State-Formation Capacity Adjudication

2026-07-13 — experiment created

• Created a fresh successor under structured_execution_and_compilers; neither prior result-bearing directory will be extended or overwritten.
• Status is in-progress. No model load, GPU mechanics call, procedural result corpus, training, evaluation, or scientific analysis has run in this directory.
• The triggering uncertainty is the post-result audit of qwen35_4b_state_carry_vs_state_bag_fullrank_delta: the raw full-rank state miss remains valid, but it did not isolate LoRA capacity because shared state-module initialization and dropout streams differed across experiments and non-capacity promotion gates failed simultaneously.

2026-07-13 — design correction: capacity-only verdict axis

• Removed Bag advantage, answer-gain, query-stratum, edge-cut, swap, and sample-more requirements from the capacity verdict. They address utilization or serial mechanism, not whether the representation formed.
• Retained the same Carry loop, procedural state targets, state-before-query placement, exact K=1 path, and joint node+phase+checksum readout.
• Froze fresh procedural splits and three fixed-final 1,500-step seeds 7411–7413. Prior rows and checkpoints are context only, never training data or result cells.

2026-07-13 — design correction: joint-first with conditional controls

• Rejected a state-only-only adjudication because it would change the original joint objective and could not answer whether LoRA rank caused that failure.
• Froze the sequential branch: three LoRA joint runs first; after a valid LoRA formation miss, three LoRA state-only controls plus three full-rank joint runs become mandatory; only a full-rank joint miss opens the three full-rank state-only controls.
• The maximum is 12 result-bearing runs, but controls and dense runs are prohibited until their identity-bound branch receipt exists.
• Joint retains answer/state/fixed-point weights 1.0/0.5/0.05. State-only omits answer loss rather than multiplying it by zero.

2026-07-13 — cross-capacity confound controls

• Froze one serialized common loop-state initialization bundle per seed. Every capacity/objective arm must reload it strictly and prove bit-identical names, shapes, dtypes, and tensor values before training.
• Replaced the divergent PEFT-versus-direct dropout paths with a common custom adaptation-hook backend. Both capacities use dropout 0.05 and the same capacity/objective-independent per-microbatch dropout schedule.
• Froze separate gradient clipping for shared loop-state parameters and adaptation parameters; dense adaptation gradients cannot change the clipping scale applied to shared modules.
• Added intact and adaptation-disabled state evaluation for every checkpoint. A capacity rescue must show a registered positive intact-minus-disabled effect.
• Setup-only overfit/readout controls must pass before result training. Failure is a repair stop, not a capacity result.

2026-07-13 — terminal-metric and conditional-contrast hardening

• Replaced pooled trajectory accuracy with exact terminal joint node+phase+checksum correctness at K equal to semantic depth. The 0.40 gate applies separately to every seed×depth cell; validation depth 1 is reported but excluded because K=1 bypasses adaptation.
• Initially froze 1,024 rows for each trigger/deep result evaluation split. Validation has 256 rows at each depth 1–4; depth extrapolation and joint family+surface shift have 128 rows at each depth 5–12. The later adversarial repair below adds the distinct 768-row sealed trained-depth split.
• Reserved seeds 73305 and 73306 as sealed 1,024-row post-trigger depth and joint-shift contrasts. They remain unscored until a valid LoRA miss is followed by all Stage-B fixed finals and trigger evaluations, including LoRA state-only, and a dedicated identity-bound seal receipt proves no prior contrast scoring access. If direct full shape misses any trigger cell, the seal mandates state-only and leaves those rows unopened. The LoRA-miss receipt licenses training but cannot license contrast evaluation, preventing conditional branch selection from contaminating the rescue comparison.
• Added deterministic-tensor output/gradient parity against pinned PEFT lora.layer.Linear before the common LoRA hook can be treated as compatible with the parent PEFT recipe. This does not load a second Qwen model.
• Named the affirmative conditional result DIRECT_FULLSHAPE_RECIPE_RESCUE; it is explicitly not mathematical rank identification.
• Replaced the arithmetically impossible 32-row “balanced” control with 48 setup-only rows: exactly two per depth 2/3/4 × query kind × training family × training template cell, seed 73991.
• Froze the exact design-boundary interface and canonical reports/design_receipt.json; it binds the intake, scientific contract/review, architecture/runbook/handoff, and config. Source/tests, implementation review, and the training lock must be clean at its recorded HEAD and are bound separately by downstream artifacts, so a mechanical repair can preserve the scientific boundary while invalidating all artifacts produced under the old implementation digest.
• Required tiered sealed-data validation: preauthorization stages may verify compressed identity but may not decompress or canonical-reopen contrast payloads.

2026-07-13 — adversarial repair: complete sealed-domain replication

• Found that the two original deep sealed splits could not retest a LoRA branch triggered solely by a trained-depth validation miss. Added contrast_validation at seed 73307 with 768 fresh rows over required depths 2–4 (256/depth). The complete sealed matrix now covers trained depth, depth extrapolation, and joint shift.
• Retained exactly six authorized capacity×seed contrast-evaluation jobs: each job evaluates all three sealed splits in intact and disabled modes. A sealed LoRA intact pass has terminal priority and emits LORA_TRIGGER_MISS_NOT_REPLICATED_ON_SEALED_CONTRAST regardless of full-rank score.
• Required category-matched replication before rescue: every LoRA trained, depth, or joint category that failed on trigger must fail again in its corresponding sealed domain. Additional sealed failures are allowed; a missing counterpart emits LORA_TRIGGER_FAILURE_CATEGORIES_NOT_REPLICATED_ON_SEALED_CONTRAST once full rank passes its absolute cells. Full-rank absolute failure still takes priority and mandates state-only.
• Made sealed-evaluation recovery executable rather than advisory. Each same-cell retry requires exactly one newly tracked, content-validated FAILED_ATTEMPT_ARCHIVED receipt appended to the existing access-ledger event; completed evaluations are ineligible for failure archival. Ledger revisions now hold a separate stable lock inode across atomic replacement, with the temporary file and parent directory fsynced, and initial access rejects an archive that predates its event.
• Hardened optimizer receipts so adaptation and common-state learning rates are logged separately and independently recomputed from the registered schedule at every exact step.
• Required post-contrast analysis to reopen the Stage-B-bound full-rank-joint, LoRA-joint, and LoRA-state-only trigger evidence and recompute their formation summaries. Stage C separately reopens the exact Stage-B LoRA-state-only and full-rank-joint predecessors before classification.
• Expanded the PEFT compatibility gate to both exact FP32/dropout-off/autocast-off and live-like bf16-autocast/dropout-0.05 output/A/B-gradient parity, with respective tolerances atol=1e-6/rtol=1e-5 and atol=2e-3/rtol=1e-2, plus a same-device-seed realized-mask protocol.
• Named ADAPTATION_DISABLED_REVERSAL for intact formation miss plus disabled formation pass. It is a reported adapter-interference diagnostic, never an adaptation-required or rescue result; both LoRA and full-rank sealed statuses remain visible.
• Recorded posttraining_and_adaptation as a secondary program fit because the experiment adjudicates update parameterization, objective interaction, and adapter dependence. Recurrent state formation remains the sole primary estimand under structured_execution_and_compilers.

2026-07-13 — implementation audit and CLI repair

• The first real CPU-smoke invocation found a pre-result false rejection: an empty authorization allowlist rejected the required omitted receipt. Implementation status was immediately returned to NO_GO; no model, result corpus, training, evaluation, or scientific analysis had run.
• Repaired authorization semantics so every preauthorization stage and LoRA-joint path accepts only the required omitted receipt, while all conditional paths require their exact canonical branch receipt. Added an early rejection for the unregistered state-only contrast cross-product.
• Hardened output construction and operator safety: missing analysis phases, state-only contrast, irrelevant phase/evaluation axes, conflicting --smoke stage selection, junk authorizations, and noncanonical outputs all fail before dispatch.
• The final implementation passes 130/130 experiment tests, including 23/23 static/CLI tests and 14/14 focused sealed-data/replay tests. An independent exhaustive audit verified all 23 normalized CLI cells and their canonical outputs. The real CPU smoke then passed with zero benchmark reads, no GPU/model load, and scientific_evidence: false.
• These changes harden execution mechanics only; they do not change the preregistered scientific design. Downstream source-bound artifacts must use the final post-repair source-contract digest.

2026-07-13 — clean-head freeze and setup artifacts

• Froze reports/design_receipt.json at clean committed head 86230605 with exact status DESIGN_FROZEN, zero benchmark reads, design identity d943b909250c2ebd377b8094bb55324a5d5ccf555c7559736526f013d248ac52, source-contract digest 903f19141ef982f0e90ea856edd72d75fb48d2dbd96e81dc166a6c32d4c14116, and training-lock digest 05546fe977583116d6169ea0dfa7b27e1184dd4a2b61d556dfb3f889d5b2b7b1.
• Generated the seven fresh deterministic splits. The tracked manifest records 12,000 training rows, 1,024 trigger rows in each domain, 768 sealed trained-depth rows, 1,024 sealed depth rows, 1,024 sealed joint-shift rows, zero cross-split structural duplicates, zero benchmark reads, and data contract 2eb5c134fb0818aeadb1cd9ac0074554cc26b5c010c248316c6e1eeb56f02291.
• Left all three sealed contrast payloads unopened after generation. Their empty access ledger has identity 238c0ace4e4cc929657de28130e0c2775f31faa2f096027e25ab8bd635ae922d.
• Created and independently reopened the three canonical common-state bundles. Seeds 7411, 7412, and 7413 have bundle SHA-256 values 2d3d1f11c78d561a4e623e0512326ad20560e10aaa192276611250b1d3b73497, c236d1f936b570447746a561286da643fe958a705aa26e17bd5fe0bbb7299c9e, and c4a20778d5ff803faff944bd3735a92869f132982bef2b51a5e5c0ecbbee5448, respectively; each tracked mirror is byte-identical to its external sidecar.
• Live preflight found exactly one idle RTX 6000 Ada, 49,138 MiB total minus 2 MiB in use, no GPU process, a compatible 76-package environment, and both required causal-convolution and flash-linear-attention kernels. All 130 experiment tests passed again. No model was loaded.

2026-07-13 — first live G0 stopped on revision-provenance defect

• The first seed-7411 LoRA G0 loaded only the requested Qwen/Qwen3.5-4B snapshot at pinned revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a, then stopped inside _load_base before wrapper construction or any mechanics probe. The training split had passed its permitted pre-G0 validator; no sealed contrast payload was opened, no positive control or training/evaluation step ran, and no canonical G0 output was created.
• Root cause: Transformers 5.13.0 resolves config.json through the exact pinned snapshot and its cache utility extracts the correct 40-character commit, but the Qwen3.5 causal-LM wrapper retains a derived text config whose runtime _commit_hash is None. The runner incorrectly treated that lossy field as the sole revision proof.
• Preserved the exact setup failure at runs/failures/g0_lora_seed7411_revision_provenance_failure.json. This is mechanics evidence only, never scientific evidence and never branch authorization.
• The fail-closed repair verifies the resolved commit and records the byte count and SHA-256 for config, tokenizer assets, safetensors index, and every indexed shard before model construction. A missing runtime config hash is accepted only after that complete proof; any non-null mismatch remains fatal. The real local snapshot currently resolves all nine files through the pinned commit.
• Because the repair changes the registered source contract, every data, initialization, and setup artifact produced under digest 903f19141ef982f0e90ea856edd72d75fb48d2dbd96e81dc166a6c32d4c14116 is invalidated and must be regenerated before retry. The frozen scientific design receipt remains unchanged.
• Preserved all 19 invalidated files (17,651,037 bytes) under the external invalidated_setup/ archive. The tracked archive receipt reopens every byte count and SHA-256 and has identity e29f2059058b621903c1bbf2933dd56aa0e500a0c73b70ffc8b19f72faa85014.

2026-07-13 — corrected-source setup regenerated

• Recreated CPU smoke, all seven data splits, the contrast-access ledger, and all three initialization bundles from committed source contract 9fd420f5f29fea2d9144bf50d3b187fc8e50d9acc9cb076656372281029614fb. CPU smoke again records zero benchmark reads, no model load, and scientific_evidence: false.
• Every procedural payload reproduced the exact prior content hash; the new source-bound manifest has SHA-256 fa871390b28a9a0119ad77957bc1b403cc7008aa7052a539e6788aa8d871be4c and data contract a57c05100ff51897dbfdc2e140b1eff6634513eaa873a222e2fd2d7f9b37ad2e. The new sealed-access ledger has no events and identity 14748750079830ea25e7935baf35a4689957b3bfb9db0b7630f643133bbfde8d.
• Seeds 7411, 7412, and 7413 preserve the exact common tensor-value digests while their new source-bound bundle hashes are fbb40b74d97fd07dd0e6590382f99834d5091e1467df1f3edda814092f386c77, b05cbafcf9a100dd80309a5b50309649372fe6841e32cdd38d37b15bdcaba6ed, and cd9923902e25f04ba5aca0ba001f036aa5ac6e8ae6792ac059d78773f22b1bf3, respectively. Every bundle reopened and each tracked receipt is byte-identical to its external sidecar.

2026-07-13 — G0 reached final persistence, then path shadowing stopped it

• The corrected seed-7411 retry passed pinned-snapshot proof and every registered in-memory G0 check, including PEFT output/gradient parity, K=1 parity, K=4/K=12 call geometry, two-step and live joint gradients, optimizer/clipping receipts, timing, finite worst-depth output, checkpoint roundtrip, RNG isolation, and VRAM headroom.
• The final _write_json(output, receipt) then received the last StateLoopOutput tensor container rather than the canonical Path: the two-step probe had reused the parameter name output as a local. Python function scoping retained that shadow through the final write. No canonical G0 receipt was created, so the in-memory pass is not durable evidence and authorizes neither positive control nor result training.
• Preserved the exact failure at runs/failures/g0_lora_seed7411_receipt_path_shadow_failure.json. The training payload was the only result payload opened; no sealed contrast was opened and no positive-control/training/evaluation stage ran.
• Renamed the local to probe_output and added an AST regression prohibiting any assignment to the canonical output parameter inside model_smoke. Audited all runtime sources: this was the only G0/path-parameter shadow. Also detached the PEFT diagnostic before scalar conversion to remove its non-semantic autograd warning.
• Archived every setup artifact bound to source contract 9fd420f5…614fb at large_artifacts/qwen35_4b_state_formation_capacity_adjudication/invalidated_setup/source_9fd420f5f29fea2d9144bf50d3b187fc8e50d9acc9cb076656372281029614fb/. The tracked archive receipt has identity 7f974f52e153eb2fe7a3985c8221cfa8f7cd2d8440eb63d608abf03a63651d47 and independently reopens all 19 files totaling 17,651,037 bytes.
• An independent audit found that all four hand-authored historical receipts had included a trailing newline in their claimed self-hash, unlike the runtime's compact canonical-JSON identity. Corrected the four identity fields without changing their evidence payloads and added a regression that revalidates every tracked runs/failures/*.json receipt against the runtime identity function.
• Final independent re-review is GO: the full suite passes 135/135, the static contract passes 27/27, both 19-file archives reopen byte-for-byte, and the post-review source contract is 3baa7b532d62bae4d9751dfe4be9c6ce314c11ca4524266f4cebea63289d5c42.

2026-07-13 — post-path-fix setup regenerated

• Recreated CPU smoke, all seven procedural splits, the seal ledger, and all three common-state bundles from committed source contract 3baa7b532d62bae4d9751dfe4be9c6ce314c11ca4524266f4cebea63289d5c42. CPU smoke SHA-256 is 8858671b077dd609141eb218b6932e9b8d7772877c2308df57c59d3d066ae467; it records zero benchmark reads, no model load, and no scientific evidence.
• The new source-bound data manifest has SHA-256 c2338a7aa2dad245683b6d3aebeef704199e9f143490830179263ea1fed57247 and data contract 5ff3255bebe18a4a73a3bd9d1db4c153e4727aa23e0e064cfd5ed3d685307968. Its exact seven-file projection reproduces the prior payload metadata at digest 3d52d6b31ffbd916cdd86495e0c14f0d11c293f868914e667d10a05acfecfba2. Independent local reopen validated all 15,072 non-contrast rows and decompressed zero sealed rows.
• The new seal ledger has file SHA-256 e2922cb1e51f58d378a5bc4fd6d49de391ea10f04fe06ac2c576424d91e2a85a, canonical identity 45f58325fbf7f362265bb6812cfeb23fc974bca5b7c25eea61d229d211446523, and events: [].
• Seeds 7411, 7412, and 7413 have bundle SHA-256 values 9ffbb459238ed42cfeef541c0b331b744318a1940a55c5a6a8bca133866774f4, 9d666538e98de6c2660d42221286c89de36fdd786d5223df4b932bf8a6425670, and 6ce43dffdc7c47dc8e8d7370e42a394bcb25d35de063bfa04ab9cea3983d4d1b. Their receipt identities are 136565d109b7b936452d9d91a4a64de565157320aa5b965e49030da05e253256, 361274c81e734c651897f5c8b69380f6dd1f2ef4906aa7c6d8191bb8a148572b, and 85699c6008c83704453acfaff8aee1de68e977eaf0fcb552151c04f206b0273d; every external sidecar is byte-identical to its tracked mirror and all tensor-value digests reproduce exactly.
• Independent audit is GO: it independently validated the source contract, manifest geometry, compressed-byte hashes, empty-ledger identity, all three bundle files and receipt mirrors, 15,072 non-contrast rows, and zero sealed row decompressions.

2026-07-13 — seed-7411 LoRA G0 passes durably

• The fresh retry wrote canonical receipt runs/setup/g0_lora_seed7411.json, file SHA-256 8495799beb226644f8c88b18f26510c7c4cfaed4117fb81d54b1cc94d4efac66, runtime identity e7394bcf48175c8232ac825a4f74f16734452602b9d8657be09ca47e293f3735, and exact status MODEL_SMOKE_PASS. The runtime's canonical loader reopens it against source contract 3baa7b53…d5c42, data manifest c2338a7a…247, design receipt d943b909…ac52, and initialization receipt 136565d1…3256.
• The pinned snapshot proof covers nine exact files at revision 851bf6e8…cd0a, 9,342,815,919 bytes, aggregate file-manifest digest 06486f260c8e9135835c768802e0e947bab0470e4454a36e138da6d10ed1fe12. The target manifest has 62 ordered modules and the LoRA bank has exactly 16,232,448 parameters.
• FP32/dropout-off and bf16/dropout-0.05 PEFT output/A-gradient/B-gradient parity each pass at zero observed error. Initial enabled-versus-disabled and both pre/post-update K=1 errors are zero. The second state-only step and the live joint probe give finite nonzero gradients to every required recurrent group and all 124 LoRA tensors, with zero base-model gradient tensors. All required Adam states are complete and finite.
• The ten-step probe is finite at 0.309 seconds/update. K=12 is finite with exactly 682 ordered calls; the setup row has zero structural overlap with result data. Destructive checkpoint restoration is exact, peak allocated/reserved memory is 11.20/11.36 GiB, and 35.63 GiB remains free after G0.
• This is setup evidence only. The model stage opened only the permitted training payload; the seal ledger remains byte-identical at e2922cb1…a85a, identity 45f58325…6523, with events: [].
• Independent receipt audit is GO: the runtime identity, all lineages, mechanics, environment, optimizer, restore, and firewall claims independently reopen with no blocker.

2026-07-13 — seed-7411 positive control stops at 0/48

• The fixed 256-update LoRA positive control reached its final gate, then raised tiny state-path overfit failed: 0.0 < 0.95. The oracle component necessarily passed its 0.99 gate because the run continued to the overfit check. No canonical positive-control receipt was created.
• Preserved the exact setup failure at runs/failures/positive_control_lora_seed7411_overfit_failure.json, identity 44397a2e278293bf54fe5d172ac4294c565a2a98ab7c8f4faaeb5ee044e8ec7c. It binds the G0, source, design, data, model, seed, 48 rows, 256 completed updates, and exact exception. It records no result payload read, no result training, and no sealed access.
• Independent shape/target tracing found no scorer-index defect. State targets and logits both have aligned [batch, step] geometry, and terminal joint scoring compares all three heads at the same final semantic step. Near chance, exact joint accuracy is 1/256, so zero of 48 is not surprising.
• The leading mechanics defect is an underpowered update path: the positive control performs one single-row microbatch per optimizer update instead of honoring the configured accumulation of 16. Across 48 rows, the failed recipe exposed each row only five or six times. It also persisted no losses, per-head metrics, gradient norms, optimizer state, or failure receipt before raising.

2026-07-13 — positive-control accumulation correction approved

• Three independent failure reviews found no target, tensor-shape, terminal-index, recurrence, gradient, or scorer defect. The failed setup control made one singleton presentation per optimizer update and omitted the globally frozen accumulation of 16. It therefore made 256 presentations, only five or six per high-entropy row, while preserving 256 optimizer updates.
• The frozen-boundary ruling permits exactly one narrow correction: the same 48 rows, canonical row hash, seed 73991, 256 optimizer updates, state-only objective, learning rate 2e-4, weight decay zero, dropout 0.05, clip thresholds, oracle threshold 0.99, final threshold 0.95, initialization, row order, and fixed-final decision now use 16 sequential singleton microbatches per optimizer update. Loss is divided by 16; each group is clipped once; the optimizer steps once; no early stop, selection, or second retune is allowed.
• The exact 4,096-event schedule gives the first 16 rows 86 exposures, the remaining 32 rows 85, and depth totals 1,368/1,368/1,360. Global indices feed the unchanged dropout-seed preimage. Fixed diagnostics at updates 0, 1, 16, 64, 128, and 256 score intact and adaptation-disabled modes while proving no parameter, train/eval-mode, or CPU/CUDA RNG mutation.
• Any reached exception now writes canonical and byte-identical tracked SETUP_CONTROL_FAILED receipts before re-raising. The receipts deny training and record zero benchmark, result-payload, and sealed-contrast access. Downstream training continues to accept only POSITIVE_CONTROL_PASS.
• The source-bound suite at the archive boundary passed 154/154 tests: 13 focused positive-control tests, eight durable invalidation tests, and 27 static/CLI/provenance tests. Independent code and GPU/runtime audits both give GO. This changes setup mechanics only; the frozen scientific design is unchanged.

2026-07-13 — source-3baa setup archived after control correction

• Before deletion, the complete source-3baa7b53…d5c42 setup reopened as exactly 20 files totaling 17,775,495 bytes with file-manifest identity 115b9dd5e810cb1bcf88e58a4da366370cbb48c4a1fc1adfcdcb37a5283e1a25 and an empty seal ledger.
• The durable archive is large_artifacts/qwen35_4b_state_formation_capacity_adjudication/invalidated_setup/source_3baa7b532d62bae4d9751dfe4be9c6ce314c11ca4524266f4cebea63289d5c42/. Its tracked mirror is runs/failures/invalidated_setup_source_3baa7b53.json, with receipt identity 1daa86e02f7d3f3c612c2f1ec01db89e4967b5b9ce7eb19ba75c752fe1e283aa. Both receipts and every archived byte were independently rehashed before current setup files were removed.
• The receipt binds invalidated source 3baa7b53…d5c42 to replacement source fe82f8cfe13d656a588c8f2766fc5519820edd884c9b855dcf7a16a14af547d5 and to the preserved 0/48 failure identity 44397a2e…e8ec7c. That failure remains outside the archive and unchanged.
• Current CPU smoke, data manifest, access ledger, initialization bundles, and G0 receipt are absent by design. No sealed payload was decompressed during archival; the archived ledger has events: []. No positive-control pass or result artifact exists.

2026-07-13 — post-archive directory cleanup hardened

• The first repository gate after archival stopped because successful deletion left empty runs/cpu_smoke/ and runs/setup/ directories on the local filesystem. Git cannot transport empty directories, so CI would have observed a different tree. The archive itself and both receipts remained exact.
• Adversarial follow-up reproduced partial deletion that could not resume and symlinked canonical roots that could redirect cleanup. The final helper uses the immutable archive receipt as its journal, verifies archive-only recovery before recreating the byte-identical tracked receipt, accepts every live source only as exact or absent, resumes unlink/rmdir/fsync interruption, and rejects unknown or tampered state, receipt/archive mismatch, special entries, and symlinks in every canonical root, ancestor, or evidence leaf.
• Twenty-five invalidation/recovery tests exercise those states. The full source-bound suite passes 171/171, both independent recovery re-audits give GO, and a real completed replay leaves the 20-file archive plus both receipt bytes and mtimes unchanged.
• No setup artifact was created under either intermediate post-archive source. The final source contract for regeneration is 1d1368cf064689322d9df7f345e67b026cecccc32d3a7b7514b82f253d434b0a.

2026-07-13 — final-source setup regenerated

• Recreated CPU smoke, all seven deterministic procedural splits, the contrast-access ledger, and all three common-state bundles under final source contract 1d1368cf064689322d9df7f345e67b026cecccc32d3a7b7514b82f253d434b0a. The CPU receipt has SHA-256 56032f75f8c3be863b43d1d6fc46a27199c8162961bfbe49b146d41e0ac7ad43, records CPU_SMOKE_PASS, zero benchmark reads, no model load, and no scientific evidence.
• The manifest has SHA-256 85286a9549b0e74111eaa0c1254f04f68cad4d75159abc8ba59c169637280cd9 and data contract 891ad7848ba23429ed5c7d66f2d6f4f85e8cc6560044b56f51f218bff957e9c8. All seven compressed payload hashes reproduce the archived source-3baa data exactly. Validation reopened all 15,072 non-contrast rows, found zero cross-split structural duplicates, and decompressed zero sealed rows.
• The contrast ledger has file SHA-256 0c03a0d8e0b665415bf956856296a2750ce1fe7facd2335d6a711f96f5aac261, identity b122d49054cd37474784a8b4ec3bcc0026e745321718dee799615e3051983c14, and events: [].
• Seeds 7411, 7412, and 7413 have bundle SHA-256 values 64f6ac313ee46335c4e8869ac3333136b4fe1739c8c59954b81f971fb96a2971, ef662c4a3f184b5d18a88142362a9803b51627e2bb59b4d0180afb331f5820ae, and 29f35ec72113039aed8c278a9166c5a70e0c9c575b88d3c077d62dacc3ce8f9a; receipt identities are bbb21fef784beb1bf9dc50de27ff41624a85904c39191f2288e94cb7a8ad10ae, a5d1a3fdc574ac4968709c65c49b59e859b03ab4226d847dde40fdbdc98ce3f7, and 42f9f66b848bcadd6e320db5421d2dfe39572793511b2b38151f6163a92e19e7. The canonical loader independently reopened each bundle, every tracked receipt is byte-identical to its external sidecar, and the three tensor-value digests exactly match the archived source-3baa bundles.
• No model-bearing stage ran during regeneration. No result training is licensed, and the sealed ledger remained unchanged throughout initialization validation.

2026-07-13 — final-source seed-7411 LoRA G0 passes

• The canonical receipt runs/setup/g0_lora_seed7411.json has file SHA-256 8c381fc7e883e384debef6ce64e55dd57aacab6f30e4a17dd6ef82d4db56aa01, identity 928e756fa6aea30104f0ed8a71c03620ed8663198a4811f6fcf9b6fda53f820c, and exact status MODEL_SMOKE_PASS. It reopens against final source 1d1368cf…434b0a, data manifest 85286a95…0cd9, design identity d943b909…ac52, seed-7411 initialization identity bbb21fef…0ae, the pinned training lock, and only Qwen/Qwen3.5-4B revision 851bf6e8…cd0a on the Transformers backend. No branch authorization was supplied or required.
• The pinned snapshot proof covers the same nine exact model/tokenizer files at the registered revision. The target manifest contains 62 ordered modules and the LoRA bank contains 16,232,448 trainable adaptation parameters, with every output factor still exactly zero at construction.
• Pinned-PEFT output, A-gradient, and B-gradient comparisons pass with zero observed error in both FP32/dropout-off and bf16-autocast/dropout-0.05 regimes. Enabled-versus-disabled output and both pre/post-optimizer K=1 errors are zero, with zero K=1 adaptation calls.
• Both state-only optimizer probes and the live joint probe give finite nonzero gradients to every required recurrent group and all 124 LoRA tensors while producing zero base-model gradient tensors. Every required Adam state is complete and finite with no registered exemption.
• The ten-step path is finite at 0.311 seconds per step. The structurally disjoint setup-only K=12 row is finite with exactly 682 ordered calls across 11 identical target cycles. Destructive checkpoint restoration returns both adaptation and common-state digests exactly and changes logits by zero. Peak allocated/reserved memory is 11.20/11.36 GiB, leaving 35.63 GiB free.
• The empty contrast ledger remains byte-identical at file SHA-256 0c03a0d8…ac261, identity b122d490…3c14, with events: []. G0 opened only the permitted training payload and created setup evidence, not scientific evidence. Independent receipt audit recomputed every lineage, all nine pinned snapshot files, parameter geometry, mechanics gate, and ledger invariant and gives GO.

2026-07-13 — corrected seed-7411 LoRA positive control passes 48/48

• The canonical receipt runs/setup/positive_control_lora_seed7411.json has file SHA-256 f71ee195b3e326abe59daca019b266932990f17adff20ceba8b414f5ac8c88ec, identity 8db4595e1473d2f26d1d99d2ac1bb9be45947720d676bf43cc697cb90b402df7, exact status POSITIVE_CONTROL_PASS, and exact G0 lineage identity 928e756f…820c. It binds final source, data, design, initialization, model, revision, backend, and training lock with no branch receipt.
• The oracle-coded readout scores 1.0. The setup corpus is the frozen 48-row factorial grid with two rows in each depth 2/3/4 × query kind × family × template cell, canonical row digest 581dadcb…ef41, and zero structural overlap with every result split.
• The corrected path completes exactly 256 optimizer updates, 16 sequential singleton microbatches per update, and 4,096 total presentations. There are 256 clips and optimizer steps per parameter group, 257 zero-gradient calls, no early stop, and no checkpoint selection. The first 16 rows have 86 exposures, the remaining 32 have 85, and depth exposures are exactly 1,368/1,368/1,360.
• Row order, dropout schedule, and optimizer-event digests are respectively b3ece16c…e6fa, 8dc49101…98c, and ae8322bc…b6b. Dropout probes at presentations 1, 2,048, and 4,096 have the registered call geometry. Optimizer probes at updates 1, 16, 64, 128, and 256 show finite positive group norms, constant learning rate 2e-4, zero base trainables, finite nonzero gradients for every required recurrent and LoRA tensor, and the registered answer-only aggregate exemption.
• Fixed intact/disabled diagnostics at updates 0, 1, 16, 64, 128, and 256 leave parameters, model mode, and CPU/CUDA random streams unchanged. Intact terminal joint accuracy progresses from 0.00 to 0.00, 0.0625, 0.75, 1.00, and 1.00. The fixed final is 48/48, exceeding the 0.95 gate; the adaptation-disabled fixed final remains 0/48.
• Adam state is complete and finite for every required parameter. Adaptation input/output and common state move by L2 norms 9.046, 8.905, and 3.425, and the final trainable-value digest differs from initialization. The receipt authorizes training but is setup evidence only: benchmark reads are zero, result/sealed payload lists are empty, and the ledger remains byte-identical at 0c03a0d8…ac261, identity b122d490…3c14, with events: []. Independent audit regenerated all 48 rows and all 4,096 schedule events, recomputed the full row/dropout digests, reopened every lineage and diagnostic invariant, and gives GO.

Current authorization

Seeds 7412 and 7413 LoRA G0 and their identical positive controls are now authorized. No result training is authorized until all three per-seed setup gates pass canonically.

2026-07-13 — seed-7412 LoRA G0 stops on aggregation-scalar precision

• Seed 7412 passed the pinned-snapshot load, zero-function and K=1 checks, K=4 call geometry, both state-only optimizer probes, a finite live-joint forward/backward, and the registered joint dropout schedule. It then stopped before joint clipping or stepping at the frozen all-groups reachability gate. No canonical runs/setup/g0_lora_seed7412.json was created.
• All 124 LoRA tensors, four initializer tensors, the step encoder, all eight sufficiency tensors, and the damping scalar had finite nonzero gradients. The frozen base remained gradient-free. aggregate_logit.grad existed and was finite but its norm was exactly zero, so the joint all-groups gate correctly failed even though the state-only-oriented helper excludes that scalar from its all_required_tensors_finite_nonzero summary.
• Preserved the failure at runs/failures/g0_lora_seed7412_live_joint_gradient_reachability_failure.json, identity ce3406f8fa788c08421687d5d6a0843a2eb7035fd254d5086123d895b2bb634c. It binds the final-source contract, design, data, model/revision/backend, seed-7412 initialization, reached and unreached checks, train-only access, zero benchmark/sealed access, and denies every downstream authorization. This is setup-mechanics evidence, not a LoRA or state-formation result.
• Root-cause review found that the FP32 aggregate_logit is sigmoid-transformed and cast to BF16 before broadcasting through the registered last-state/mean-state convex mix. Two otherwise matched seed-7411 G0 runs produced scalar gradients whose de-sigmoided magnitudes are exactly 2^-12 and 2^-11, while seed 7412 produced zero with an allocated gradient tensor. This is consistent with a one-example BF16 projection/reduction quantization defect and argues against simple graph disconnection or magnitude underflow; the regenerated live gate remains the falsification test.
• An unchanged retry is prohibited. The only reviewed repair keeps the existing BF16 mean arithmetic, forms the scalar convex mix in FP32, and casts its completed result once back to model dtype. It preserves K=1, parameters, initialization, row, masks, objective, schedule, thresholds, and both capacity arms; the nonzero joint-scalar gate remains strict. Automatic G0 failure persistence must also be added before retry.
• Any repair changes the source digest. After repair tests and independent review pass, every current source-1d1368cf…434b0a setup artifact must be archived, setup regenerated, and G0/control replayed from seed 7411. Seed 7413 and all result-bearing stages remain prohibited.

Current authorization

No model-bearing stage is authorized until the aggregation-precision and G0-failure-persistence repair is reviewed and the source-bound setup is regenerated. No result training is authorized.

2026-07-13 — aggregation precision and G0 persistence repair passes review

• Implemented the reviewed narrow numerical repair under source contract d4269bf34f7c80affcc8c1e8a33fee9afddcd912d1bd9dead223e520ee108b36. The registered recurrence states and their mean retain the live BF16 dtype. Only the FP32 sigmoid scalar's convex mix of the already-computed last and mean states runs in FP32 with autocast disabled, followed by one cast back to model dtype. K=1, state-only optimizer exemption, parameters, RNG, calls, rows, objectives, schedules, thresholds, and capacity symmetry are unchanged.
• A deterministic width-256 BF16 cancellation case gives the legacy scalar gradient exactly zero and the production helper's gradient 0.0450000092, matching the analytic FP32 derivative. The same check passes directly on the registered RTX 6000 Ada. Random width-2,560 CUDA probing confirms the intended forward effect is at BF16 rounding scale and common to both capacities. The frozen live G0 gate remains strictly nonzero; a repeated zero after regeneration is a hard stop, not permission to add rows or weaken the gate.
• Refactored G0 into a thin persistence guard and instrumented attempt. Every reached failure binds only the lineage actually validated, retains structured gradient/dropout progress, denies all downstream authorization, and re-raises. One serialization is written into two independently fsynced staging files; the source-qualified mirror is installed first and the canonical failure second without replacement. Their final inodes are independent, and existing files, broken symlinks, symlinked ancestors, and partial-link states fail closed.
• The setup invalidation helper accepts canonical G0 pass or failure receipts. Reached lineage must be exact; null lineage is allowed only at its corresponding early stage and may not contradict the completed-check list. A failed G0 requires its identical nonsymlink mirror and cannot coexist with a positive control. The historical seed-7412 manual receipt remains tracked-only and unchanged.
• Source-contract version 6 includes the new tests. The full suite passes 201/201: 38 invalidation/ recovery, 14 G0 persistence/authorization, three aggregation precision, and 27 static/CLI/ provenance tests. Independent numerical, failure-persistence, and real-artifact archive audits give GO.
• A read-only rehearsal validates the real 21-file source-1d1368cf…434b0a setup totaling 18,288,790 bytes and trigger identity ce3406f8…b634c against replacement source d4269bf3…8b36. Nothing has yet been moved. The next action is to publish this repair, archive that exact setup, regenerate all source-bound artifacts, and replay from seed 7411.

Current authorization

Archival of source-1d1368cf…434b0a setup is authorized after this repair is published and green in CI. No model-bearing stage or result training is authorized until replacement setup is regenerated.

2026-07-13 — source-1d setup archived under the repaired source

• Commit 799fb98b passed both repository-validation and site-publish workflows before mutation.
• Archived the exact 21-file source-1d1368cf…434b0a setup (18,288,790 bytes; files identity d05f410c…9d88) under replacement source d4269bf3…8b36, using the preserved seed-7412 G0 failure identity ce3406f8…b634c as the external trigger.
• The durable and tracked archive receipts are byte-identical at SHA-256 e899bc46…9484, receipt identity 13cdcaec…2050b. Canonical CPU-smoke, generated-data, initialization, G0, and control paths were deleted only after archive verification. The generated-data directory retains only its tracked .gitignore; both tracked setup directories are absent; all source-1d initialization payloads now live only in the immutable archive.
• Replayed the archival command as a read-only idempotency check; it reopened the complete archive, trigger, tracked receipt, and cleanup postconditions without rewriting the archive.

Current authorization

Replacement-source CPU smoke, deterministic data/empty-ledger generation, and all three shared initialization bundles are authorized. No G0 or model-bearing stage is authorized until that entire setup is regenerated and strictly reopened under source d4269bf3…8b36. No result training is authorized.

2026-07-13 — replacement-source setup regenerated

• CPU smoke passed under source d4269bf3…8b36 without loading a model; receipt SHA-256 is 1d5a57c9…6fdb.
• Regenerated all seven deterministic splits with zero cross-split structural duplicates and zero benchmark reads. Manifest SHA-256 is e935c31a…1e57, data contract is 8e95991b…d5b, and the access-ledger identity is eaf7ba23…b3cb with events: []. The strict reopen checked every compressed payload but decompressed no sealed contrast split.
• Regenerated and canonically reopened all three shared initialization bundles. Bundle SHA-256 / receipt identity: seed 7411 539e8e6c…5dc26 / a79d0f9b…07f8a; seed 7412 bf916363…15e9d / 28e7d6db…36d86; seed 7413 59699025…2593c / c4e8657b…40ffa. Every sidecar is byte-identical to its tracked mirror, and every tensor-value digest exactly equals the corresponding source-1d bundle.

Current authorization

Seed-7411 LoRA G0 is authorized as the first model-bearing replay under source d4269bf3…8b36. Later seeds, positive controls, and all result training remain blocked until their exact predecessor setup gates pass and are committed.

2026-07-13 — replacement-source seed-7411 LoRA G0 passes

• Canonical receipt SHA-256 5127c305…ecedd, identity 185835ee…3216, under source d4269bf3…8b36 and manifest e935c31a…1e57.
• The repaired aggregate_logit has a present, finite, nonzero live-joint gradient norm 5.328678525984287e-5. Every one of 124 LoRA tensors and every other required recurrent group is finite/nonzero; the frozen base has no gradient tensors.
• Both K=1 parity probes are exact with zero adaptation calls, the K=12 setup path is finite with the exact 682 adaptation calls, and destructive checkpoint restoration returns exact tensor digests and zero logit error. The setup receipt records train-only access, zero benchmark/sealed access, and authorizes only the seed-matched positive control.

Current authorization

The seed-7411 LoRA positive control is authorized. Seeds 7412/7413 G0 and all result training remain blocked until this setup-control replay passes and is committed.

2026-07-13 — replacement-source seed-7411 positive control passes

• Canonical receipt SHA-256 9dcd28ee…14850, identity 6a1394d9…cefa, with exact lineage to G0 identity 185835ee…3216 and manifest e935c31a…1e57.
• Oracle readout accuracy is 1.0. The fixed-final intact score is 48/48 exact terminal triples and all 144 trajectory steps; adaptation-disabled is 0/48 exact terminal triples. The path therefore still exercises the LoRA update after the FP32 aggregation repair.
• The control completed exactly 256 optimizer updates, accumulation 16, and 4,096 singleton presentations. It opened no result payload, benchmark, or sealed contrast and remains setup-only evidence.

Current authorization

Seed-7412 LoRA G0 is authorized under source d4269bf3…8b36. Seed 7413, later positive controls, and all result training remain blocked until their exact predecessor setup gates pass and are committed. The unchanged nonzero aggregation-gradient gate is mandatory.

2026-07-13 — replacement-source seed-7412 LoRA G0 validates the precision repair

• Canonical receipt SHA-256 3ef95869…aa2ca, identity 737a8b39…0a89f, under source d4269bf3…8b36 and manifest e935c31a…1e57.
• The historically exact-zero aggregate_logit gradient is now present, finite, and nonzero at 6.673135794699192e-5. All other live-joint trainable groups remain finite/nonzero, the frozen base has no gradients, K=1 is exact, K=12 is finite with 682 calls, and checkpoint roundtrip error is zero.
• The unchanged gate passed without a tolerance or exemption. This supports the preregistered BF16 projection/reduction explanation and resolves the mechanics question; it is not scientific evidence about LoRA capacity.

Current authorization

The seed-7412 LoRA positive control is authorized. Seed 7413 and all result training remain blocked until the seed-7412 control passes and is committed.

2026-07-13 — replacement-source seed-7412 positive control passes; result boundary stays closed

• Canonical receipt SHA-256 c76da222…a254, identity 02a329d9…669a, with exact lineage to G0 identity 737a8b39…0a89f and manifest e935c31a…1e57.
• Oracle readout accuracy is 1.0. The fixed-final intact score is 48/48 exact terminal triples; adaptation-disabled is 0/48. The run completed exactly 256 optimizer updates, accumulation 16, and 4,096 singleton presentations, with parameter changes in the LoRA and common-state groups.
• The receipt records zero benchmark, result-payload, and sealed-contrast access and remains setup-only evidence under source d4269bf3…8b36.
• A read-only pre-result authorization re-audit then found that generic receipt loading does not re-enforce every producer-written model/backend, access, scientific-evidence, and downstream authorization field. It also found stage-ordering firewalls that exist in the runbook but are not all enforced at the execution boundary. No result was run.
• The seven fail-open classes are preserved under the still-current source contract as a nonauthorizing audit-failure receipt at identity 18892121…3a13. It is the canonical trigger for archiving source-d426 setup after the repaired source is published.

Current authorization

No further model-bearing stage or result training is authorized. Preserve this source-d426 setup, repair and adversarially test the authorization consumers, archive every invalidated source-d426 artifact, regenerate source-bound setup, and replay its gates in order before scientific execution.

2026-07-13 — source-v8 published; source-d426 setup durably archived

• The first source-v7 archive preflight stopped before mutation because successful legacy positive- control receipts omit the later explicit result-evaluation deny bit. Source v8 accepts only that legacy-safe omission or exact boolean false, rejects true, null, and integer substitutes, and removes the redundant unignored per-source lock while retaining the shared runs/run.lock.
• Source-v8 reviewed implementation f9364c36…b9d873 and full contract 7991d46a…b1cc88 pass 357/357 tests and the exact machine GO. Commit ee729def was pushed to main; both Validate Repository and Publish Research Site completed successfully before archival resumed.
• The registered transaction preserved 23 source-d426 files totaling 18,927,960 bytes. The archive files identity is 1538f2f2…ec3ed0, tracked receipt file SHA-256 is 9aa04d35…efc1a1, and receipt identity is e7a71362…818b77; it binds replacement source 7991d46a…b1cc88 and trigger identity 18892121…3a13.
• Independent verification recomputed every payload hash and both receipt identities, matched exact archive and 23-leaf zero-quarantine membership, confirmed empty canonical CPU/setup directories, only .gitignore under generated data, and no live initialization bundle. An immediate idempotent rerun revalidated the same evidence without further mutation.

Current authorization

Only archive-checkpoint validation, commit, push, and workflow verification are authorized. Do not regenerate CPU smoke, data, initialization, or any model-bearing setup until that checkpoint is green. No result training is authorized.

2026-07-13 — source-v9 makes retained setup directories clone-stable

• Repository validation correctly rejected the archive's empty retained runs/cpu_smoke and runs/setup directories because Git cannot represent them and CI would see a different tree.
• Source v9 adds one tracked empty .gitkeep per directory. The archiver requires an empty, inode-distinct stable regular file; excludes it from archive evidence; permits no other residue; preserves it through cleanup; and revalidates it at the terminal postcondition. Missing, nonempty, hardlinked, symlinked, and extra entries fail closed.
• Reviewed implementation 1c73fbf9…d5ad3b and full source contract 5629a3a4…99e236 have exact machine GO. No setup was regenerated and no model or sealed payload was opened during this structural repair.

Current authorization

Only combined source-v9/archive-checkpoint validation, commit, push, and workflow verification are authorized. Regeneration and every model-bearing or result stage remain blocked until both workflows are green.

2026-07-13 — source-v9 non-model setup regenerated

• Combined source-v9/archive commit 9932a560 passed both Validate Repository and Publish Research Site before regeneration began. The box exposed one idle RTX 6000 Ada with 49,140 MiB total; the pinned environment and required training extensions passed preflight.
• CPU smoke passed without loading a model at SHA-256 1655354b…45a38, source 5629a3a4…99e236, and zero benchmark reads.
• All seven deterministic splits regenerated with zero structural duplicates. Manifest SHA-256 is 957013ad…a4517, data-contract identity is 3677a0a3…ae761, and the sealed-access ledger remains events: [] at identity eb8028bf…9c84dc. Post-generation inspection read only manifest and compressed-byte metadata, not sealed rows.
• All three shared initialization bundles reopened through the canonical loader, have byte-identical external sidecars and tracked mirrors, and exactly reproduce their archived tensor-value digests. Bundle SHA-256 / receipt identity: seed 7411 5ed9d5c6…0e1b8 / 74ddebb1…31413; seed 7412 15366ea6…dcb2c / 01dd7e7c…1ee4b; seed 7413 bda608a2…bf8b4 / a7ab7d5d…e912f.

Current authorization

Only non-model setup validation, commit, push, and workflow verification are authorized. LoRA G0, positive controls, result training, evaluation, and analysis remain blocked until both workflows are green.

2026-07-13 — source-v10 seed-7411 LoRA setup pair passes

• Non-model setup commit 33abfe33 passed both required workflows before the model-bearing command.
• Seed-7411 G0 passed at file SHA-256 efde2db9…420f9, receipt identity 0fe46a0c…7a383. It binds the pinned Qwen revision, source v10, manifest 0b1cca35…7422a, and seed-7411 initialization e202efb8…3e0d8.
• The exact 62-target, 16,232,448-parameter LoRA setup passed both PEFT parity regimes with zero observed output/A-gradient/B-gradient error. K=1 and destructive checkpoint restoration have zero logit error. The live joint probe gives finite nonzero gradients to all 124 LoRA tensors and every required common-state tensor, no base tensor, and aggregation-scalar norm 2.2352195e-5; K=12 is finite. Peak allocated memory was 11.197 GiB.
• The seed-matched positive control passed at file SHA-256 04b7f995…a34f6, identity 18d02610…ebe89: oracle accuracy 1.0; fixed-final intact 48/48; disabled 0/48; exactly 256 updates, accumulation 16, and 4,096 presentations; adaptation and common-state parameters both changed.
• Both receipts record zero benchmark and sealed-contrast access and are setup-only evidence. The control's seed-specific training authorization does not satisfy the three-seed setup barrier.

Current authorization

Only seed-7411 setup-pair validation, commit, push, and workflow verification are authorized. Seed 7412 G0/control may begin only after both workflows are green. Result training, evaluation, and analysis remain blocked.

2026-07-13 — source-v10 seed-7412 LoRA setup pair passes

• Seed-7411 setup commit d0642d4a passed both required workflows before seed 7412 began.
• Seed-7412 G0 passed at file SHA-256 e2ef4951…fc18f, receipt identity 4af55cc3…e30de, with exact lineage to initialization bundle ab0b70c1…9169b and the unchanged source/manifest.
• Both PEFT parity regimes, K=1 before/after, and destructive checkpoint roundtrip have zero observed error. All 124 LoRA tensors and every required common-state group have finite nonzero live-joint gradients, the base has none, aggregation-scalar norm is 1.0440836e-4, and K=12 is finite.
• The seed-matched positive control passed at file SHA-256 1097d31b…c41e5, receipt identity ae4fdb5f…e461b: oracle accuracy 1.0; fixed-final intact 48/48; disabled 0/48; exactly 256 updates, accumulation 16, and 4,096 presentations; parameter values changed.
• Both receipts record zero benchmark and sealed-contrast access and remain setup-only evidence. The complete setup barrier still lacks seed 7413.

Current authorization

Only seed-7412 setup-pair validation, commit, push, and workflow verification are authorized. Seed 7413 G0/control may begin only after both workflows are green. Result training, evaluation, and analysis remain blocked.

2026-07-13 — source-v10 seed-7413 LoRA setup pair passes; setup matrix complete

• Seed-7412 setup commit 41d587ec passed both required workflows before seed 7413 began.
• Seed-7413 G0 passed at file SHA-256 b1a42d2e…5b9256, receipt identity 1909508f…b47e43, with exact lineage to initialization bundle 7dca25ea…a7c5c and the unchanged source/manifest.
• Both PEFT parity regimes, K=1 before/after, and destructive checkpoint roundtrip have zero observed error. All 124 LoRA tensors and every required common-state group have finite nonzero live-joint gradients, the base has none, aggregation-scalar norm is 1.5991002e-3, and K=12 is finite.
• The seed-matched positive control passed at file SHA-256 5ea3a552…fa01c7, receipt identity 7ed7332a…4f85cc: oracle accuracy 1.0; fixed-final intact 48/48; disabled 0/48; exactly 256 updates, accumulation 16, and 4,096 presentations; parameter values changed.
• All three source-v10 LoRA G0/control pairs now pass. Every setup receipt records zero benchmark and sealed-contrast access and remains setup-only evidence; no scientific training has started.

Current authorization

Only final setup-matrix validation, commit, push, and workflow verification are authorized. After both workflows pass, run the three Stage-A LoRA joint cells in frozen seed order. Evaluation, analysis, and all conditional branches remain blocked until their exact upstream barriers authorize them.

2026-07-14 — Stage-A seed-7411 LoRA joint training completes

• Complete setup commit 5b7c22fe passed both required workflows before the first scientific cell began. The source-v11 digest remained exactly 5a8ed26d…6666, main was aligned, and the sole RTX 6000 Ada was idle with 48.6 GiB free.
• The source-v11 repaired launch boundary durably progressed from its authorized attempt through terminal COMPLETE; there was no retry or partial-recovery path. Exactly 1,500 fixed updates ran in 9,217.945800065994 seconds. Interim scientific metric values were not inspected during execution.
• The production classifier reopens the cell as COMPLETE with no errors. Run SHA-256 is 19cac1e2…7aa60, run identity 6db22c2f…f81f1, checkpoint-metadata SHA-256 0c68e7a5…14966, and checkpoint identity 174f3c69…cf9e.
• Adaptation-state SHA-256 is 9af35963…9e71f; loop-state SHA-256 is 333f8c44…50a5; optimizer journal SHA-256 is 45664bad…e5a3; metrics SHA-256 is 2b97d188…8ade. Each of the four tracked mirrors is byte-identical to but inode-distinct from its external counterpart.
• All 1,500 optimizer rows report finite gradients and zero base trainable parameters. The cell binds setup barrier eb1fa85a…bc77, seed-7411 G0/control, source v11, and the exact initialization. It opened only train, no benchmark or sealed contrast. It does not authorize evaluation until all three Stage-A training cells are complete.

Current authorization

Only seed-7411 training-cell validation, commit, push, and workflow verification are authorized. After both workflows pass, run Stage-A seed 7412. Evaluation, analysis, and conditional branches remain blocked on the complete Stage-A barrier.

2026-07-14 — Stage-A seed-7412 LoRA joint training completes

• Seed-7411 training commit 72d526d8 passed both required workflows before the second scientific cell began. Source v11 remained exactly 5a8ed26d…6666, local and remote main were identical, and the sole RTX 6000 Ada was idle.
• The authorized seed-7412 attempt reached terminal COMPLETE without retry or recovery. Exactly 1,500 fixed updates ran in 9,202.148542404175 seconds. Interim scientific metric values were not inspected during execution.
• The production classifier reopens the cell as COMPLETE with no errors. Run SHA-256 is 0bb1680e…1e05, run identity f7d049c6…d0ca, checkpoint-metadata SHA-256 1b588d9f…e9fe, and checkpoint identity 90750967…0324.
• Adaptation-state SHA-256 is cba9fb80…8c74; loop-state SHA-256 is 03f4b811…8076; optimizer journal SHA-256 is 32109aa7…948a; metrics SHA-256 is ce8e7d32…348a. Each of the four tracked mirrors is byte-identical to but inode-distinct from its external counterpart.
• All 1,500 optimizer rows report finite gradients and zero base trainable parameters. The cell binds setup barrier eb1fa85a…bc77, seed-7412 G0/control, source v11, and the exact initialization. It opened only train, no benchmark or sealed contrast. It does not authorize evaluation until all three Stage-A training cells are complete. The realized-artifact suite passes 363/363.

Current authorization

Only seed-7412 training-cell validation, commit, push, and workflow verification are authorized. After both workflows pass, run Stage-A seed 7413. Evaluation, analysis, and conditional branches remain blocked on the complete Stage-A barrier.

2026-07-14 — Stage-A seed-7413 LoRA joint training completes; barrier complete

• Seed-7412 training commit 07547f4e passed both required workflows before the final Stage-A cell began. Source v11 remained exactly 5a8ed26d…6666, local and remote main were identical, seeds 7411/7412 classified COMPLETE, seed 7413 classified ABSENT, and the sole GPU was idle.
• The authorized seed-7413 attempt reached terminal COMPLETE without retry or recovery. Exactly 1,500 fixed updates ran in 9,527.480490207672 seconds. Interim scientific metric values were not inspected during execution.
• The production classifier reopens the cell as COMPLETE with no errors. Run SHA-256 is 8bbdb3b3…8b59, run identity 07f91939…362b, checkpoint-metadata SHA-256 25936f83…fbfd, and checkpoint identity d65d5bb3…e68d.
• Adaptation-state SHA-256 is 2ed8167f…9a65; loop-state SHA-256 is 97d59a7f…b298; optimizer journal SHA-256 is 44f9eaa2…fc9c; metrics SHA-256 is 6afe4d38…cd38. Each of the four tracked mirrors is byte-identical to but inode-distinct from its external counterpart.
• All 1,500 optimizer rows report finite gradients and zero base trainable parameters. The cell binds setup barrier eb1fa85a…bc77, seed-7413 G0/control, source v11, and the exact initialization. It opened only train, no benchmark or sealed contrast. The realized-artifact suite passes 363/363.
• All three Stage-A cells classify COMPLETE with no errors. Their complete training-barrier identity is 31d7f6f3…6451.

Current authorization

Only final Stage-A training validation, commit, push, and workflow verification are authorized. After both workflows pass, evaluate the three LoRA joint checkpoints on the trigger set in frozen seed order, then analyze lora_joint. Conditional branches remain blocked on that analysis receipt.

2026-07-14 — seed-7411 trigger evaluation completes; analyzer path defect found pre-result

• Complete Stage-A commit 664d6574 passed both required workflows before evaluation began. The reached-training barrier authorized the exact seed-7411 trigger invocation on source v11.
• Evaluation completed in 978.507162809372 seconds with exactly 3,072 intact and 3,072 adaptation-disabled rows. Summary identity is fd7a516e…9264; intact rows SHA-256 is 59e55adc…f6c; disabled rows SHA-256 is 65323163…745. The receipt binds checkpoint 174f3c69…cf9e, reached-training barrier 6d506285…589e, and source 5a8ed26d…6666.
• The evaluator opened only validation, depth_extrapolation, and joint_holdout; it records zero benchmark reads, no sealed-contrast payload, and no contrast-ledger event. Scientific values were not inspected.
• The independent production _load_evaluation reopen stopped before consuming result rows because _canonical_expected_path(ROOT / config["paths"]["large_artifacts_dir"]) rejects the registered ../../large_artifacts/... value as non-lexical-canonical. The CLI analyzer necessarily reaches the same path. This is a deterministic mechanics defect, not a LoRA result.
• Changing source v11 would invalidate the source-bound setup/training/evaluation graph. Preserve it unchanged. Freeze a separate, self-contained analysis-recovery experiment that consumes the exact v11 receipts and changes only the internal registered-path resolution boundary before opening any scientific values.

Current authorization

Only seed-7411 evaluation validation, documentation, commit, push, and workflow verification are authorized. After both workflows pass, freeze and adversarially review the separate analysis-recovery consumer. Remaining trigger evaluations may resume only after that consumer is mechanically ready; no result value, LoRA verdict, sealed contrast, or conditional training arm is yet authorized.

2026-07-14 — recovery published; seed-7412 trigger completes; operator visibility deviation

• The separate qwen35_4b_state_formation_analysis_recovery consumer passed 10/10 focused tests and its non-result smoke at file SHA-256 02f6f927…43f7, receipt identity 30353be5…2f8f, and recovery source contract 6ab26016…e94b. It changes only the exact registered external prefix and clean descendants before calling the unchanged v11 canonical repository/no-symlink validator.
• Recovery commit a6360cc1 and its fully enriched date follow-up e35e071e passed both repository workflows. The producer source remains exact v11; no result value was opened by the recovery smoke.
• The fixed seed-7412 LoRA joint trigger command then completed once, without retry, in 1,008.6487309932709 seconds. Summary file SHA-256 is a18ba761…f4e2, receipt identity 03eaf96b…597d, intact rows SHA-256 dd3f9424…cc4f, and disabled rows SHA-256 743f821a…c085; both modes contain exactly 3,072 rows. The receipt reopens checkpoint 90750967…0324, has exact K=1 bypass, opens only the three trigger payloads, and records zero benchmark/sealed-contrast access and no branch authorization.
• During post-run integrity validation, an intended metadata-only jq projection included the whole modes object. That object contains per-split scientific metrics, so their values were printed before seed 7413. This was an unauthorized operator-visibility deviation, not a model, data, receipt, or analyzer defect. It is documented separately and cannot be undone.
• No metric was classified, pooled, compared to the gate, or used to choose an action. Seed 7413, the analysis code, seed order, fixed finals, thresholds, and all conditional branches were frozen before exposure. To prevent adaptive stopping, seed 7413 remains compulsory and must run unchanged after this exact evaluation/deviation record is published and both workflows pass.

Current authorization

Only seed-7412 evaluation/deviation documentation, full validation, commit, push, and workflow verification are authorized. After both workflows pass, run the fixed seed-7413 LoRA joint trigger evaluation. Aggregate analysis and every conditional branch remain blocked on the complete exact three-seed evaluation matrix. The terminal report must disclose imperfect operator blinding.

2026-07-14 — seed-7413 trigger completes; exact Stage-A evaluation matrix complete

• Seed-7412 evaluation/deviation commit 27166420 passed both required workflows before the fixed seed-7413 command began. Source v11 remained exactly 5a8ed26d…6666, the seed/checkpoint/runner were unchanged, and the sole GPU was idle.
• The authorized seed-7413 trigger evaluation completed once, without retry, in 986.8842830657959 seconds. Summary file SHA-256 is 411af149…80a, receipt identity b35b9e14…da70, intact rows SHA-256 d62c43a9…b30c, and disabled rows SHA-256 683156e9…6cb; both modes contain exactly 3,072 rows.
• The receipt strictly reopens checkpoint metadata 25936f83…fbfd, checkpoint identity d65d5bb3…e68d, target-cell receipt 07f91939…362b, training barrier 6d506285…589e, exact source v11/config, and the fixed evaluation axes. It reports exact K=1 bypass, opens only validation, depth_extrapolation, and joint_holdout, and records zero benchmark or sealed-contrast access and no downstream authorization.
• The corrected post-run projection selected metadata and row-lineage fields only; no seed-7413 scientific metric was printed or inspected. The prior seed-7412 visibility deviation remains disclosed and does not change the frozen analysis or branch rule.

Current authorization

Only seed-7413 evaluation documentation, validation, commit, push, and workflow verification are authorized. After both workflows are green, run the already-published recovery consumer exactly as --phase lora_joint. Its original v11 receipt alone classifies Stage A and authorizes the next registered branch; no conditional training, sealed contrast, or source edit is authorized before it.

2026-07-14 — recovered immutable analysis mandates Stage B capacity controls

• Seed-7413 evaluation commit b326f6cd passed Validate Repository and Publish Research Site before analysis. Local main and origin/main matched, and the producer/recovery frozen contracts were unchanged.
• The recovery consumer ran exactly once as --phase lora_joint. It changed only _canonical_expected_path, delegated all scientific work to immutable producer v11 analysis.py, and emitted RECOVERED_V11_ANALYSIS_COMPLETE at sidecar identity d068482a…f40e.
• The producer output has file SHA-256 cb9fee75…818a and receipt identity b973bc01…a862. Status and verdict are LORA_JOINT_MISS_CONTROLS_REQUIRED; next stage is exactly run_lora_state_only_and_fullrank_joint.
• Zero of 57 required seed×split×depth cells passed the 0.40 terminal joint-state threshold. The maximum intact required accuracy was 0.0234375; per-seed maxima were 0.0234375, 0.015625, and 0.015625. Trained, depth-extrapolation, and joint-shift categories all miss. Adaptation contrast is ADAPTATION_CONTRAST_UNCERTAIN, with neither a consistent positive intact effect nor a disabled reversal.
• The result establishes a valid failure of the registered LoRA joint recipe, not a rank-causal conclusion. No sealed contrast was opened. The previously disclosed seed-7412 visibility deviation did not change the frozen analysis, and imperfect operator blinding remains a terminal limitation.

Current authorization

Only analysis publication, validation, commit, push, and workflow verification are authorized. After both workflows are green, Stage B must run all three LoRA state-only controls and all three direct-full-shape joint arms under the exact analysis authorization. Full-rank G0/positive-control setup precedes result training; sealed contrast remains prohibited.

2026-07-14 — first Stage-B G0 exposes downstream copy of the path defect

• Analysis commit 719c8e5a passed both workflows before the exact source-v11 full-rank seed-7411 G0 command began.
• The attempt failed at branch_authorization before its first completed check or model load. The branch consumer recomputes LoRA-miss evidence and reaches v11's same nonlexical registered external prefix. Canonical/mirror failure file SHA-256 is 47305826…2c71, identity 070c23af…aa24; both are byte-identical and inode-distinct.
• The receipt records no model load, training/evaluation, benchmark, sealed contrast, scientific evidence, or downstream authorization. Producer v11 refuses overwrite until the pair is archived.
• A separate branch-recovery experiment passed 14/14 tests and exact downstream smoke at identity d1135ea2…49b5, then archived the failure at identity ff478d40…0ec3 without removing either producer path. Its checkpoint must be published and green before verified retirement.

Current authorization

Only recovery/archive documentation, validation, commit, push, and workflow verification are authorized. After green CI, retire the exact producer failure pair through the frozen recovery, publish that retirement, and require both workflows green before retrying full-rank seed-7411 G0.

2026-07-14 — branch failure archived and exact producer paths retired

• Recovery/archive commit bdedabf4…b6b2 passed both workflows.
• The retirement transaction verified that commit contains the exact archive copy, archive receipt, producer canonical, and producer mirror. It then removed only the two producer paths while holding their exact inodes; the recovery archive remains unchanged at 47305826…2c71.
• Retirement receipt SHA-256 is 6e4c8ee3…53ad, identity c9abdc59…eae7, with zero model, training/evaluation, benchmark, or sealed access.

Current authorization

Only retirement validation, documentation, commit, push, and workflow verification are authorized. After both workflows are green, retry full-rank seed-7411 G0 through the frozen branch-recovery wrapper. Positive control and result-bearing cells remain blocked on that passing G0 checkpoint.

2026-07-13 — source-v9 seed-7411 G0 stops before model load; cache proof repaired

• Non-model setup commit ff4a8b9b passed both repository workflows before the exact seed-7411 LoRA G0 command began on an otherwise idle RTX 6000 Ada.
• G0 stopped at model_setup before model load, wrapper construction, mechanics, training, or evaluation. The general stable reader refused chat_template.jinja because Hugging Face's standard cache snapshot is a symlink into blobs/. The canonical failure and source-qualified mirror are byte-identical at file SHA-256 39ec9625…46ec7 and identity 30af333c…9cfe9; they record only train validation, zero benchmark/sealed access, and no downstream authorization.
• Source v10 keeps general repository/artifact reads no-symlink. Its dedicated cache proof requires exact snapshot basename/revision and only canonical ../../blobs/<content-id> targets, hashes an inode-distinct stable blob while revalidating the held snapshot/link, and rejects traversal, non-content targets, hardlinks, mixed roots/revisions, malformed indexes, and missing shards.
• The real pinned cache proof passes all nine files totaling 9,342,815,919 bytes at exact revision 851bf6e…cd0a, file-set identity 06486f26…d1fe12, without loading the model. The authorization context also now preserves body StableArtifactError diagnostics instead of mislabeling them as an aliased implementation review.
• Reviewed implementation a5a494b7…6f1c4a and full source contract 979a9012…f394b7 pass 360/360 tests and exact machine GO.

Current authorization

Only source-v10 validation, commit, push, and workflow verification are authorized. After both workflows pass, archive every source-v9 setup artifact with the source-qualified G0 failure as trigger, publish and validate that archive checkpoint, then regenerate from CPU smoke. No model-bearing retry or result stage is authorized before those gates.

2026-07-13 — source-v9 setup durably archived under source v10

• Source-v10 commit 3756ce29 passed both repository workflows before archival began.
• The registered transaction preserved 20 source-v9 files totaling 17,655,138 bytes. Files identity is 7360b00d…1f2650, tracked receipt file SHA-256 is 086d35af…be14e, and receipt identity is 8d5fe94d…33ad5c; it binds replacement source 979a9012…f394b7 and G0 trigger identity 30af333c…9cfe9.
• Independent verification recomputed every payload/receipt hash, matched exact archive and 20-leaf zero-quarantine membership, confirmed only structural sentinels in canonical CPU/setup directories, only .gitignore under generated data, no live initialization bundle, and the retained byte-exact source-qualified G0 failure mirror. Immediate idempotent replay changed nothing.

Current authorization

Only archive-checkpoint validation, commit, push, and workflow verification are authorized. Do not regenerate source-v10 CPU/data/initialization or run a model until both workflows are green. No result stage is authorized.

2026-07-13 — source-v10 non-model setup regenerated

• Archive commit 9c1fadde passed both Validate Repository and Publish Research Site before regeneration began.
• CPU smoke passed without loading a model at file SHA-256 ebfb68fe…17bc5, source 979a9012…f394b7, and zero benchmark reads.
• All seven deterministic splits regenerated with zero cross-split structural duplicates. Manifest SHA-256 is 0b1cca35…7422a, data-contract identity is 5fba6c3c…c252, and the sealed-access ledger remains events: [] at identity d0b9eda7…17e04. Inspection read only receipt, manifest, and compressed-byte metadata; no sealed row was decompressed.
• All three shared initialization bundles reopened through the canonical loader, have byte-identical but inode-distinct external sidecars and tracked mirrors, and exactly reproduce their archived source-v9 tensor-value digests. Bundle SHA-256 / sidecar receipt identity: seed 7411 e202efb8…3e0d8 / 9f6923d4…147abd; seed 7412 ab0b70c1…9169b / cf90d157…806e9; seed 7413 7dca25ea…a7c5c / 791aa7ec…6201.

Current authorization

Only non-model setup validation, commit, push, and workflow verification are authorized. LoRA G0, positive controls, result training, evaluation, and analysis remain blocked until both workflows are green.

2026-07-13 — first Stage-A launch stops PREPARED; source-v11 repairs publication

• Final source-v10 setup commit 555cf0b5 passed both required workflows before the exact seed-7411 Stage-A command began.
• The launch stopped before model load, train-payload access, optimizer construction, or training. Its durable journal reached only PREPARED; the canonical external directory existed but had no entries, and the tracked result directory was absent. This is not scientific or LoRA evidence.
• Root cause: atomic_write_json passed both a repository-relative full destination and its relative parent to the no-clobber publisher. Descriptor traversal therefore looked for the parent path a second time beneath itself and raised FileNotFoundError; the wrapper mislabeled it as an overwrite refusal.
• The source-v10 journal is preserved byte-exact at SHA-256 209d4c1f9183359535a07fe5bef3f535ebc13e2f00d4ddcf10c289501b1912df, journal identity ef72ba4e…f51c9, and attempt authorization d5c14f38…254ff. Failure receipt identity is 6b23f95d…b8c2; it records zero benchmark/sealed/result-payload access and no model or training.
• Source v11 normalizes no-clobber JSON destinations, publishes a leaf relative to its held parent, preserves the stable-I/O cause, and rebinds validated training/checkpoint/evaluation outputs to canonical absolute paths. It recovers only an exactly empty PREPARED output directory and rejects any nonempty markerless collision.
• Three new regressions reproduce the relative-path bug, exact mkdir-only crash, and unsafe nonempty state. The full suite passes 363/363. Reviewed implementation is 7d6cd93f…d278; full source contract is 5a8ed26d…6666. No model, GPU, benchmark, result payload, or sealed split was used in the repair.

Current authorization

Only source-v11/failure-evidence validation, commit, push, and workflow verification are authorized. After both workflows are green, archive all source-v10 setup with the preserved launch failure as trigger, publish that checkpoint, retire the stale PREPARED paths, and replay setup from zero under v11. No result-bearing stage is authorized.

2026-07-13 — source-v10 setup durably archived under source v11

• Source-v11 repair/catalog commit 94ec3780 passed both required workflows before archival began.
• The registered transaction preserved 25 source-v10 files totaling 19,566,021 bytes: all generated data and the empty sealed-access ledger, CPU receipt, all three initialization bundle/sidecar pairs, all three tracked initialization mirrors, and all three LoRA G0/control pairs.
• Files identity is f767eb80…d4b91; tracked receipt file SHA-256 is ab583fc1…e217; receipt identity is 252be000…5d6a3. It binds replacement source 5a8ed26d…6666 and launch-failure trigger identity 6b23f95d…b8c2.
• Independent postconditions leave only .gitignore under generated data and the two structural .gitkeep sentinels under CPU/setup. An immediate idempotent rerun reproduced the same 25-file receipt without mutation. No model, GPU, benchmark, result payload, or sealed split was opened.

Current authorization

Only archive-checkpoint validation, commit, push, and workflow verification are authorized. After both workflows are green, retire the canonical empty Stage-A output and stale PREPARED journal only after revalidating their published failure/journal copies, then regenerate source-v11 setup. No model or result-bearing stage is authorized meanwhile.

2026-07-13 — stale PREPARED attempt names retired after durable publication

• Archive checkpoint 24733d34 passed both required workflows before retirement.
• The canonical journal still matched its published source-v10 copy byte-for-byte at SHA-256 209d4c1f…12df, with one event whose head was exactly PREPARED. The external result directory still had zero entries, and the tracked result directory was absent.
• The source-v10 failure receipt and source-v10 setup archive receipt were rehashed to identities 6b23f95d…b8c2 and 252be000…5d6a3. Only then were the stale canonical journal and empty external directory retired. The published journal copy remains immutable under tracked failures.
• The retirement receipt records both retired paths, their preconditions, the green archive commit, zero benchmark/sealed/result/model/training access, and no downstream authorization.

Current authorization

Only retirement-checkpoint validation, commit, push, and workflow verification are authorized. After both workflows are green, regenerate source-v11 CPU/data/empty-ledger/initialization setup and publish that checkpoint before any model-bearing gate. No result-bearing stage is authorized.

2026-07-13 — source-v11 non-model setup regenerated and strictly reopened

• Retirement commit aa85086f passed both required workflows before regeneration began.
• CPU smoke passed without loading a model at file SHA-256 d46b32bd…0192. The deterministic data manifest has file SHA-256 d104a9c0…a22c, data-contract identity 43363814…6669, zero structural duplicates, and zero benchmark reads. The contrast ledger remains empty at receipt identity 01b2724b…41f3f.
• The seven compressed payload hashes and row counts reproduce the frozen procedural substrate, including 12,000 train rows, 1,024 validation rows, and the three still-sealed contrast payloads. Validation inspected only manifest/compressed-byte metadata; no sealed row was decompressed.
• All three initialization bundles reopened through the canonical loader. Each external sidecar is byte-identical to but inode-distinct from its tracked mirror, and each 15-tensor value digest exactly reproduces source v10: seed 7411 5338572e…af31, seed 7412 30aac3e3…e14, seed 7413 04054449…efeb. Bundle SHA-256 / sidecar identity: seed 7411 a03486a1…c77 / fc997954…c6b; seed 7412 3ae97463…021f5 / 688825b1…b26b; seed 7413 1167ebeb…8b2f2 / f7b95a1c…45ddc.
• Regeneration used source contract 5a8ed26d…6666, loaded no model, read no benchmark, and created no result evidence.

Current authorization

Only source-v11 non-model setup validation, commit, push, and workflow verification are authorized. After both workflows are green, replay seed-7411 LoRA G0 and its positive control. Later setup and all result-bearing stages remain blocked by their frozen ordered barriers.

2026-07-13 — source-v11 seed-7411 LoRA setup pair passes

• Non-model setup commit bcd57181 passed both required workflows before seed 7411 began.
• LoRA G0 passed at file SHA-256 9292794f…1264d, receipt identity da06c198…94d0f, with exact lineage to initialization bundle a03486a1…c77, manifest d104a9c0…a22c, and source 5a8ed26d…6666.
• Both PEFT parity regimes, K=1 before/after, zero function, and destructive checkpoint restoration have zero observed error. All 124 LoRA tensors and every required common-state group have finite nonzero live-joint gradients, the base has none, the aggregation-scalar norm is 6.167606625240296e-5, and the K=12 path is finite. The ten-step probe measured 0.2844161748886108 seconds/step with 11.197 GiB peak allocation and 35.629 GiB free after G0.
• The seed-matched positive control passed at file SHA-256 ee1de715…c98d8, receipt identity c830e65b…0c259: oracle accuracy 1.0; fixed-final intact 48/48; disabled 0/48; exactly 256 updates, accumulation 16, and 4,096 presentations; parameter values changed.
• Both receipts record zero benchmark and sealed-contrast access and remain setup evidence rather than scientific results. The complete setup barrier still lacks seeds 7412 and 7413.

Current authorization

Only seed-7411 setup-pair validation, commit, push, and workflow verification are authorized. Seed 7412 G0/control may begin only after both workflows are green. Result training, evaluation, and analysis remain blocked on the complete three-seed setup barrier.

2026-07-14 — recovered full-rank G0 passes; handoff repair frozen

• The branch-recovery retirement checkpoint e69c1960…c79 passed both workflows before retry.
• Full-rank seed-7411 G0 passed at file SHA-256 cdc90cd…c68f, identity e1f1c906…f89dc: K=1 parity before/after optimizer and checkpoint restoration have zero error; all 62 direct-full-shape delta tensors have finite nonzero live-joint gradients and complete finite optimizer state; K=12 is finite; and about 22.1 GiB remains free.
• The G0 authorizes only its positive control and records zero training/evaluation, benchmark, contrast, or scientific evidence. First-recovery STARTED/COMPLETE file hashes are f0c78055…3c87 / 755ad561…f941.
• The positive-control handoff stopped before wrapper STARTED publication because the frozen first recovery tests the former failure pathname rather than its exact bytes/status. No producer or model work occurred.
• The additive byte/status-aware handoff suite passes 8/8. Its no-model smoke passes all six controls at file SHA-256 0e0409de…6e5b, identity d6fecf0a…de4c, under source contract 4d2ffde3…8acd.

2026-07-14 — full-rank seed-7411 setup pair passes

• Handoff commit 627254f1…d892 passed both workflows before the positive control.
• Producer control file SHA-256 is a0d17e2e…7a16, identity 6708a4d4…9649; handoff STARTED/ COMPLETE file hashes are b1b91cb2…ff2b / e0887536…13f1, with COMPLETE identity af0dd15c…2262.
• Oracle and intact fixed-final joint state are 48/48; adaptation-disabled joint state is 0/48 after exactly 256 optimizer updates, accumulation 16, and 4,096 presentations. The 892,272,640 full- rank parameters and shared-state parameters both changed; early stop and checkpoint selection are false.
• G0/control are setup-only evidence and record zero result-payload, benchmark, contrast, or scientific-evidence access.

Current authorization

Only seed-7411 full-rank setup documentation, validation, commit, push, and workflow verification are authorized. After both workflows are green, run the seed-7412 full-rank G0 and positive control through the handoff wrapper. Result training remains blocked on the complete setup matrix.

2026-07-14 — full-rank seed-7412 setup pair passes

• Seed-7411 setup commit c507488c…3c06 passed both workflows before seed 7412 began.
• G0 file SHA-256 / identity are 10bf22fc…1d18 / 62ecb79e…951dc; handoff COMPLETE identity is d61d6441…9246. K=1 and checkpoint-roundtrip error are zero, all 62 deltas have complete finite optimizer state, joint gradients are finite/nonzero, K=12 is finite, and 22.1 GiB remains free.
• Control file SHA-256 / identity are 1cbbd823…8510 / 6575e1d2…4554; handoff COMPLETE identity is 6334d72f…c14a. Oracle/intact are 48/48 and disabled is 0/48 after exact 256 updates, accumulation 16, and 4,096 presentations; full-rank/shared-state parameters changed.
• Both receipts remain setup-only evidence with zero result-payload, benchmark, contrast, or scientific-evidence access.

Current authorization

Only seed-7412 setup documentation, validation, commit, push, and workflow verification are authorized. After both workflows are green, run the seed-7413 full-rank G0 and positive control. Result training remains blocked on the complete setup matrix.

2026-07-14 — full-rank seed-7413 pair passes; setup matrix complete

• Seed-7412 setup commit 92473a53…8530 passed both workflows before seed 7413 began.
• G0 file SHA-256 / identity are 021a8444…d635 / 4d2316d3…2ff0; handoff COMPLETE identity is fbfb282f…5aef. K=1/round-trip error are zero, all 62 deltas have complete finite optimizer state, joint gradients are finite/nonzero, K=12 is finite, and 22.1 GiB remains free.
• Control file SHA-256 / identity are 8a4af0d6…fde8 / 976f28ef…14df; handoff COMPLETE identity is ccff1520…58c9. Oracle/intact are 48/48 and disabled is 0/48 after exact 256 updates, accumulation 16, and 4,096 presentations; full-rank/shared-state parameters changed.
• All three full-rank G0/control pairs now pass and remain setup-only, with zero result-payload, benchmark, contrast, or scientific-evidence access.

Current authorization

Only complete setup-matrix documentation, validation, commit, push, and workflow verification are authorized. After both workflows are green, run the mandatory three LoRA state-only and three full- rank joint Stage-B training cells under their exact producer barriers. Evaluation remains blocked on the complete Stage-B training matrix.

2026-07-14 — source-v11 seed-7412 LoRA setup pair passes

• Seed-7411 setup commit 5536f785 passed both required workflows before seed 7412 began. The next remote commit touched only another experiment and generated indexes; source contract 5a8ed26d…6666 was reverified after fast-forward.
• Seed-7412 G0 passed at file SHA-256 c19a6944…6b82d0, receipt identity e2229ad2…da380, with exact lineage to initialization bundle 3ae97463…021f5 and the unchanged manifest/source.
• Both PEFT parity regimes, K=1 before/after, zero function, and destructive checkpoint restoration have zero observed error. All 124 LoRA tensors and every required common-state group have finite nonzero live-joint gradients, the base has none, the aggregation-scalar norm is 4.465342499315739e-5, and the K=12 path is finite. The ten-step probe measured 0.2639180898666382 seconds/step with 11.197 GiB peak allocation and 35.629 GiB free after G0.
• The seed-matched positive control passed at file SHA-256 1a8d263b…a27a37, receipt identity fdb1a16e…cd72a3: oracle accuracy 1.0; fixed-final intact 48/48; disabled 0/48; exactly 256 updates, accumulation 16, and 4,096 presentations; parameter values changed.
• Both receipts record zero benchmark and sealed-contrast access and remain setup evidence rather than scientific results. The complete setup barrier still lacks seed 7413.

Current authorization

Only seed-7412 setup-pair validation, commit, push, and workflow verification are authorized. Seed 7413 G0/control may begin only after both workflows are green. Result training, evaluation, and analysis remain blocked on the complete three-seed setup barrier.

2026-07-14 — source-v11 seed-7413 LoRA setup pair passes; matrix complete

• Seed-7412 setup commit 47e160bf passed both required workflows before seed 7413 began.
• Seed-7413 G0 passed at file SHA-256 87faadfa…aca4d0, receipt identity 1fa2b054…eec8d5, with exact lineage to initialization bundle 1167ebeb…8b2f2 and the unchanged manifest/source.
• Both PEFT parity regimes, K=1 before/after, zero function, and destructive checkpoint restoration have zero observed error. All 124 LoRA tensors and every required common-state group have finite nonzero live-joint gradients, the base has none, the aggregation-scalar norm is 1.5512113459408283e-3, and the K=12 path is finite. The ten-step probe measured 0.281365180015564 seconds/step with 11.197 GiB peak allocation and 35.629 GiB free after G0.
• The seed-matched positive control passed at file SHA-256 e2e70edf…fc5b57, receipt identity 8d74b90d…d689b7c: oracle accuracy 1.0; fixed-final intact 48/48; disabled 0/48; exactly 256 updates, accumulation 16, and 4,096 presentations; parameter values changed.
• All three source-v11 LoRA G0/control pairs now pass. Every setup receipt records zero benchmark and sealed-contrast access and remains setup evidence rather than a scientific result.

Current authorization

Only final setup-matrix validation, commit, push, and workflow verification are authorized. After both workflows pass, run the three Stage-A LoRA joint cells in frozen seed order. Evaluation, analysis, and all conditional branches remain blocked until their exact upstream barriers authorize them.