Qwen3.5-4B Deep-Advantage MOPD

2026-07-12 — intake and pre-output design

• Created a new result-bearing directory rather than modifying the completed two-teacher experiment.
• Selected the deep-only branch because deep independently passed both same-prefix audit contrasts in both predecessor blocks; MOPD remained untested only because quick was separately mandatory.
• Reused the exact immutable 40/60 soup by SHA-256 instead of constructing a numerically new starting checkpoint. New training artifacts have a separate external root.
• Froze two new route blocks, the unchanged strict deep-over-quick-and-student rule, five-update locality, four 60-deep/20-soup rounds, three primary seeds, and unconditional final comparisons against sources, router, controls, and sample-more.
• Added two direct mechanism controls: deep targets on one-to-one matched non-deep-selected states, and quick targets on the exact selected states. The off-policy continuation control and parameter soups remain.
• Copied the parent harness and procedural gym, then adapted quotas, route gate, target-cache inventory, locality mixture, controls, and confirmation without generating task-model output.
• Passed 50 isolated tests and all 14 family selftests; verified the exact quick, deep, and immutable-soup file hashes. The smoke receipt contains no task-model generation.
• Committed the complete frozen design at 1ef1f5ad, pushed it to shared main, and wrote runs/preregistration_receipt.json with byte hashes for all frozen files before any Qwen load.

2026-07-12 — pinned-model and installation preflight

• Passed all four pinned-runtime semantic probes and a finite Transformers training forward pass; vLLM resolved the registered full/piecewise graph geometry.
• Revalidated exact quick, deep, and soup checkpoint hashes and merge receipts. On the eight fixed canary prompts, every installed checkpoint differed from base; quick/deep differed on 8/8, soup/quick on 8/8, and soup/deep on 7/8.
• The installation gate authorizes the fresh two-block route qualification. No route evidence or training output exists yet.

2026-07-13 — fresh route qualification passes

• Generated two disjoint 192-state soup blocks and 9,216 continuations: three policies × four selection + four audit branches. All state, replay, model, engine, and branch-hash checks passed.
• Deep routed on 28/26 states. Independent audit deep-minus-soup was +0.16499/+0.12205 by block (pooled +0.14209, one-sided lower bound +0.12297); deep-minus-quick was +0.20003/+0.14203 (pooled +0.16910, lower bound +0.15337). The frozen deep gate passed.
• Quick also passed diagnostically on 29/18 routes in this new replication, unlike its predecessor reversal. The locked treatment remains deep-only; the result is evidence for, not permission to improvise, the later cross-fitted two-teacher design.
• Downstream authorization is exactly locality_pilot. No update has run.

2026-07-13 — five-update exact-logit locality passes

• The first two online candidate batches supplied 55/60 required deep routes, so the frozen runner opened a third batch rather than weakening the quota. Three batches yielded 90 deep routes among 576 failed states.
• The final round manifest contains exactly 60 deep capability units, 20 soup anchors, and 60 disjoint non-advantage controls. Control matching used 57 exact cells and three family/kind fallbacks. Its SHA-256 is f4a1eb0848804ddef874ece0afef89a1ea39d84f2717b971575f2ff5f021e0e8.
• Cached all three policies for 140 samples and 35,147 active positions. The cache SHA-256 is 20c21a5bb8e8b8058d0b7377929b37fd5e2eca8e55598d61b6ac827503761e76.
• The pilot completed all five updates with the exact 15-deep/5-soup mixture. Training mean corrected top-50 loss was 0.05242; held-probe loss improved 0.04773→0.02947 and overlap improved 0.84840→0.85163.
• Exact batch-of-one probes measured centered non-target drift 0.02760, relative entropy drop 0.03112, and target loss 0.01293→0.01170. Every frozen check passed. The probe is one midpoint token per consumed unit, so the authorization is literal and deliberately not described as global token-position invariance.
• Downstream authorization is exactly four_round_mopd. No capability result, control comparison, or benchmark event exists yet.

2026-07-13 — seed-42 integration round-1 cache recovery

• Full-dose round 0 passed: 20/20 updates, exact 60-deep/20-soup geometry, mean corrected loss 0.05669, probe loss 0.08318→0.05112, and non-decreasing overlap. Its merged receipt is 8432e6391ce1f4ce328938163984f490bc424325832c891a58a5dbf35cb06920.
• Round 1 used three fresh candidate batches. It found 81 deep routes among 576 failed states and froze the exact 60-deep/20-anchor/60-control quotas (56 controls exact-cell, four family/kind).
• Target-cache construction stopped before loading any policy because one matched route-control episode tokenized to 3,203 > 3,072. This was an implementation omission, not a registered scientific stop: the fixed training length existed, but completion-preserving prompt fitting did not.
• Recovery keeps the frozen 3,072 budget and every completion/target position, deterministically left-truncates only oldest prompt tokens, records the exact cut in cache/training receipts, and still fails if a completion leaves no causal prompt token. Regression coverage reproduces the exact 3,203→3,072 case. Existing states, branch ledgers, routes, quotas, and model outputs are reused unchanged.

2026-07-13 — live integration integrity audit

• Round 1 subsequently passed its frozen gate: 20/20 updates, mean corrected loss 0.04901, probe loss 0.03915→0.02020, and overlap 0.84121→0.84604. Exactly 60 deep and 20 anchors were consumed with zero prompt truncation. The sole 131-token cut remains a cache-only matched route-control sample.
• An independent math audit confirmed the teacher-top-50 corrected reverse-KL value, gradient, causal indexing, masks, quotas, and no-hint construction. It also exposed prospective fail-open edges before they affected a primary update.
• Cache creation now fails before model load if a capability or anchor prefix would be shortened, and every trainer independently rejects any shortened sample selected by its arm. Per-role truncation counts are receipted. The known round-1 route control therefore requires a deterministic full-prefix replacement from the already-frozen candidate pool before that control may run; the primary manifest/cache remain immutable.
• Cache creation, resume, and training now bind stage, frozen config hash, top-k, and exact quick/deep/soup paths plus model-config and merge-receipt hashes. Existing round-0/1 caches pass the strengthened validator.
• Fixed a pre-control implementation error: off-policy pressure probes had taken the first eight lexicographic units (5/3 and 1/7 capability/anchor in actual rounds 0/1). They now use and receipt the registered deterministic 6-capability/2-anchor geometry. The frozen initial-objective-loss scaling definition is unchanged.
• Full-round probe entropy contracted 10.28% in round 0 and 12.33% in round 1. This is not a registered full-round stop—the 10% entropy ceiling is specific to locality—so no post-hoc gate was added. It is preserved as a collapse-risk warning for confirmation and final interpretation.
• All 58 experiment tests pass after the safeguards. Seed 42 remains in progress; no capability or control comparison exists yet.

2026-07-13 — deterministic control-only full-prefix overlay

• Implemented the required route-control recovery without changing the frozen primary manifest or target cache. The control path replays the original matcher from all 15 hashed candidate artifacts, applies only the registered 3,072-token full-prefix eligibility condition, and fails if the filter changes any nonoffending match.
• On the actual round-1 evidence, 490/495 non-deep candidates are eligible and the replay changes exactly one of 60 pairs. The cut episode-256a1dbfee96673bcc5a8066 is replaced at the same cache index by the first legal exact-cell candidate, episode-2e451ff7c44b165288e7c8f4 (2,907 tokens, no cut). The other 59 controls and the 56-exact-cell/4-family-kind geometry are unchanged.
• Only the future non_advantage_route arm can consume the derived cache. Wrong-teacher, primary, and off-policy paths retain their original inputs. The overlay copies 139 samples byte-semantically and scores only the one new sample under the same quick/deep/soup target policies; transitive receipts bind the source cache, manifest, candidate files, tokenizer, and replacement.
• A no-GPU end-to-end dry run against the real round-1 inputs produced the expected sample at cache index 138 and passed both provenance validators. Nine focused rematch tests and all 79 current experiment tests pass. Actual policy scoring and control training have not run, so no control result exists.

2026-07-13 — seed-42 four-round integration passes

• Rounds 2 and 3 completed the frozen seed-42 trajectory. Each of all four rounds used three fresh candidate batches, selected exactly 60 deep units and 20 soup anchors, consumed them once over 20 updates, and passed finite-loss plus non-decreasing-overlap gates.
• Deep-route supply by round was 90/81/78/83. Mean corrected losses were 0.05669/0.04901/0.04855/0.05404; probe losses were 0.08318→0.05112, 0.03915→0.02020, 0.03476→0.01893, and 0.04873→0.02793. The terminal integration receipt records four completed rounds and gate pass; the round-3 merge receipt is 88512a57ebb190f0392118a30258eee5fb3bc58d5d34ae04e384afc8842f9122.
• Probe entropy drops were 10.28%/12.33%/8.90%/11.42%. The locality ceiling is not a registered full-round gate, so no post-hoc stop was added; the contractions remain an explicit collapse-risk warning.
• Hardened the later control/confirmation chain without altering frozen science: one canonical route-control matcher, atomic score-last external raw artifacts, raw-to-score semantic replay, exact full confirmation geometry, and raw hashes retained through benchmark execution. An independent audit found no remaining blocker; 116 experiment tests and repository CI pass.

2026-07-13 — NF4 versus bf16 diagnostic warns against trainer inference

• Ran the preregistered interpretation-only diagnostic after seed 42 and code stabilization. It validated all four fixed 6-deep/2-soup probes, exact LoRA attachment/replay, 7,970 natural target positions, and unchanged artifacts.
• NF4 objective gain averaged +0.02191, but the explicit bf16 merges averaged -0.000224. Only 15/32 unit gains agreed in sign (46.88%), Pearson gain correlation was -0.152, and mean midpoint update cosine was 0.407.
• Endpoint top-1 agreement was 31/32, illustrating the trap: endpoint outputs can look close while the measured update direction is not. The diagnostic has no scientific gate and no downstream authorization. It does not change seed 42's registered pass; it makes sealed vLLM procedural confirmation the only acceptable capability verdict.

2026-07-13 — seed-43 round-1 pass and byte-preserving recovery

• Seed 43 passed round 0 with 20/20 updates, mean corrected loss 0.05638, probe loss 0.08318→0.04706, and non-decreasing overlap. Round 1 reached the exact 60-deep quota after two fresh batches (52 quick diagnostics, 272 abstentions) and froze 60 deep, 20 soup anchors, and 60 matched controls (56 exact-cell, four family/kind).
• After the second batch's student-state receipt was durably complete, the first quick-branch subprocess stopped before model construction because a concurrent confirmation hardening temporarily removed the legacy _engine_protocol symbol imported by branch_states.py. No branch draw began and no state was regenerated. The exact compatibility function was restored, direct acquisition-import regressions were added, and the same integration command reused every authenticated batch-0/batch-1 byte before starting the untouched quick branch.
• Round 1 then passed: 20/20 updates, zero prefix truncation, mean corrected loss 0.05172, probe loss 0.05210→0.02899, and overlap 0.83395→0.83775. Its merge receipt is 2160ecd97c4ee2eda7a29f09048802b953e99d7f026420a653d22c8c009db449. The seed-43 integration receipt remains deliberately in progress at two of four rounds; this infrastructure interruption is not a scientific stop and no capability conclusion exists yet.

2026-07-13 — parameter-control outputs now fail closed

• A pre-control transition audit found that the soup25/50/75 merge receipts bound their source adapters and output safetensors, but resume and independent authorization did not rehash those output bytes. Corruption could therefore weaken a sealed comparator without invalidating the controls receipt.
• Weighted merges now receipt an exhaustive, sorted recursive inventory of every regular inference artifact (weights, model configuration, tokenizer, and nested load assets), rejecting symlinked or non-regular paths. Both the controls runner and the independent authorization path require exact current names and hashes plus the frozen model, revision, adapters, mixture weight, and merge semantics.
• Regression coverage fails on mutated, missing, extra, and symlinked weight, configuration, tokenizer, and nested artifacts, including both resume and independent-audit call paths. All 199 experiment tests pass in the pinned training environment. No parameter-control output existed when this was fixed, so no result or frozen scientific choice changed.

2026-07-13 — seed-43 round-2 pass

• Round 2 required all three frozen candidate batches. The first two were only four examples short of quota; the final supply was 82 deep, 76 quick, and 418 abstentions over 576 failed states. The assembler consumed exactly 60 deep units and 20 soup anchors once each, with 60 matched controls (57 exact-cell, two family/kind, and one kind/level).
• The all-policy cache bound 140 samples and 35,266 active target positions to the exact quick, deep, and soup models. Capability, anchor, and route-control prompt truncation were all zero.
• Training completed 20/20 updates with mean corrected top-k loss 0.05588. The held probe improved from 0.04249→0.02281 loss and 0.83777→0.84112 top-k overlap, so the frozen round gate passed. Probe entropy contracted about 10.66%; as in prior full rounds, this is retained as a collapse-risk warning, not promoted into a post-hoc stop.
• The round-2 bf16 merge receipt is 08136cf7e1a8b4b46b8ac6ffae3422c9f38737263fe1d86ab20adf698b0156b6. The tracked seed-43 receipt now records three completed rounds and remains deliberately in progress while round 3 runs; no capability conclusion exists.
• Several long generation subprocesses span Git commits because unrelated confirmation/control hardening was checkpointed while the GPU remained live. The exact vLLM runner hash, package lock, model configuration, engine geometry, state/branch hashes, and token ledgers were unchanged and all registered engine checks passed. The commit-marker churn is provenance metadata, not a generation-protocol change.

2026-07-13 — seed-43 four-round integration passes

• Round 3 reached the exact 60-deep quota after two candidate batches, with 43 quick diagnostics and 281 abstentions. It froze 60 deep, 20 soup anchors, and 60 controls (55 exact-cell and five family/kind), with zero prompt truncation across every role.
• Training completed 20/20 consume-once updates with mean corrected loss 0.05130. Held-probe loss improved 0.07297→0.04417 and top-50 overlap improved 0.83025→0.83288; the frozen round gate passed. Probe entropy contracted 13.11%, retained as the same non-gating collapse-risk warning.
• Seed 43 is terminal at four completed rounds with gate pass. Deep-route supply was 90/60/82/60 and mean losses were 0.05638/0.05172/0.05588/0.05130. The terminal merge receipt is 4af497550de22d9bbafdd9de97dd95eabeb6b16b6fa9a7516bf78c4c719d6ecf.
• The independent manifest→cache→trainer→adapter→merge audit passes across all four rounds; the tracked integration receipt hash is 885419b1fd8bafb4d16aa56369d91168e0dc0aca92416cc7a4361eeaad2eba11. This supports replication of route supply and optimizer stability only. It is not evidence of deployed capability gain; seed 44 and sealed comparison remain mandatory.

2026-07-14 — seed-44 four-round integration passes

• Seed 44 completed all four frozen integration rounds. Deep-route supply was 90/82/69/65; rounds 0 and 1 required three candidate batches and rounds 2 and 3 required two. Every round froze 60 deep capability units, 20 soup anchors, and 60 disjoint controls. Control matching was 57/3, 56/3/1, 57/3, and 54/6 across exact-cell/family-kind/ kind-level tiers respectively.
• All four rounds completed 20/20 consume-once updates with mean corrected losses 0.05674/0.05561/0.04284/0.05147. Held-probe loss improved in every round: 0.08318→0.04540, 0.05051→0.02258, 0.06825→0.04500, and 0.05601→0.02952. Top-50 overlap also increased in every round.
• Probe entropy contracted 11.05%/9.53%/10.62%/7.55%. The registered full-round gate still passes, but the round-0 and round-2 contractions retain the material non-gating collapse-risk warning; no post-hoc stop is added.
• The terminal merge receipt is 33ae673db2abda3bfee69f311f5d5d5b8e1bda29fb2c1b286b3adbe514d4ba00. The independent manifest→cache→trainer→adapter→merge audit passes across all four rounds; the tracked integration receipt hash is ff329ebccdc888689b6d6c985a558e66a2385aaa701babf8724d61444428bf1f.
• Three optimizer seeds now establish safe completion of the registered update and replicated fresh strict-deep route supply after round 0. They do not establish capability gain, causal routing, superiority to any comparison, retention, transfer, or composability; sealed confirmation remains mandatory.

2026-07-14 — confirmation model-byte boundary hardened before controls

• A fail-closed pre-control review found two related gaps: semantic controls authorization did not reauthenticate the frozen quick/deep/soup source checkpoints, and the prior confirmation path could recompute and adopt model bytes after authorization when it created global ADMISSION. No controls authorization, admission, or confirmation output existed when found.
• model_provenance.py now authenticates the committed source receipt at ancestor commit 37dc74ef, the exact seven-file model root, the exhaustive weight/inference inventory, and only the frozen source or local tokenizer profile. Symlinked leaves/directories/ancestors, nested extras, non-regular files, and mutated/missing/extra config, generation, chat-template, tokenizer, receipt, or weight artifacts fail closed.
• The semantic controls authorizer now seals one canonical 13-arm map: three sources, three primary seeds, three trained controls, three parameter soups, and the exact soup alias under sample-best-of-8 decoding. It recomputes this map immediately before no-clobber publication. Confirmation requires exact arm-map equality, rehashes all arms immediately before and after global ADMISSION, and rehashes each selected arm before STARTED and after generation. Admission is built only from the authorized map.
• The independent integration/control/benchmark audits and benchmark runner now use the same model authenticator. Future LoRA merges record the exhaustive inference inventory; legacy source and completed primary receipts remain admissible only because every byte matches a fixed canonical profile.
• All 209 experiment tests pass, including explicit mutation transitions after authorization and after admission. This is evidence-pipeline hardening only; it changes no treatment, gate, or capability conclusion.

2026-07-14 — non-advantage-route control arm completes (interim)

• The four-round full-prefix non-advantage-route arm completed 20/20 consume-once updates in every round. Mean corrected losses were 0.05393/0.05036/0.04990/0.04619; every frozen full-round gate passed.
• Held-probe loss improved 0.06090→0.03392, 0.04644→0.02432, 0.04979→0.02781, and 0.03674→0.02730. Top-50 overlap increased in all four rounds. Round-0 probe entropy contracted 10.47%, retained as a non-gating collapse-risk warning; later contractions were 7.79%, 16.34%, and 4.33%.
• The full-prefix overlay reproduced the original matched mapping byte-for-byte in rounds 0, 2, and 3. Round 1 deterministically replaced the sole original state that violated zero-truncation eligibility; no training role was truncated. The terminal merge receipt is 99e4d3258f450173204466bd4a2b4f1dfadfc54d706008e6fc3944a5f7bd57f5.
• This is an explicit in-progress checkpoint while wrong-teacher, off-policy SFT, and parameter controls run. The aggregate tracked controls receipt does not yet exist, so this arm is not admitted to confirmation and supports no causal or capability conclusion.

2026-07-14 — control-runtime recovery before off-policy merge

• All four wrong-teacher training rounds and merges completed. Off-policy SFT round 0 then completed 20/20 consume-once updates and passed its frozen round gate, but the parent orchestrator stopped before merge: the documented python3 entrypoint ran under the system interpreter while strengthened canonical off-policy replay lazily required transformers from .venv.
• This was an orchestration-environment failure, not a scientific stop. No off-policy merge or later round had started, and the complete round-0 adapter receipt was preserved for semantic replay rather than retrained.
• Every model stage now re-execs under the pinned .venv before importing dependency-bearing validators. The guard preserves the exact arguments, accepts the normal virtual-environment Python symlink, and fails before stage work if the runtime is absent. The README documents this boundary and two regression tests cover re-exec and no-re-exec paths.
• All 211 experiment tests pass after the fix. Resumption must first replay and validate the existing round-0 receipt under .venv; this recovery changes no treatment, seed, target, update, or gate.

2026-07-14 — all matched controls complete and independently audit

• Resumption replayed the existing off-policy round-0 adapter under .venv, validated it, and proceeded directly to merge without retraining. The other three off-policy rounds then completed normally.
• All 12 trained-control rounds completed 20/20 consume-once updates and passed their frozen gates. Non-advantage MOPD mean corrected losses were 0.05393/0.05036/0.04990/0.04619; wrong-teacher losses were 0.07040/0.06537/0.06047/0.06949. Probe loss improved and top-50 overlap non-decreased in every MOPD control round.
• Off-policy SFT mean CEs were 0.10926/0.11021/0.09851/0.09942; all registered update gates passed and probe loss improved in every round. The frozen off-policy gate contains no top-k-overlap or CE threshold, so no MOPD criterion was applied post hoc.
• The 25%/50%/75% deep parameter soups each applied 128/128 nonzero modules and recorded exhaustive inference-file inventories. Their receipt hashes are aa42d8e67ee87f8cfd937404bf7c43daa5064483185e2fde72c20d3cb2f43d0d, 1599289b9b83932b6d0a5553daa7907ca2b95a22139140f22e5020f0e6b5280a, and 95b607b4e644f2236b1f01498101dac5a274e008ba1c1fae26e0133234feeb1b.
• The tracked controls receipt passes and hashes to 103ef4cc0b24d7c10666b6f0adfcd4dfae4720415c7fbbc76b681ab79162640b. A separate authorization-path audit replayed every canonical ledger and reauthenticated all model bytes. This authorizes requesting sealed comparison; it is not a capability or causal-routing result.

2026-07-14 — semantic controls authorization sealed

• The no-clobber semantic authorizer independently replayed all prerequisite ledgers, reauthenticated every model byte, and sealed the exact 13-arm confirmation map at 709694b7d770b5cbb09afe8b932bba3891ab4fea39c54c625fc84c5da973072d.
• The complete control-code inventory hashed to 0dff95e5f12847ca36435cad6bcbe752e97a67215968cebb68d52c5cd47efe54 both immediately before and after publication. The authorization receipt hashes to f4a5456844adeafd39e2e4f2a8036ed9fff2c78830b2eab9d4a7bfa1300d2278.
• The only licensed downstream action is sealed confirmation evaluation. Global ADMISSION.json and every confirmation score remain absent at this boundary, so this receipt supports no capability, superiority, retention, or causal-routing claim.

2026-07-14 — sealed confirmation admitted; first arm started

• The live confirmation entrypoint reproduced the no-clobber semantic authorization exactly, then independently reauthenticated the full 13-arm model map before and after atomic global admission publication.
• runs/confirmation/ADMISSION.json hashes to 18c019e92fb6b7f7caed0b0f916b958d528b36b9a30607c2890e6b9385d0125d. It binds block seeds 98700/98800, the exact authorized arm map, evaluator inventory, configuration, and authorization receipt.
• Block-0 deep evaluation entered STARTED; its external journal hashes to c1eb2878bf936d95b62885309257aa1c98e39bbfa9096f3e6a6cf01d3a2a32f1. No generated bundle or score exists at this in-progress checkpoint, so it changes no capability, comparison, retention, or routing conclusion.

2026-07-14 — attempt-1 deep transaction quarantined before score

• Block-0 deep completed 4,032 atom rows, 384 episode rows through 18 turns, and 19 contiguous inference-call bundles. Semantic finalization then failed closed before scores.json with confirmation call journal output lacks a scored runner field; the terminal QUARANTINED marker hashes to 7c93540429dc340c1b62d4cce9e5d2aa82fa3439e615d5ffb3ae4d5300e33580.
• A key-presence-only audit found 6,879 journal outputs. The 1,392 outputs that carried retained_thinking_token_ids satisfied the strict schema; the other 5,487 lacked only that field. VLLMRunner._ordinary_output omitted it, while harness._slim intentionally defaulted the absent optional evidence to []. Thus the generated/scored projection and strict journal schema had diverged; model outputs were not malformed.
• No prompt, gold, generated text, item score, aggregate score, or performance comparison was inspected. The entire transaction, original authorization, and admission are archived as attempt 1 and will not be reused. The failure receipt hashes to 2e645322ead3fbbdf58760849fe17def81fd12b62cdfa4b6c58808e24612ed41.
• Recovery is limited to making the ordinary runner output satisfy the already registered strict schema without changing text or scoring, adding a direct regression, issuing fresh no-clobber authorization/admission, and rerunning both sealed blocks from an empty current tree. This is evidence-pipeline failure, not capability evidence or a scientific stop.

2026-07-14 — ordinary-output journal schema repaired

• The strict call-journal validator remains unchanged. The ordinary runner path now emits retained_thinking_token_ids: [], exactly matching the scoring harness's pre-existing projection for a path that never constructs a forced-close continuation. The corrected runner hashes to 1e065b9c3718e4d2353dc3215928b936ed3fdaeb0d1e04ce96f215d4c9331054.
• The change does not alter model calls, sampled IDs, decoded text, prompt bytes, task seeds, token budgets, answer/action extraction, scores, engine geometry, or model artifacts. A direct regression invokes the naturally closed budget path and requires both retained and injected evidence fields.
• All 212 experiment tests pass, including the full call-journal semantic replay suite. Attempt 1 remains terminal and no archived byte will be reused. A fresh no-clobber authorization/admission, bound to this corrected source inventory, is required before any GPU rerun.

2026-07-14 — frozen recovery-review placement corrected

• The first fresh authorization request stopped immediately, before model hashing or output publication, because the post-quarantine review had been appended to preregistered reports/design_review.md. The frozen-design audit correctly rejected its changed hash; no authorization file was created.
• The original design review is restored byte-for-byte at its registered hash cdeaad5b7bb363d0d650af6080d2d0be6c41d0873328e810965d017da271dd5b. The operational analysis now lives in the additive, non-frozen reports/confirmation_recovery_review.md.
• Repository checks pass after the correction. This was a pre-authorization evidence-pipeline stop and changes neither design nor result.

2026-07-14 — corrected runtime receives fresh authorization

• Fresh no-clobber semantic authorization passed after independently replaying every integration/control ledger and reauthenticating all 13 arm models. The arm-map hash remains 709694b7d770b5cbb09afe8b932bba3891ab4fea39c54c625fc84c5da973072d; model selection and decoding are unchanged from attempt 1.
• The corrected 56-file control inventory hashes to 690a5b5e345f7f4070f731eb69f7a6d9adf7f8148d4b4c07e8234cb8916250ef both before and after authorization and explicitly binds runner hash 1e065b9c3718e4d2353dc3215928b936ed3fdaeb0d1e04ce96f215d4c9331054. The new authorization receipt hashes to 2b9b86aa76bfb87169a2c70313f967f20c13a09e62fbab25069120e29f0ef9f1.
• Fresh admission, raw output, and scores remain absent. This authorizes only a new empty-tree sealed campaign and contributes no performance evidence.

2026-07-14 — attempt-2 confirmation admitted from an empty live tree

• The live confirmation entrypoint reproduced the corrected no-clobber authorization exactly, independently reauthenticated all 13 arm models, and atomically published fresh runs/confirmation/ADMISSION.json at hash 6424b68d01420154d10b7a999332eb4b9d44fca3cad63cf7266f098b9d9c990e.
• At this checkpoint, no external confirmation file, evaluator process, GPU allocation, generated row, or score existed. The parent was still performing its post-admission arm-byte recheck. This is in-progress provenance evidence only and changes no scientific conclusion.
• The post-admission check then passed and block-0 deep entered STARTED; the external journal hashes to ca44441a2784a79d314b7c359dde6f136617d6bf68301b9dd0741fe3325b8d38. No generated bundle or score existed when this start marker was recorded.

2026-07-14 — attempt-2 block-0 deep completes and validates

• Block-0 deep generated 4,032 atom rows and 384 episode rows through all 18 turns. The generated journal hashes to adc322ff06f1c585653467c7f8b745e6bc89aad5103dbd31e29312cbad8ac09c; atom and episode raw files hash to 80fe4abcb8f09fff268f514e4b9fc2cfb3718d566b537988a2fab50d46339754 and 2e5b0b467e9a543d468fc7759675c022b6c4b29cbd5cdf6e8109335003a8c150.
• The unchanged strict call-journal validator passed, directly resolving the attempt-1 evidence-pipeline question. COMPLETE.json hashes to d029202d1f11626bdf76ef6f6607b5e7033fc58584d852a83cc30da48dc6663a and binds score hash f6d7c0271aa2d7acdc04e1e62e54dfb79d7a49f244fe332e898b4b813937d1e6.
• The parent accepted registered aggregates 0.580068 (n=3,072) on the deep stratum and 0.810029 (n=1,344) on the quick stratum, with 4,261,790 sampled tokens. This is a single source-arm anchor, not treatment evidence; block-0 non-advantage evaluation started next with external journal hash b5b1a6c6b19d95630c066ba0767fbaa62f8627836a03deb8384ab71003c7286a.

2026-07-14 — block-0 non-advantage control completes

• The non-advantage control completed the same 4,032-atom/384-episode sealed geometry and passed strict journal validation. GENERATED.json and COMPLETE.json hash to 42b5c16414a6d3eb1c013990ea47cae02c132c3b983908023bccf107298eb947 and c614b53218d0c46e89e9ebc185441b7bea6af5c6b364d8a8eda55acef79e0f3d; the score hashes to d3e7b57fb437c3ca4194b74f8e7f4fa63ab067d50e62188de8e07d6f6d7ba7d8.
• Registered means are 0.565173 deep and 0.790951 quick, respectively −0.014895 and −0.019078 versus the block-0 deep source anchor, with 4,468,793 sampled tokens. This is a one-block control/source comparison, not treatment evidence or a causal advantage-routing result. Off-policy SFT evaluation started next with external journal hash fbd64d2cd877e2e84a811337541e1d60de63a3d680ad8fad67fcf785e9db4e2e.

2026-07-14 — block-0 off-policy SFT control completes

• Off-policy SFT completed the sealed geometry and strict journal validation. GENERATED.json and COMPLETE.json hash to f865801c755fb2fc1ba2ff81d1dd4f780d432b99116b69986ecf06d535a8f8ae and b23d72356ab3cedf7794462d6bc6d403705eaa1b964eda4893ec263e5b5195de; the score hashes to 44e8e830e0a05c5d903b0638fcc756ea19203b80044b26ce70948b5d6c5ae002.
• Registered means are 0.577559 deep and 0.797585 quick, −0.002509 and −0.012444 versus the deep source anchor, using 4,342,736 sampled tokens. It is +0.012386 deep and +0.006634 quick versus non-advantage, but remains one control on one block; treatment and replication conclusions are sealed. Primary seed 42 started next with external journal hash fc7fd5c948893412f06d181a647a45ff9b183b5f833d291089b046096acc9ee4.

2026-07-14 — block-0 primary seed 42 completes below source

• Primary seed 42 completed the sealed geometry and strict validation. GENERATED.json and COMPLETE.json hash to a3ecc145f38c4413e27bd1f3b56bc868836129069dd647b30b6bdafd96159caf and 104601605d2ad0b5c12f41947242989c2cfdd11d24dedd954abc0134cf955274; the score hashes to 585f99d71141a70de1db609b827b767218068645d31db0f7396324152c232841.
• Registered means are 0.577005 deep and 0.795225 quick, −0.003063 and −0.014804 versus the deep source anchor, using 4,422,572 sampled tokens. It is also −0.000554 deep and −0.002360 quick versus off-policy. This is an unfavorable first treatment block, not a terminal conclusion: the frozen decision requires all treatment seeds, comparators, and both blocks. Primary seed 43 started next with external journal hash 81047c1f3329eb71fd76dfa47a4a2c6dd87627c2bcc227761567b155d400d3ef.

2026-07-14 — block-0 primary seed 43 repeats negative deep sign

• Primary seed 43 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 32c006cfa1e970cf845b2b1d8af23d606a314f1ee09aeba49a1466a0721a2d95 and 44c7f6c239d00fb512f3c474ca20f711a7d85f39b3fd7f1607729a97caaa36bb; the score hashes to 9ebf64444ce068a14de943d2c3c28c34e2dc567a23197e7307c93a029d42111e.
• Registered means are 0.576097 deep and 0.804167 quick, −0.003971 and −0.005862 versus the source, using 4,408,848 sampled tokens. The deep sign repeats seed 42's unfavorable result; seed 44 and block 1 remain required. Primary seed 44 started next with external journal hash a607b9b93f7c629aff92083f308c53cb0058e353f00a4ee7f2603bcd4e4082e2.

2026-07-14 — all three block-0 treatment seeds are below deep

• Primary seed 44 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 49ca7c8cfb9597b2b8ec8b047a82ce9de321f11cf323d695ff7d068d44b0c1af and 2cb75d45ba30974e3457212001377804a951cb422ad2f62951326e752b8b9dc9; the score hashes to cf3021207260a30357df7dc52d74c1b75fb81a315a04d22fb8424f03dfff4a45.
• Registered means are 0.572010 deep and 0.805422 quick, −0.008058 and −0.004607 versus the source, using 4,393,570 sampled tokens. Seeds 42, 43, and 44 all have negative block-0 deep signs, so the frozen success condition is already unreachable. The preregistered campaign nevertheless requires all arms and both blocks for the terminal control ranking and receipt.
• The block-0 quick source evaluation started next with external journal hash 0ab077980fc97ee8cc5481575a0059896f686735539903ea6634cb15f59b154d.

2026-07-14 — block-0 deep source also dominates quick source

• The quick source completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 8b826d1202add12808efd9951f964598b2f01bfe2775ae23b2abb6fe64621185 and a51a444f0d467dfa6005fda65ddb80c62acfdb7e75acdd3c64cde3c488a2fb80; the score hashes to 089bbdb35d9749cece3fe23bead1f3da25149c5409994d2ca1c64de4364ab08b.
• Registered means are 0.529284 deep and 0.775684 quick, −0.050784 and −0.034344 versus the deep source, using 4,773,735 sampled tokens. The deep source is therefore the better source on both block-0 strata. All three MOPD seeds lie between the two sources but fail to improve the better source.
• The frozen soup comparator started next with external journal hash b2ca6c3aeb2348510d78c99ba7efefe417684a14e5ae946e5a11316fbedb699d.

2026-07-14 — block-0 MOPD does not dominate its soup initialization

• The immutable 40/60 soup completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 243d7fb8d00b0ccb430811cdb9017650395139dcdf0504eb630716adc1728da8 and 742f388e187847f1e74654a855cd3529cc6145cf392f8aa676dc63b8942251ca; the score hashes to 57dce804f32f7e0c2c0c18b23cb170075698eb9216587a34ce8818cbc7eb090c.
• Registered means are 0.571744 deep and 0.811128 quick, −0.008324 and +0.001099 versus the deep source, using 4,392,180 sampled tokens. Primary seeds 42/43/44 are only +0.005261/+0.004353/+0.000266 above soup on deep and are −0.015903/−0.006961/−0.005706 below it on quick. The treatment does not dominate its own initialization on this block.
• Soup25 started next with external journal hash 4bd39562939e2d3b726ce5221e22370fed9f2accdb6ddee8b702a9c558a8c2b6.

2026-07-14 — block-0 soup25 is dominated by the initialization

• Soup25 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 777084b5a15287b83302d20a6b40ba828c4a72f9276056e3eebf6d1ca04b065a and f98a8d5e6581253c68b2e900f23cb53d1e740aff27ef9181ef6729fe842e0a28; the score hashes to 574617cd1646ec8d00f7cb3d6861d3c89a6a892ecb454d2c79d08d2069bf3487.
• Registered means are 0.543311 deep and 0.801212 quick, −0.036757 and −0.008817 versus the deep source, using 4,693,690 sampled tokens. Soup25 is also −0.028434 deep and −0.009916 quick versus the immutable 40/60 initialization, so it is dominated on both strata.
• Soup50 started next with external journal hash b55a6675fea796006dc99e44bf7b24f42b7344f6ceae11d163c568d260fd544d.

2026-07-14 — block-0 soup50 traces the same Pareto trade

• Soup50 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to da8d590fadceff3aa4403da29d1b60aca3b99427133826e937cafe072ac0230d and 8c2e437777096f6e024a6a65ac08f6a0e661e3abc853803fba9fa91412fb4e02; the score hashes to 9226553f6189d59fbac461c8eb4faf8e6299bc22c941210ef2b9195fb2effafa.
• Registered means are 0.562536 deep and 0.812542 quick, −0.017532 and +0.002514 versus the deep source, using 4,494,796 sampled tokens. Relative to the immutable 40/60 initialization, soup50 is −0.009208 deep and +0.001415 quick. Every MOPD seed is higher on deep and lower on quick, so neither checkpoint class dominates the other.
• Soup75 started next with external journal hash 4e2c0077e11ab05405721899382e64609824af62146372d229ce4c1ede874263.

2026-07-14 — block-0 soup75 matches or dominates MOPD

• Soup75 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to f7257cc5f275aed3aaabfcd1058578471a1c9778d1b5e5b02913a10bd8467d8a and e9b38a1aa7194347df551857c49ee3499bceb1c850db6bdcecf166d54f61fd93; the score hashes to 90ec218905010e0469cf306e2b5bb901d06d5853334959515c98745bde74adad.
• Registered means are 0.576786 deep and 0.807131 quick, −0.003282 and −0.002897 versus the deep source, using 4,332,931 sampled tokens. Soup75 dominates primary seeds 43 and 44 on both strata. Seed 42 is only +0.000219 deep and −0.011906 quick versus soup75.
• The 40/60 soup best-of-eight sample-more comparator started next with external journal hash f085310f00e629c80f4fd387aa44a077b9e6ae34dac133136029ed866d6d1745.

2026-07-15 — block-0 sample-more decisively beats MOPD

• Soup best-of-eight completed its sealed 4,032-atom/384-episode transaction. GENERATED.json and COMPLETE.json hash to 64f2c7b8d15bce5911308f8662d2d884252fa4351fe9b46b6b285ffcdcc74740 and 1c6c4656d1ed7f85c9551e3c534bfd82f2cedaee1ab762ad40ab9eb17822f4af; the score hashes to d2270cde58d691ba19ec80f7c0796438805b117589755cc8a796afb9abc34a80.
• Direct item-weighted means are 0.770908 deep and 0.946205 quick, using 35,600,602 sampled tokens. The frozen equal-cell projection is 0.770805 deep, 0.946205 quick, and 0.858505 joint. Seed 42 trails by −0.180246, −0.150980, and −0.165613, respectively.
• The interim values recorded for earlier arms are direct score-file summaries; the terminal analyzer uses equal cell macros. Replaying that frozen projection confirms seeds 42/43/44 trail deep on the block-0 deep macro by −0.000481/−0.005426/−0.004869, so the already-unreachable success rule is not caused by row weighting.
• Wrong-teacher started as the last block-0 arm with external journal hash 13913bd80710ef29a2b141718bf1fa71c4395bd54c92981fea6a6151c0d5125b.

2026-07-15 — block 0 completes with a small teacher-identity signal

• Wrong-teacher completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 5cef623a5b72921d272251a41ea153b60d0f69fb0ecc365324eb6be8465d7c0e and 5a1469de05795e5aa592a0c8ea55ed63da2d594b66aea6a3022602d4782d6a41; the score hashes to 2c41e25dac43862cd7216cb6a92e66b67c0b3940656c954b05790e4e04a744e7.
• Direct item-weighted means are 0.571148 deep and 0.788833 quick, using 4,446,949 sampled tokens. Equal-cell macros are 0.587505 deep, 0.788833 quick, and 0.688169 joint. Seed 42 exceeds them by only +0.003054, +0.006392, and +0.004723, respectively.
• The correct teacher therefore matters directionally on block 0, but the effect is much smaller than the treatment's deficits to the deep source, immutable soup, soup75, and sample-more. This cannot rescue the frozen gate.
• Block 1 began with the deep source; its external STARTED journal hashes to 20a2516780884ee672aa21f1e684ebe000770942e18d34f8954ac38ebcef07c8.

2026-07-15 — block-1 deep source anchor completes

• Deep completed the second sealed block. GENERATED.json and COMPLETE.json hash to 565e023ab8854e70ea0de6093a80c13b0cf60688fe17bab1c121d2213238711e and 9db347c86f5de9a782a5f41425b11d061a7d8f88f423cbf0a3ede36c59553510; the score hashes to 0a6c94de3f1a04126b4fa6c5d2d1e21ffa0e09636320093c7b71d36dc36a1678.
• Direct item-weighted means are 0.584024 deep and 0.811023 quick, using 4,228,153 sampled tokens. Equal-cell macros are 0.600187 deep, 0.811023 quick, and 0.705605 joint. This is a source anchor only.
• Block-1 non-advantage started next with external journal hash e3951bead0ccf17014cf1479f258dfd0b7e68f055eab5eb5419d737c7724d77f.

2026-07-15 — block-1 non-advantage remains below deep

• Non-advantage completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 9d6460a6ae69b1fb8333cac05992b8e5a0c6b8617ecab7cb94200e7cd3841863 and 6c85dac0a1464cb729be4f551de2964267aac313652b0cf6be6a7d26a8833bda; the score hashes to a596c3541ab9816c15cd7dd8e3fc1f58c7c0845ccb2c62aad7c6218a79930e17.
• Direct item-weighted means are 0.578032 deep and 0.795058 quick, using 4,369,088 sampled tokens. Equal-cell macros are 0.596955 deep, 0.795058 quick, and 0.696006 joint, respectively −0.003232, −0.015965, and −0.009599 versus deep.
• The non-advantage control is below deep on both blocks. Off-policy SFT started next with external journal hash 4a2626981d694d6f2740a138b3d4e4a62a69ab6feb835b4b0e4d680de20c29bd.

2026-07-15 — block-1 off-policy trades deep for quick

• Off-policy SFT completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 388288924eedd4b824354351d9427988ee91d675261f31b762e91024f582f5b8 and 7a5dfe15cd57303a96514a9a912e9f9ff072f31c2109fdbfb7e52bfc9dd9400c; the score hashes to b3a3cacd73fd8e42cfee98d2674dd7c413fbca479c88aed41f854cd9718d1edc.
• Direct item-weighted means are 0.583994 deep and 0.800776 quick, using 4,282,251 sampled tokens. Equal-cell macros are 0.603801 deep, 0.800776 quick, and 0.702289 joint: +0.003614, −0.010247, and −0.003316 versus deep. It is a small Pareto trade rather than a source improvement.
• Primary seed 42 started next with external journal hash be153537b495033edafb9c84d808bb67378bb1d3fca97f3e73aac268d3069bdf.

2026-07-15 — block-1 seed 42 independently trails deep

• Primary seed 42 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 96475fa27b93709f939ac37571de8759495ff7f2cabf9b1b73f0b55303d5db9c and e2be8c88c46aff569d48fbfca32fba7f4dc1332771ff4dce81579c91eef136a5; the score hashes to 240cab29fe579e99e40b6a0f1b72b8fce22574ddf2780167a81416c01496892b.
• Direct item-weighted means are 0.575202 deep and 0.806729 quick, using 4,404,223 sampled tokens. Equal-cell macros are 0.592386 deep, 0.806729 quick, and 0.699558 joint, respectively −0.007800, −0.004294, and −0.006047 versus deep.
• Seed 42 therefore has a negative deep-cell sign versus deep on both sealed blocks. This independently reinforces the frozen stop while the remaining arms continue for the terminal receipt and control ranking.
• Primary seed 43 started next with external journal hash 62cca0f1c558cccff5f7a490aee4472188f167bbfd55f170b0aa989f638046a9.

2026-07-15 — block-1 seed 43 makes a tiny non-dominating trade

• Primary seed 43 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 22a67d4b264ac64a20a5ba9a38080f43c23ba29ea1402ad8798f939917c33b70 and 58916aab2b52bfeee0ce44043a49b6a2f613f57406f115e485aa74a8086ce72a; the score hashes to af371468fef92b97cc931408722e8eac68152b052e5d1f6d875847633124d674.
• Direct item-weighted means are 0.582282 deep and 0.806832 quick, using 4,367,189 sampled tokens. Equal-cell macros are 0.601865 deep, 0.806832 quick, and 0.704349 joint, respectively +0.001678, −0.004191, and −0.001256 versus deep.
• The tiny deep-cell gain is paid for by quick loss and does not dominate the source. It also misses the registered positive block-1 joint direction versus deep required of the replication seed.
• Primary seed 44 started next with external journal hash 394c423432ad25f86e7270de2f896ce60b87bf8b6b2b2565eb72f477e832381f.

2026-07-15 — all block-1 treatment seeds fail source domination

• Primary seed 44 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 34ac9f0c0f804fc094f4aad024f6b239bceb5aca04fa09a28ad4c8342bd9daf2 and f71411e8742c022c9b000c7c7cbaba08e592770599fe1cb4eb8dc76fffc44b1f; the score hashes to b6e6e887e2cf3ae230a0ee6fbb4ddb8286f68a1fcb31681306af593a76260485.
• Direct item-weighted means are 0.571220 deep and 0.807583 quick, using 4,359,813 sampled tokens. Equal-cell macros are 0.590461 deep, 0.807583 quick, and 0.699022 joint, respectively −0.009726, −0.003440, and −0.006583 versus deep.
• All three treatments fail to dominate deep on block 1: seeds 42/44 trail on both strata and seed 43 buys only a tiny deep-cell gain with quick and joint losses. Remaining arms rank controls; they cannot reopen the frozen gate.
• The quick source started next with external journal hash 8f441c4df5e781a713415b6190ee181c865276a3b0e3e6ffe906c62ae4853ab6.

2026-07-15 — deep is the better source on both sealed blocks

• Quick completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 2d4d6539f4ce0636ef2eade37b64b8d9a778c2525acb2cbe594b5d138deb9a65 and 40d777ee9d8054c66a4805d5c0b3ef6fb00cb88f1f712b6f932f614836c36786; the score hashes to e820adcc5728e05994c0a2190e267701833f98263745bb4b3a5c23deb892ad48.
• Direct item-weighted means are 0.529900 deep and 0.777722 quick, using 4,736,614 sampled tokens. Equal-cell macros are 0.551467 deep, 0.777722 quick, and 0.664595 joint, respectively −0.048720, −0.033301, and −0.041010 versus deep.
• Deep is the better source on both strata in both blocks. The treatments beat quick on block 1 but do not beat deep, locating them between or near their sources rather than beyond the source frontier.
• The immutable 40/60 soup started next with external journal hash 8b6f61cdf306d8d74ff73bf7960fe7f043a1fa5016c625543f91ea976c379b8f.

2026-07-15 — treatment beats initialization locally but not across blocks

• The immutable soup completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to a4d7ecd6f0fee02e30e6e844f223cf8f1b3b6ed6d63ecd1cf79d1bac2923fae3 and 46768d2d95411f2831623803eaf84dd2ab8783b1432bab36fde3dae4a3995097; the score hashes to cefed360bea80ec79731ca3b891d49ea391bfa2a9a3130a020e95185dcaa6536.
• Direct item-weighted means are 0.569971 deep and 0.800865 quick, using 4,363,638 sampled tokens. Equal-cell macros are 0.589461 deep, 0.800865 quick, and 0.695163 joint, respectively −0.010726, −0.010158, and −0.010442 versus deep.
• Seeds 42/43/44 beat soup on both block-1 strata, with joint gains +0.004394/+0.009185/+0.003859. This block-local update signal does not replicate block 0, where no treatment dominated soup, and it remains below the better-source frontier.
• Soup25 started next with external journal hash 8dfbec59f535168b34097f2f56ea25237dc5f19f9a1bdbfddb6aea8c13f6d698.

2026-07-15 — soup25 is dominated on both sealed blocks

• Soup25 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 1fdfef844ceedcf74bbd48667835b805e2c8c24e32848959951ab2cf59c29a07 and 36957d441d8df25fd03e68dd8eaa2de5b1bdb8e885dac7d223f76f10eb2f9d22; the score hashes to 6c0d19733ad9a3dbba9c7b0364276625c6b74b7e5fda181d28e3ccc740a76d00.
• Direct item-weighted means are 0.543777 deep and 0.792800 quick, using 4,688,835 sampled tokens. Equal-cell macros are 0.562769 deep, 0.792800 quick, and 0.677784 joint, respectively −0.026693, −0.008066, and −0.017379 versus the immutable soup.
• Soup25 is dominated by the initialization on both sealed blocks and cannot explain the treatment checkpoints; every block-1 treatment dominates it.
• Soup50 started next with external journal hash fd036a60b739b9b3bf484956de8f0ab01d657bbe563824412ea8989cc3203afe.

2026-07-15 — soup50 is dominated on block 1

• Soup50 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 4e57dcf508c7216435eecc074aee79a8657b7f48d987a5e65a7756b13fa0a47e and 2158d5451c2df2412c9b2431ad76b97b3e6a19633644c6162774fef309e5803d; the score hashes to 58674e67b7bfa0ddd9d7fb65668bb9ae0e8565c3573832311a2b2673b016eded.
• Direct item-weighted means are 0.565327 deep and 0.795188 quick, using 4,451,128 sampled tokens. Equal-cell macros are 0.586316 deep, 0.795188 quick, and 0.690752 joint, respectively −0.003145, −0.005677, and −0.004411 versus the immutable soup.
• Soup50 is dominated by the initialization and by all three treatments on both block-1 strata. Unlike block 0's soup50 Pareto trade, this point adds no frontier explanation for the treatment checkpoints.
• Soup75 started next with external journal hash c9ff00e7325bff53b964cfd2ce556336a187e2e3036fc80aa75e5cb5a0395460.

2026-07-15 — soup75 exceeds the primary on block 1

• Soup75 completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to 55d92933db0da2e9de7e16bc5474f21d88c4f33cb3ce38e773706cc30ffc0b07 and 8876a7bde424166d46a6e3e0f37d11ee1333f7c25fa5ec5bb6a38d3ec77af8d2; the score hashes to 7367f35384ec5f50e7d02ba1d7a5f30ce2fde6390f8e6ca8bc0ad86369dc0682.
• Direct item-weighted means are 0.576310 deep and 0.807503 quick, using 4,289,527 sampled tokens. Equal-cell macros are 0.595053 deep, 0.807503 quick, and 0.701278 joint, respectively +0.005592, +0.006638, and +0.006115 versus the immutable soup.
• Soup75 beats primary seed 42 by +0.002667 deep, +0.000775 quick, and +0.001721 joint on block 1. Together with its block-0 joint advantage, ordinary interpolation exceeds the registered primary on both blocks. Seed 43 trades higher deep for lower quick; soup75 is also +0.002256 joint versus seed 44.
• Soup best-of-eight started next with external journal hash 092c63c98849093b209da1a9089aee7a0f805d978286f2c39bef30af730553fe.

2026-07-15 — block-1 sample-more decisively wins again

• Soup best-of-eight completed its sealed transaction. GENERATED.json and COMPLETE.json hash to a7b79d4bb34a98788c8d19272e9421ba74cbc12985de95d0a1f5f3f43dd89371 and 386c0587218badb42833b3ede837a9795874985c1dda0fea2a1830c76e1e1e07; the score hashes to c74e6f2f4225b220f99567aa666f9281310d7e492d17ad2552c2cd18bfd6c75e.
• Direct item-weighted means are 0.780596 deep and 0.957211 quick, using 35,135,000 sampled tokens. Equal-cell macros are 0.787636 deep, 0.957211 quick, and 0.872423 joint. These exceed seed 42 by +0.195249, +0.150482, and +0.172865, and exceed the deep source by +0.187449, +0.146187, and +0.166818.
• Averaged across both sealed blocks, sample-more scores 0.779220 deep, 0.951708 quick, and 0.865464 joint. It beats seed 42 by +0.187748, +0.150731, and +0.169239, and the deep source by +0.183607, +0.141182, and +0.162394. The matched-compute loss is therefore large, replicated, and not a weighting artifact.
• Wrong-teacher started as the final confirmation arm with external journal hash 7f71ae4083bb45f1bda0c2148b0f9e061ce67dddd2446e7735bf222fb82959a5.

2026-07-15 — terminal confirmation rejects capability installation

• Wrong-teacher completed sealed generation and strict validation. GENERATED.json and COMPLETE.json hash to a6315632da88d74deaf1b6c90958b60e4cc6139fc81f04224b9b9be2b895d365 and 38357e20c5a7a8bae8110e44c928763fa0340b9c8f9245019427bcbf884d989a; the score hashes to 2667a17fe7395c6296a35b74d22396a94bfa724e1f675c219e2b298f322aa7ea.
• Its direct item-weighted means are 0.568835 deep and 0.795695 quick, using 4,408,929 sampled tokens. Equal-cell macros are 0.591619 deep, 0.795695 quick, and 0.693657 joint. Seed 42 beats it by +0.005901 joint in block 1 and +0.005312 pooled, with one-sided 95% LCB +0.000099.
• The terminal manifest hashes to aeedc23524fa81e92ffb9de4e05829e897f64ca2a5fb53c0d6af364f87aadfdf. The frozen analyzer authenticated all 26 arm-block score sets, passed every protocol check, and produced analysis hash faa1ab0d5f5af16ac252cc795ab98913bd5cc2e656040e2701e7c2b90018cdd3.
• Primary seed 42 has pooled joint deltas −0.006845 versus deep (one-sided 95% LCB −0.012839), −0.001300 versus soup, −0.001872 versus off-policy SFT, −0.003706 versus soup75, and −0.169239 versus soup best-of-eight (LCB −0.175468). Seeds 43/44 are also negative versus deep by −0.003450/−0.005660 pooled joint. All source-stratum domination cells fail.
• Seed 42 does beat quick (+0.033284 joint), non-advantage MOPD (+0.005619), the visible router (+0.010066), and wrong-teacher (+0.005312). These controls establish that routing and teacher identity contain directional signal, not that the update crosses the source frontier.
• Retention passes its 0.02 regression ceiling. Untouched transfer improves by +0.015625 on brinework and +0.010590 on spindle. These successes do not rescue the capability gate.
• The analyzer exited with the preregistered negative code 4 and emitted stop_before_benchmark_cli. No benchmark file, content, authorization, or run was opened. The experiment is terminal and finished.
• Strategic read: direct-bf16 deployment-parity microtraining with an explicit merge-survival/source/interpolation/sample-more gate is now prerequisite to more MOPD. Cross-fitted direct advantages, adaptive allocation, and a third untouched block remain necessary for two-teacher integration, but are not sufficient until the causal update kernel survives deployment.