exact-exposure commit 821d50d4, the replay control independently continued the authenticated parent for exactly 40/40 steps. It encoded 320/320 rows with zero skips, completed in 297.3 trainer seconds (318.70 wrapper seconds), and reported final train loss 0.3873. Receipt/log/adapter-config/adapter-weight hashes are 3a9cc1ea...6d49, 3bedc341...f25, dce1095c...f8f6, and 5840757d...b1c; the adapter is 169,903,320 bytes. The preflight binds clean pushed main at 821d50d4, the frozen stream, and the original parent adapter. This is an authenticated training event, not capability evidence.

From the separately pushed-green control checkpoint 2c78e655, the candidate then continued the same original parent for exactly 40/40 steps. It encoded 320/320 rows with zero skips, completed in 298.5 trainer seconds (315.33 wrapper seconds), and reported final train loss 0.5838. Receipt/log/adapter-config/adapter-weight hashes are 6aa5c3f1...9871, c8572c88...202a, 6915787d...7f50, and 2072c5c8...39bc; the adapter is 169,903,320 bytes. Its receipt embeds and reauthenticates the published control prerequisite while proving the candidate warm start remained the original parent. Training loss is not a capability comparison. From pushed-green local-design commit 3b8b46aa, the replay-control adapter was explicitly merged into the pinned full composite. All 128 expected LoRA modules had nonzero deltas at scale 2; TF32 was disabled, and delta-norm sum/max were 160.3068/2.8268. Run-receipt/log/external-receipt/weight/tree hashes are 751a0152...f72f, 8a438197...281b, bcb0060e...53e2, e48ed4a0...ae17, and d1a8336d...6027. The complete seven-file tree passed exact config/tokenizer and file-set authentication. This is a deployment artifact, not a capability result; no local evaluation has run.

From the separately pushed-green control-merge commit 6c551000, the restart candidate was merged with the identical method. It applied 128/128 nonzero modules at scale 2 with TF32 disabled; delta-norm sum/max were 160.3715/2.8331. Run-receipt/log/external-receipt/weight/tree hashes are 2956fa41...8ea7, e138a06c...b483, 97edeb08...6df6, d704af19...49a9, and 9f64dc55...4a1b. This is also authenticated deployment construction, not capability evidence. No local output has been observed.

The separately reviewed fresh-local design now freezes 26 new procedural items at seed 88,010, exactly two per universal skill, and compares the unchanged replay parent, matched-exposure replay continuation, and counterfactual-restart candidate. Source/input/design-receipt SHA-256 values are 7b69473b...975f, 6efefc92...15e2, and 124bbf99...2db5. Every arm uses the same explicit-composite vLLM runner and geometry. Promotion requires the candidate's absolute 17/26 floor plus strict wins over both controls on total correct and the six execute/induct/probe items. Complete seven-file composite tree manifests, pre/post-arm model and Git authentication, strict 78-row receipt shape, and durable failure receipts close deployment and transaction ambiguity. Aggregate seed 78,140 remains sealed. This is frozen design evidence, not a capability result.

The one authorized local event then completed from pushed-green candidate-merge commit 64e01a6d. All 78 requests ran through the same vLLM backend and frozen geometry. Parent/replay/candidate scored 17/16/15 correct, parsed 21/22/25, contacted the cap 5/4/1 times, and solved 2/2/0 of the six execute+induct+probe items. The candidate passed the parse, cap, and route-abstention checks but missed the 17/26 accuracy floor and was 0/2 separately on execute, induction, and probe. It lost all four strict comparisons with both controls. Local/promotion receipt hashes are 39fe68b9...de9e and 4c381fbd...6759; promotion is empty, benchmark_data_read=false, and aggregate seed 78,140 remains sealed.

This is a terminal local negative. The clean restart package materially improved bounded emission—four fewer caps and four more parses than the parent—but did not install the selected execution/induction policy and erased both probe successes.

Interpretation

Task-level on-policy failure selection plus clean recomputation fixes the predecessor's long-prefix and target-exposure confounds, but it still does not install broad competence at this dose. The intervention changes termination behavior more reliably than semantic execution: candidate parse/cap metrics improved sharply while correctness fell below both controls and target skills collapsed to zero. This rejects the balanced 52-restart package under exact exposure, not all on-policy learning. A successor should distill short verified trajectories that are already in the model's policy support—successful sampled siblings of greedy failures—rather than another hand-authored oracle trace surface.

Knowledgebase Update

• Program backlog/evidence and shared synthesis: record the terminal local negative and the policy-supported successful-sibling successor constraint.
• Claim ledger: unchanged; no held-out benchmark was opened and no broad capability claim is warranted.

Artifacts

• idea_intake.md — closest duplicate and mechanism decision.
• reports/preregistration.md — frozen identities, gates, and checkpoint order.
• reports/design_review.md — adversarial pre-rollout review.
• data/rollout_tasks_seed77114.jsonl — executable truth and oracle restarts.
• data/parent_rollout_input_seed66114.jsonl — oracle-free vLLM input.
• data/rollout_task_manifest.json and data/design_receipt.json — freshness and design receipts.
• data/failure_inventory_seed66114.json — complete frozen failure inventory.
• data/counterfactual_restart_source.jsonl — 52 clean selected restarts.
• data/restart_selection_receipt.json and data/selection_summary.json — quota and composition receipts.
• data/sft_blend.jsonl and data/predecessor_stream_manifest.json — self-contained authenticated replay lineage.
• data/source_token_lengths.json — exact trainer-encoder source measurements.
• data/stream_manifest.json, data/replay_control.jsonl, and data/counterfactual_restart_candidate.jsonl — exact integral partition and frozen streams.
• data/stream_token_receipt.json — independent final-stream exposure validation.
• reports/compute_review.md — second adversarial review and control-only authorization.
• reports/local_design_review.md — explicit-composite and fresh-local adversarial review.
• data/local_tasks_seed88010.jsonl, data/local_input_seed88010.jsonl, and data/local_design_receipt.json — executable truth, hidden-free input, and frozen local protocol.
• runs/local/seed88010.json and seed88010_promotion.json — complete grading and empty-promotion receipts; sibling raw output/metadata/log files preserve all three arms.
• src/vllm_runner.py — pinned same-backend runner with explicit-composite gate.

Report

Rendered from reports/report.md

Current status

The model-free design, one authenticated parent rollout, frozen failure selection, exact-exposure stream freeze, both paired training events, and the separately reviewed fresh-local protocol, and both current-arm explicit merges are complete. No benchmark event ran. The fresh local event is a terminal negative and the aggregate gateway remains sealed.

The active hypothesis is that task-level on-policy failure selection can help when the supervised example restarts cleanly before the error and target exposure is matched exactly to replay. See preregistration.md and design_review.md for the frozen contract.

Evidence

• 624 fresh truth-audited tasks, balanced 48 per universal skill.
• Zero prompt overlap with predecessor collection/local sources and prior reserved local seeds.
• Explicit merged replay-parent authentication.
• Unit-tested removal of the parent's failed prefix from trainable rows.
• Design receipt SHA-256: e861cd647c5a39df893366b948a39fc2bf67ac08e1b1fe704a69032597ffae24.
• Parent event: 624/624 completions, 304,013 sampled tokens, 879.9 tok/s, 394.96 seconds. Rollout receipt SHA-256: 1d35c63a70d53d8803666cb8c30f4d0efffd884c7f6ab04adceaf8b05442b381.
• Selection: 602 eligible, 228 hard failures, 52 selected (four per skill), with 40 hard and 12 budget-only rows. All 52 are full oracle restarts from the original prompt and zero contain a parent prefix.
• Exact final-arm equality: 320 rows, 297,731 forward tokens, 126,796 loss-bearing target tokens, absolute loss mass 27,632.8, zero skips, and 200 byte-identical aligned replay rows per arm.
• Stream hashes: control 7a8d4566...b5078, candidate 28deb20e...3190, manifest 7ba55045...91de1, and independent token receipt 52a761ef...170.
• Candidate minus control is zero on the three preregistered axes, answer targets, close targets, and parent prefixes. Its 16,414-token target-span difference is zero-weight forced-close composition and is disclosed in compute_review.md.
• Replay control: 320/320 rows, zero skips, 40/40 steps, train loss 0.3873, complete 169,903,320-byte adapter. Receipt/log/adapter hashes are 3a9cc1ea...6d49, 3bedc341...f25, and 5840757d...b1c.
• Restart candidate: 320/320 rows, zero skips, 40/40 steps, train loss 0.5838, complete 169,903,320-byte adapter. Receipt/log/adapter hashes are 6aa5c3f1...9871, c8572c88...202a, and 2072c5c8...39bc. Its receipt binds the published control prerequisite and independent original-parent warm start.
• Fresh local protocol: seed 88,010; 26 new tasks, two per all 13 skills; hidden-free runner input; identical explicit-composite vLLM geometry across unchanged parent, replay control, and candidate; complete model-tree authentication; strict absolute and two-control-relative promotion rules. Source/input/design hashes are 7b69473b...975f, 6efefc92...15e2, and 124bbf99...2db5. This is design evidence only.
• Replay-control composite: 128/128 nonzero merged modules at scale 2; exact seven-file tree hash d1a8336d...6027; merged weight e48ed4a0...ae17; run/external receipts 751a0152...f72f / bcb0060e...53e2. This authenticates deployment and is not a capability result.
• Restart-candidate composite: 128/128 nonzero merged modules at scale 2; exact seven-file tree hash 9f64dc55...4a1b; merged weight d704af19...49a9; run/external receipts 2956fa41...8ea7 / 97edeb08...6df6. This likewise authenticates deployment only.
• Fresh local result: parent/replay/candidate scored 17/16/15 correct, parsed 21/22/25, and contacted the cap 5/4/1 times. Execute+induct+probe subtotals were 2/2/0 of six. Candidate passed parse/cap/route-abstention mechanics but missed the accuracy floor, was 0/2 on each target kind, and lost every strict comparison with both controls. Local/promotion hashes are 39fe68b9...de9e / 4c381fbd...6759; promotion is empty and aggregate seed 78,140 remains sealed.

Interpretation

Removing the parent's failed trajectory and exactly matching forward tokens, loss-bearing targets, loss mass, update count, and shared replay rows did not rescue on-policy failure selection. The candidate learned a shorter, more parseable termination policy—mean sampled tokens fell to 414 from 448 for parent and 436 for replay—but correctness fell, and the two probe wins present in both controls were erased. Clean oracle restarts were therefore in-distribution enough to shape emission but not to transfer the required execute/induct decisions.

This rejects the complete balanced 52-restart curriculum at this dose, not every on-policy objective. The next result-separated mechanism should use verified successful sibling trajectories sampled from the same model on greedy-failure tasks, so supervision stays within policy support while retaining fresh procedural truth, exact-exposure replay, and the unchanged local gate.

Terminal disposition

No further event is authorized in this experiment. Preserve the negative and move any successor to a fresh directory with fresh collection, training, local, and conditional aggregate seeds.

Experiment log 10

Reproduce

Smoke test

PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -B experiments/qwen35_4b_universal_failure_selected_restart_target_match/scripts/run.py --smoke

Full run

checkpointed sequence below; never run stages back-to-back without publishing and waiting for both workflows

Run steps are documented inside the experiment folder (README and scripts).

Browse the experiment folder on GitHub ↗

Related

Related experiments

Nearest by shared name, tags, and programs.

• 2026-07-14 Policy-Supported Successful-Sibling Universal Curriculum
• 2026-07-13 Low-Density Token-Matched Universal Curriculum
• 2026-07-14 Natural-Language State-Table Universal Curriculum
• 2026-07-14 Search-Scaffold Universal Curriculum
• 2026-07-14 Residual-Skill Successful-Sibling Universal Curriculum
• 2026-07-14 → 15 Fresh-Surface Budget-Commit Universal Curriculum