Final v3 plan: frozen one-shot pose robustness

Original decision: 2026-09-29. Execution amendments: 2026-10-09, 2026-10-12 and 2026-10-19. Delivery: 2026-12-18. Status: final planning decision; implementation and measured campaign gates remain to be completed. This document replaces the v3 draft. Decision 0004 records the change of direction from v2. It does not constitute a final-test lock, a completed experiment, or a change to the deployed tracker.

The October 19 researcher-authorized ADR0008 governs the next Week 5 execution. Freeze one prospective runtime baseline; require exact repeats and initialization/batch pairing within that profile, without requiring runtime-derived scales, tensors or initialization to equal historical CPU outputs. Original panel IDs, fallback, manifests, algorithms, scientific settings and all other checks remain fixed. Bind the immutable amendment/runtime record and fresh scientific inventory to the unchanged r2 protocol/design/source hashes, preserving all locks and artifacts. Accepted setup/rehearsal remains valid where its bound inputs are unchanged; refresh runtime-dependent and dynamic checks. Only a bound core change requires a new revision through the existing correction path. Allocations may resume after recording the amendment and passing required preflights; selection, final training and novel opening remain gated. The two non-progress diagnostic attempts remain historical and are not reset. Earlier runtime recovery instructions below describe their dated checkpoints; ADR0008 controls this specific prospective change.

October 22 accepted milestone: independent review accepted job 8189505's prospective CPU runtime, three fresh initializations per seed 7/17/29, and exact two-process geometry for the original 43 panel IDs and 430 tensors per repeat. Original identities, fallback, scientific settings and r2 source/locks remain fixed; allocated replay, accounting, resources and seals passed. Receipt study-v3/operations/week5-runtime-baseline-acceptance/20261022T154102Z/verification.json has SHA-256 16106ecafe347dc8d31795e14a88b267e716f82e494311b4753462cfcd88f8aa. Week 5 remains incomplete. Next is the fresh source-bound GPU diagnostic under this baseline, then fresh parity/five caches and independent validation. Three complete paired pilots, nine corruption paths/projections/results, full admission, independent recertification and measured schema-2 capacity remain required. The two historical non-progress diagnostics stay recorded; selection, final training and novel opening remain closed. See the current execution checkpoint.

The October 9 researcher-authorized amendment adopts the specifications below before selection or novel opening. Scientific settings remain fixed; completing currently disabled stages requires a successor execution release and new provenance bindings. Preserve the September release, Week 4 closure and all original evidence. Sections 5–10 govern the remaining work; they do not assert that the new implementation, capacity checks or deployment have happened.

The October 12 researcher-authorized ADR0007 permits command-local reuse of successful numerical validation with complete content rehashing, and separate CPU/GPU capacity accounting. It requires a fresh execution-r2 release and regeneration. The independently reviewed implementation passed the stable full local suite: 1,373 tests with two historical AVSL skips. Source release f101f0e9d9bd64275d2088deab1b0bdcfcc53620 is frozen and pushed. The separate SCC installation and allocated CPU rehearsal 8044010 passed independent completion review: 128 checks, no failures or skips, successful 194-second accounting, nine resource samples and intact seals. Verification receipt week5-r2-rehearsal-verification/20261013T002413Z/verification.json has SHA-256 963babe02eba70d34f3f9c9aa47af21145de387ca4d5a1ad6ccf9c226fb43381. This fixture-only evidence does not admit scientific execution. Preparation 8045152 also passed independent completion review: successful 44-second accounting, three commands, seven resource samples, both seals, all eight byte-identical manifest identities and original → r1 → r2 scientific equality. Receipt week5-r2-preparation-verification/20261013T063819Z/verification.json has SHA-256 0a3823e66a97192c5665185edbb908bf5b44c6dd9b451328d95d96476a84d785. The new protocol SHA-256 is 18f7d17fa81e7ae3a6b1e5deddf42d9eda589d96e5a13968188a8ce275bfb95f. Supplementary CPU checks 8048540 completed October 13 at 08:31:43.006302 UTC on apolkovnikov-pub@scc-h30: 86 checks passed without skips, successful 87-second accounting, three commands, seven resource records and both seals. Independent review completed at 20:43:00.762966 UTC, verifying 108 metadata bindings, 616 source bindings and actual locked-ID dependence tables for 18,929 development queries and 18,884 primary novel IDs, each over 20 actions. Novel poses and outcomes remained unopened. Receipt week5-r2-additional-verification/20261013T204221Z/verification.json has SHA-256 7703859e50361123de9054fba446c8384d423b8c549b69eab92720314c46c3d2. Geometry job 8070391 was submitted at 20:46:55.956765 UTC and initially observed queued with four CPU slots at 20:49:22.424937 UTC. Its launch receipt week5-r2-launches/20261013T204651Z/launch.json retains SHA-256 a8c55f5a6054996d49cb2abab304a8b8925914a4cf0495a2f0368f84e4177452. The due inspection at 22:05:07.547370–22:05:19.824007 UTC found terminal failure: bpnlab-pub@scc-kc1, four OMP slots, 20:51:02–20:54:51 UTC, 229 scheduler seconds, accounting failed=0/exit_status=1. The tracked checker and exact within-revision repeats passed, but the external wrapper rejected fixed cross-revision identity. That attempt produced no accepted successor geometry completion. Metadata diagnosis found 2,051 candidate own_scale scalars differed; the fixed 43 panel IDs, selection and manifests remained unchanged. All 43 clean tensor hashes matched. Four selected scales and 129 coordinate-jitter hashes differed across revisions; all 430 hashes were exact within each revision. Diagnosis week5-r2-geometry-diagnosis/20261013T220749Z/diagnosis.json has SHA-256 d9d8b549c159206754d8f8b21eaaac1e6b5f656dee54959c23611717aa70534b; discriminator week5-r2-geometry-diagnosis/20261013T221341Z/clean-tensor-discriminator.json has SHA-256 f0dc0f44ba52c114e76b95ac044dd1ed63fd14fc62c8e1b9efd57c80f886ddcb. At that earlier geometry checkpoint, CPU dispatch was an untested recovery hypothesis. Separate native/default diagnostic job 8072666 was submitted October 13 at 22:49:45.058490 UTC after 127 local checks passed (60 parent, 27 worker, 40 launcher). Initial inspection at 22:52:26.185888–22:52:38.052404 UTC observed queued (qw), four slots, with no operation, log or resource records. No numerical diagnostic result was produced. Launch week5-r2-launches/20261013T224936Z/launch.json has SHA-256 91358af8b0e0f3c04765c7fd9f5d4d5264c32cbb7dbaec2c5a2f7dfa917332a3. Its diagnostic-only 1-GiB bound comes from 67,480,406 bytes of metadata/log/source bindings; the existing 17,675,344-byte archive is read only and never copied. Fresh quota/filesystem reserve passed; this does not establish campaign capacity. The October 13 at 8:00 PM EDT check was not executed after SSH disconnected; no result was inferred during the gap. After access was restored, inspection on October 14 at 19:40:13.233256–19:40:25.519918 UTC found diagnostic 8072666 failed before its parent started: neuro-pub@scc-h32, four OMP slots, October 13 at 22:54:00–22:54:02 UTC, two scheduler seconds, accounting failed=0/exit_status=1. SGE copied the wrapper to its spool directory; the helper lookup used that directory and failed before either worker ran. Independent metadata-only review at 19:56:38.194670 UTC verified zero operations, resource records or worker outputs and 139 artifact/617 external/616 source bindings. No CPU capability comparison ran in job 8072666; the geometry cause remains unproven. Failure receipt week5-r2-geometry-runtime-failure-verification/20261014T195626Z/verification.json has SHA-256 c9e2d25740334923a537c36c82b0d2574f6b90a74b7d45ebf26ddc0c3aac371d. A separate reviewed operational remedy derives the unchanged parent helper from the canonical bound input packet directory. Its 64 focused checks and review passed; the original parent/worker and scientific settings/tolerances are unchanged. New immutable tools study-v3/operations/week5-r2-geometry-runtime-diagnostic-remedy-tools/20261014T200830Z were staged and verified October 14 at 20:08:33.599282 UTC. Tools manifest SHA-256 c6404952ea81aa28b04c2603b6b4b7190cc85c1c759ff08410ee2dd7f63c1d77; input packet SHA-256 32289c531a0330f93dad42b7974022181eb494d775d6b70825a80a7a1877d716; staging verification SHA-256 ae31768ad55b7107b113c581aa6aff0e8690f7e1b86ecdd575a40bd72dd87316. At the staged checkpoint on October 14 at 20:08:33.599282 UTC, the launcher had not been executed. That checkpoint and the earlier 4:41 PM EDT launch/check boundary remain historical. The new guarded launcher executed once: diagnostic remedy job 8102231 was submitted October 14 at 20:43:11.969137 UTC. The fresh pre-submit scheduler check completed by 20:43:10.988132 UTC with no active jobs; it precedes submission and does not establish the new job state. Launch week5-r2-launches/20261014T204308Z/launch.json has SHA-256 82427510eb77a8bc0875753d343783a1e9a383f05bef736aa2c01654ee4037f2; submission intent SHA-256 7c690ce8410753e592a06c555dfb87c4b638ca28e344552a1d246c8c2fdeee6d. Independent submission-metadata review at 20:50:34.444145 UTC rehashed 763 metadata and 616 frozen-source bindings, checked the packet/intent/command and both launch seals. Receipt week5-r2-geometry-runtime-remedy-submission-verification/20261014T205034Z/verification.json has SHA-256 63aa23c6e514f9cd1757021354fa1262b8f48f89af10e8a8015259e885dbb471. Fresh quota 826,630,000,000 bytes and filesystem free 994,956,017,664 bytes passed the diagnostic-only reserve. Its 1-GiB bound derives from 67,481,168 bytes of metadata/log/source/receipt allowances; the existing 17,675,344-byte archive is read only in allocation, with no copy. This does not establish campaign capacity. The request remains four OMP CPU slots, 32 GiB, 30 minutes, no GPU and the reviewed Gold-6426Y/AVX2 selector. At the submitted-only checkpoint, no post-submission state or result had been observed; the earlier 5:45 PM EDT (21:45 UTC) first-check boundary is historical. Actual inspection on October 14 at 22:25:31–22:25:43 UTC found completed diagnostic remedy job 8102231, completed at 20:49:21.682189 UTC. Accounting records neuro-pub@scc-ke1.scc.bu.edu, four OMP slots, failed=0/exit_status=0, 17 whole-job CPU wall seconds and 9.340 actual CPU seconds. Independent completion review at 22:34:10.047781–22:34:14.167272 UTC verified both seals, two unchanged worker commands, dispatch evidence, the final wrapper marker, six resource records and 184 artifact/617 external/616 frozen-source bindings. Receipt week5-r2-geometry-runtime-remedy-queue-record-verification/20261014T223409Z/verification.json has SHA-256 f077737e6781d002d92458d1a743d83c772f7ef028676ec987ad54455ce34d36. The four-case recovery hypothesis is unsupported: native AVX512 exactly matched the failed-fresh four scales, 40 tensor hashes and 36 damage rows. DEFAULT recovered the original four scales but only 36/40 tensor hashes and 32/36 damage rows. Clean and all masking records matched the original; four residual jitter hash/damage-row differences across two cases remain. This does not establish the complete geometry cause or a permanent CPU profile. The first completion verifier rejected only the allocation QUEUE representation and produced no accepted receipt; its source and failure remain preserved. The reviewed replacement accepts the scheduler's bare queue name with a separately verified host, or its exact qualified forms. Host, job, PE, slot, thread, scientific and numerical checks remain unchanged; 93 focused tests, Ruff/format and independent review passed. Geometry and every dependent stage, full allocated A/corruption replay, scientific admission and measured CPU/GPU capacity remain held. At the October 14 completion checkpoint, no further profile, helper or job had been claimed; the earlier 7:27 PM EDT scheduler/preflight boundary is historical. Separate libm diagnostic tools were staged and verified at 23:26:33.802337 UTC after 173 local checks and independent review. Tools manifest SHA-256 90c9336085f01d6b423a04a6087aff381b1527952cc86bea57e67c3cefd43093; input SHA-256 1d6ea033970b2b1b49ca6548180e5e6e5f8c097cb722a2466db49b4809d4a4ff. The new guarded launcher executed once: libm diagnostic job 8104416 was submitted October 14 at 23:29:21.658353 UTC. Launch week5-r2-geometry-libm-launches/20261014T232917Z/launch.json has SHA-256 03fa95ee91fb6c6686b792e111d3033bef8af8e5a12cca740be813023c87caf5; intent SHA-256 574b8f072e3364b6650f01cead24e0bf60fae09eae8cda94bf574e938d3760b9. Independent submission review at 23:34:29.161902 UTC rehashed 815 metadata and 616 frozen-source bindings, checked exact intent/source/tools/input/command and complete before/after launch seals. Receipt week5-r2-geometry-libm-submission-verification/20261014T233429Z/verification.json has SHA-256 0b6a6e05b49bd698cc4d0d317d5142d319d3ec6a5677448e0b90803e19938b33. Fresh quota 824,700,000,000 bytes and filesystem free 992,888,225,792 bytes passed the diagnostic-only reserve: 67,667,170 bytes rounded to 1 GiB. The existing 17,675,344-byte archive is read only in allocation, with no archive copy or runtime ELF output copies. This does not establish campaign capacity. The request is four OMP CPU slots, 32 GiB, 30 minutes, no GPU and the reviewed Gold-6426Y/AVX2 selector. The four-case diagnostic pairs fresh DEFAULT control and DEFAULT plus a version-appropriate libc/libm FMA-disable request; actual runtime, ELF and selected function implementation must be proven in allocation before interpreting numerical metadata. No permanent runtime profile or complete geometry cause is established. The accepted 8102231 negative result and all earlier failures remain preserved; geometry and every dependent scientific, replay and capacity gate remain held. The pre-submit scheduler check ran at 23:29:20.578046–23:29:20.793929 UTC; it precedes submission and does not establish the new job state. At the submitted-only checkpoint no post-submission state or result had been observed; the October 15 at 00:31 UTC first-check boundary is historical. The first observation ran once on October 15 at 00:32:02.955881–00:32:16.614103 UTC. Diagnostic 8104416 completed October 14 at 23:34:33 UTC on apolkovnikov-pub@scc-h30.scc.bu.edu: four OMP slots, failed=0/exit_status=0, 21 whole-job CPU wall seconds and 14.493 actual CPU seconds, peak RSS 695,096 KiB. Independent terminal review at October 15, 02:34:06.243155 UTC passed its diagnostic-only accounting/command/seal/final-marker checks, six resources and 240 artifact metadata/617 external metadata/616 frozen-source bindings. Receipt week5-r2-geometry-libm-terminal-verification/20261015T023357Z/verification.json has SHA-256 9f45a91a0ba63431f22bad35b6402fde9faabcde0d8b53682b4c448f9b6b3277; retained compact receipt SHA-256 dc5d3db0011bea0c1f3cae4d814e5d0ec2919588a441afda45d61ccdc2684fd6. Classification is inconclusive-dispatch: both DEFAULT control and FMA-off matched the original four scales, 36/40 tensor hashes and 32/36 damage rows. Neither required-symbol dispatch nor default math code offsets changed; the intervention did not establish the intended implementation difference. This supports neither acceptance nor rejection of the libm cause hypothesis and selects no permanent profile. The earlier 8102231 unsupported recovery result and all failed attempts remain preserved. The next hourly boundary after observation was October 15 at 01:32:16.614103 UTC; no repeat scheduler query is needed for terminal review. At the later access-blocked checkpoint, SCC access was found disconnected after an attempted full-receipt export returned no receipt. Only accepted compact/capture evidence was then retained, with no full local JSON export or refreshed SCC status claimed; restoration was pending and no new job was claimed. That checkpoint remains historical. The researcher restored the existing SCC connection on October 15. The full 238,862-byte verification JSON was recovered at 15:03:07.054586 UTC from seven checked metadata chunks, retaining its exact 9f45a91a0ba63431f22bad35b6402fde9faabcde0d8b53682b4c448f9b6b3277 identity. Four bound JSONs (child inputs, control runtime, FMA-off runtime and summary) were recovered and hash-checked against that receipt by 15:05:07.897329 UTC. Independent read-only review confirmed exact transport, the original comparisons and 240 metadata/617 external/616 source counts. SCC status was refreshed at 15:13:30.934012 UTC, SHA-256 bbf4d01268ac8db236ceba3424aa2167a741caba72c8b930166d57ca0fbd1230. Recovery read only metadata: no scheduler query, scientific archive or runtime binary was read. Last actual scheduler batch remains October 15 at 00:32:05.186540–00:32:16.493406 UTC. The earlier 5644 recovery command stopped at its export-size guard and remains preserved without rerun; the earlier ac786 export remains ambiguous and preserved.

Recovered runtime metadata records glibc 2.28 stable, baseline GLIBC_TUNABLES unset and candidate glibc.tune.hwcaps=-FMA_Usable, with identical required/default versioned math offsets before/after both conditions. Classification remains inconclusive-dispatch; recovery establishes neither a libm cause nor a permanent profile. One data-free resolver sensitivity design is selected for local implementation and independent review only: two serial fresh children, GLIBC_TUNABLES unset versus exact glibc.cpu.hwcaps=-FMA_Usable, changing only the namespace from the prior request. It reads no NPZ, runs no historical worker or numerical experiment, and imports no Torch in the children; the supervisor imports Torch to retain inherited safeguards. Actual package/loader provenance and required-version offset sensitivity must be established in allocation. Producer tools passed 140 local checks, Ruff/format/Bash syntax and independent review. Separate metadata-only submission, observation and terminal controls passed 75 synthetic checks, Ruff/format and independent review. Both exact source sets are reviewed and frozen locally; nothing is staged or submitted. No new job, resolver result, permanent runtime or scientific acceptance is claimed. Geometry, all dependent stages/full allocated replay, scientific admission, measured CPU/GPU capacity, Week 6 selection, final training and novel opening remain held. Source, recipes, tolerances, IDs and fallback remain frozen. Never rerun an executed launcher or observer. Its output can only be diagnostic metadata; it cannot produce eligible scientific artifacts or pass a gate. Source, scientific settings and tolerances remain frozen; there is no blind geometry retry. Fresh successor scientific evidence, full admission and measured CPU/GPU capacity remain pending. Preserve the measured no-go and all execution-r1 evidence.

Following the final prospectus, the December study asks whether a full contextual training recipe reduces one-shot retrieval degradation relative to a standard contrastive recipe when only novel pose queries are corrupted. Use one dataset, a frozen generic MotionBERT encoder, three required objectives, three paired seeds, a fixed 20-epoch budget, and nine synthetic corruption conditions. A null or negative result completes the scientific objective.

1. Final decisions

Fixing 20 epochs removes the draft's adaptive plateau rule, including its failure to distinguish a plateau from a decline. The pilots measure the whole declared duration. Their learning curves are reported as fixed-budget diagnostics; this plan does not claim that 20 epochs establishes convergence. If the 20-epoch campaign does not fit, stop for a documented, result-blind amendment before selection. There is no automatic reduction to ten epochs.

The regularizer-only control is omitted to bound implementation, storage, and researcher work. Its omission is a scope/capacity decision, not a negative experimental result. Do not add it after viewing scores. The 26-method v2 benchmark, encoder fine-tuning, Multi-Similarity, extra datasets, alternative batch sizes, and real pose-estimator shifts are future studies. Preserve their existing code and evidence; do not describe this study as completing v2.

The retained transfer recipe follows the base paper's cross-dataset appendix rather than a new pose-specific search. Physical batches preserve its within-batch neighborhood operation. Frozen-backbone transfer also has precedent in MotionBERT, Section 4.6/Table 4. Neither paper establishes optimal pose epochs, seed count or a useful-effect threshold. Keep the declared budget and report uncertainty instead of inventing such guarantees. Complete candidate and cost reporting follows Dodge et al..

2. Evidence and integration base

The local checkout inspected for this plan is c60c0fe with existing uncommitted Week 3 and v3 work. Remote main, independently checked on September 29, is 47ca528a99048278693e28019f52d8e924d7e270. Implement v3 in a clean worktree based on that commit or a reviewed successor. Do not reset, merge over, or copy the dirty local tree wholesale. Reconcile useful local changes against already committed Week 3 work individually.

The committed Week 3 execution record and compact evidence establish two byte-identical 95,001-row clean frozen extractions on an L40S: 1,758.60 and 1,765.83 seconds, 3,315,915,790 bytes each. Development caches contain 76,013 training rows, 20 gallery rows, and 18,929 primary query rows. This establishes the clean extraction path, not head-training or corruption runtime. The v2 fine-tuning forecast is not a v3 runtime estimate.

A fresh, read-only inspection through the authorized remote tmux session observed the following; no GPU job or dataset inference was launched:

These are time-specific observations. Before pilots, selection, and opening, repeat the audit against a registered list of every historical local/remote artifact root and release. An unreadable or unresolved root blocks the seal gate. Absence in the inspected locations is not proof about an unknown copy.

Seal terminology: the current aggregate loader deserializes and validates all annotations in the shared pickle before selecting auxiliary examples. Historical evidence therefore supports no novel preprocessing for retrieval, embedding, visualization, scoring, or tuning access; it cannot establish that novel pose bytes were never deserialized. Record this packaging fact in the v3 decision. No novel poses or outcomes were read by this planning inspection. For the new final extraction path, authorization and atomic opening must occur before the call that deserializes novel inputs. Auxiliary container preparation must be documented separately and must expose only authorized auxiliary IDs to preprocessing, diagnostics, or model code. If the historical audit discovers actual novel representations, scores, or outcome-informed decisions, halt: a fresh namespace cannot restore a sealed confirmatory test.

October 8 supersedes the initial feasibility uncertainty with verified clean caches and three valid pilots on release 2897f4fe4c1fb1857432a0e92a7391dde68307f5. The dated Week 4 record reports successful independent validation and 1,008.48 decimal GB of quota free at one inspection. These observations are not future reservations. Historical root retirement was resolved through dated evidence and researcher confirmation; retain that resolution instead of recreating empty roots. Audit active roots and the retirement receipts at later gates. Original release evidence does not authorize a successor release automatically.

3. Dataset and fixed episode

Retain the adopted source contract: 113,945 usable HRNet annotations and 535 disjoint official missing-skeleton exclusions reconcile to 114,480 nominal captures. Bind the exact aggregate, missing list, source inventory, pinned NTU protocol, checkpoint, and upstream hashes. The original adopted input protocol hash is c8b08b6867bc14dc0a947ebfa94e40b16ea3f0dae38dfa6d86187afecb4fd45f. It remains historical provenance, not authorization to execute v3.

• Novel actions: 1, 7, 13, 19, 25, 31, 37, 43, 49, 55, 61, 67, 73, 79, 85, 91, 97, 103, 109, 115; retain all 20 official exemplars exactly.
• Development-validation actions: 2, 8, 14, 20, 26, 32, 38, 44, 50, 56, 62, 68, 74, 80, 86, 92, 98, 104, 110, 116.
• Selection training uses the other 80 auxiliary actions. Final training uses all 100 auxiliary actions. There is no training-subset contingency.
• Preserve the validated development anchor/query identities. Its original anchor rule minimizes SHA256("protocol-v1|dev-anchor-v1|" + sample_id) within each action. Do not change that prefix to obtain new v3 anchors. The historical episode hash is 88f30f96e35dedfd34b42f8e8ee2b561db44bbaf1f93d82d0522ffe011abba46. Generate v3-bound manifests with the same ordered IDs and exclusion rule; retain the original episode and hash as provenance.
• Primary queries exclude the gallery sample and all synchronized views of its underlying performance, keyed by setup, performer, repetition, action. The exact-official secondary set excludes the exemplar itself only. Derive and lock both exact lists/counts from metadata before opening.

The existing manifest audit records 18,944 usable novel rows, 18,924 exact-official queries, and 18,884 primary queries. Verify these counts and exact ordered IDs against the bound manifests; do not reconstruct the primary set by assuming three views per performance. Keep the gallery order fixed for deterministic cosine ties.

Use the adopted deterministic [2 people, 100 frames, 17 H36M joints, 3 channels] float32 tensor: camera normalization, tracking, COCO-to-H36M conversion, uniform temporal sampling, missing-person padding, then spatial normalization. Derived-joint confidence is the minimum source confidence. Random movement is off. Preserve the declared local confidence correction; do not claim upstream confidence preprocessing is byte-identical. Encoder parity uses the same local input tensor on both paths.

4. Fixed model, recipes, and training

The generic pretrained MotionBERT checkpoint remains frozen and in evaluation mode. Pool its representation over time and people, retaining 17 × 512 = 8,704 pre-head values. Train a biased Linear(8704, 2048) followed by L2 normalization: 17,827,840 trainable parameters. Dropout is zero.

These are fixed transfer recipes, not starting settings to adjust after diagnostics. The contextual coefficients and different contrastive-only positive margin follow Liao et al., Appendices G and H.5. Batch 32, dimension 2,048, and 20 epochs are pose-study adaptations. The source image results do not validate these as pose optima. SupCon's single-view temperature-0.07 implementation is also an adaptation of the original method. Neither comparison isolates the contextual term from margins and regularization.

Lock the numerical implementation before pilots:

• Context fits Equation 5's same-label target y, resolving the conflicting printed Algorithm 1. Normalize embeddings; use squared distance max(2 - 2*cosine, 0) and include self in neighbor selection.
• Retain detached kth-distance thresholds and first-stage neighbor/non-neighbor counts, the prescribed constant-alpha comparison backward, reciprocal expansion with its non-detached denominator, and symmetric output. Average the off-diagonal contextual squared error over all n² entries.
• Average positive and negative contrastive hinges separately over strictly active pairs. Empty active sets contribute zero. The mean-similarity regularizer includes the diagonal. Explicitly specify the existing denominator clamp to at least one for empty complements as a finite local convention, and test degenerate/tied neighborhoods.
• Derive independent hand-calculated labeled forward and backward fixtures from the paper. Do not use an implementation mirroring production as the sole oracle. Ordinary finite-difference checks cannot validate an intentionally heuristic discontinuous comparison gradient.
• The unlicensed contextual repository remains reference-only. Do not copy, adapt, vendor, or redistribute its code. Preserve permitted MotionBERT compatibility-port headers, attribution, and license files.

All arms use float32 features/parameters and constant-rate AdamW: betas=(0.9, 0.999), eps=1e-8, amsgrad=False, weight decay 1e-4 on weight and bias; no scheduler, mixed precision, clipping, or augmentation. Pin the PyTorch implementation/runtime and deterministic settings, including CUDA configuration before any CUDA initialization. Use PyTorch linear default initialization under each paired seed and archive its state hash.

Retain the physical P=8, K=4 sampler. An epoch is ceil(N/32) batches; epoch randomness is derived from seed plus epoch, with class/example queues shuffled at the epoch start and modular cycling within the epoch. It is a balanced sampling budget, not a promise that every sample is visited once. Persist the exact ordered batch plan. Pair it and the initial head state across arms and learning-rate candidates within a seed.

Engineering pilots and health criteria

Before fresh parity/cache generation and the first scientific pilot, freeze the complete canonical v3 protocol: recipes, numerical conventions, fixed duration/grid, exact identities, corruption/analysis semantics, and fallback value/evidence binding. Calculate that fallback from authorized auxiliary poses during preparation. Freeze the pilot specification against this same protocol hash with 20 epochs, seed 7, and learning rate 3e-4. Run one pilot for each arm on the full 76,013-row development training set. Save immutable weight snapshots and complete clean development scores at epochs 5, 10, 15, 20, plus epoch-zero diagnostics. Pilots are explicitly ineligible for learning-rate selection, even where settings match a candidate. Final selection runs are fresh attempts.

Required wiring checks use a fixed constructed 32-row, eight-label learnable fixture, paired initialization, and physical batches. In 1,000 updates at the pilot optimizer settings, each arm must remain finite and reach 100% self-excluded same-label nearest-neighbor retrieval. This checks memorization of a deliberately learnable fixture; there is no required accuracy floor on real development data.

On real pilots require finite losses, gradients, and updated parameters at every step, the exact prescribed update count, and nonzero projected vectors. At epochs 0/5/10/15/20, require normalized-vector error at most 1e-5 and mean_q ||z_q - mean(z)||² > 1e-6 on the complete development query set, summarized in float64. This screens total numerical collapse; it does not establish a useful representation or justify tuning. Zero individual gradients from inactive terms are allowed.

Archive loss-component magnitudes, active positive/negative hinge fractions, neighborhood and reciprocal counts, empty-complement counts, gradient norms, embedding variance, cosine quantiles, and actual timing/memory/checkpoint bytes. Use fixed diagnostic sampling points, including the first batch and each epoch boundary. Poor accuracy, continued learning, or overfitting curves are reportable findings, not permission to change recipes or pick an epoch. A nonfinite/collapse or mathematical failure blocks selection until corrected under the amendment policy.

Selection and final training

Run 3 arms × {1e-4, 3e-4, 1e-3} × {7,17,29} = 27 development runs at 20 epochs. Rank candidates separately within each arm by mean epoch-20 clean development top-1 across all three seeds; exact ties prefer mean MRR, then the lower learning rate. Use unrounded metric values for ranking. There is no best-epoch selection, early stopping, or corrupted-score selection.

A candidate needs three valid seeds. An unresolved numerical failure makes that candidate ineligible; a two-seed mean cannot compete. Finish the declared attempt matrix and retain all failures. If an arm has no valid candidate, stop for a result-blind amendment. Report any unequal number of eligible candidates; all arms retain the same attempted selection budget.

Lock the three selected configurations, selection evidence, and timestamp. Train nine fresh heads on all 95,001 auxiliary rows, at 20 epochs, with the selected rates and the three paired seeds. Retain the epoch-20 checkpoint; do not choose a checkpoint using novel data. Do not reuse a pilot/development head as a final head.

An infrastructure retry uses the identical configuration, seed, initialization, and batch plan in a new attempt directory, from scratch. No resumable optimizer state or v2 segmentation framework is required for v3. Verify complete runs fit an allocation before selection. Preserve every attempt, including failed or partial output. A verified implementation defect requires a dated result-blind correction and rerunning every affected comparison, not only the unfavorable arm. Parameter changes require a new declared protocol/grid.

Before selection freeze an attempt ledger with exactly 27 matrix slots and separate attempt IDs. Classify failures using logs and integrity evidence before ranking scores:

The selection table contains every slot, attempt/retry lineage, eligibility reason, updates, initialization/batch hashes, unrounded clean metrics and costs. Show all nine three-seed candidate means or explicit ineligibility. An endpoint rate winning means best among the tested rates; it does not expand the grid. Poor accuracy alone never invalidates a numerically valid run.

5. Nested corruptions

Corrupt only queries, after the complete deterministic preprocessing above and before the frozen encoder. Each family starts from the clean tensor; families are never composed. The gallery stays clean. No deletion of frames, interpolation, extra normalization, or clipping follows corruption.

Hash UTF-8 "protocol-v3\0" + sample_id + "\0" + family with SHA-256. Let u be the unsigned big-endian integer from the first eight digest bytes, and seed63 = u & (2**63 - 1). Family names are coordinate_jitter, joint_mask, frame_mask. Severity is excluded. Pin generator/runtime versions and archive corruption-spec and output hashes.

Severity zero returns a byte-identical copy without needing a torso scale. For jitter use the lower median positive finite distance between shoulder midpoint (joints 11/14) and hip midpoint (4/1), over clean person/frame observations whose four confidences are positive. When a sample has no valid own scale, use one fixed fallback: the lower median of valid sample-level scales from the 80-action clean development-training set. Calculate it once, record the contributing IDs/count/hash and value, and freeze it before pilots. Do not recompute it from validation, all-100 final training, or novel queries. If no positive training fallback exists, the preselection gate fails.

Freeze a development-primary geometry panel before corrupted scoring. For each of the 20 development actions choose the two smallest SHA256("protocol-v3|geometry-panel-v1|" + sample_id) values, breaking ties by sample ID: 40 base rows. Add at most eight sentinels from that same query set: minimum-positive and maximum observed support; minimum-positive and maximum second-person support; minimum and maximum valid own torso scale; an observed row with invalid own scale; and an all-missing row. Resolve ties by the same hash/ID order, deduplicate, and record absent categories without substitution. Freeze the sorted ID list, selection-code hash and manifest hash outside Git. This 40–48-row panel is an engineering coverage choice, not a power sample. Constructed fixtures cover absent categories, boundary frames and degenerate scales without changing the real-data set. Verify unchanged shapes, exact mask counts and nesting, one shared jitter field, severity-zero identity, missingness preservation, boundary frames, two-person behavior, own-scale failure/fallback, and deterministic repeats in fresh processes. Record nominal damage and newly masked previously observed joint-frame/person observations, including zero-effective-damage cases. Repeat the panel's complete clean-plus-nine-condition tensor generation in two fresh processes under the pinned environment and require exact array/output hash agreement, alongside the invariant checks. Retain per-case results and fail the gate on any violation; do not use a model-accuracy acceptance floor. During the already-required full 18,929-row passes, record geometry summaries for every row, not just the panel. Observed support means clean confidence > 0 at a person/frame/joint location. For masks report newly zeroed observed count, original observed count and their ratio; a zero denominator is recorded as undefined, not zero. For jitter report per-coordinate RMS displacement sqrt(sum_observed(dx**2 + dy**2) / (2 * N_observed)) / scale, using the actual own/fallback torso scale and unchanged confidence. No observed locations yields an undefined value, counted separately. Summarize min, median, 95th percentile and max over finite per-row values (using NumPy linear quantiles for median/p95), undefined counts, zero-damage counts and fallback count/fraction of all queries. These summaries describe damage, not exclusion or severity-calibration rules. Never drop a troublesome novel query or redraw its corruption. A nonfinite source/artifact defect stops the relevant gate rather than silently excluding a sample.

By October 18, run all nine full-size development query corruption paths and retain their artifacts. Time one full-size pass per family at maximum severity end to end, including source verification, preprocessing, encoder, serialization, projection of the three pilot heads, and complete retrieval. The other cells verify complete identities, invariants, and deterministic outputs. Corrupted development scores may be logged as evaluator evidence but cannot select recipes, severities, learning rates, or duration.

These are post-preprocessing tensor stress tests, not actual camera occlusion, frame recovery, or a new pose estimator. The experiment measures incremental head-objective effects on an encoder already pretrained with pose masking/noise.

6. Locked analysis

Use cosine retrieval and query-weighted top-1, MRR, and R@5, with a fixed gallery ordering for score ties. With one relevant exemplar, mAP equals MRR; do not present it as independent evidence. Report every method, seed, clean/corrupted cell, and both query definitions.

For the primary query set define

drop(m,s,c) = clean_top1(m,s) - corrupt_top1(m,s,c)
Delta = mean over 3 seeds and 9 cells of
        [drop(contextual,s,c) - drop(contrastive,s,c)].

Negative Delta means less degradation for the contextual recipe. Queries have equal weight within each cell; seeds and the nine cells each have equal weight. The nine-cell mixture is a prespecified synthetic stress mixture, not an estimate of real-world corruption prevalence.

Freeze the following analysis before selection and rehash it at opening:

1. Compute a per-query effect by averaging the paired correctness-drop difference over the fixed three seeds and nine cells.
2. Group queries by (setup, performer, repetition, action). Sort actions, group keys, and sample IDs deterministically. In each bootstrap replicate, independently within each of the 20 fixed actions, sample its original number of groups with replacement. Include every available camera view of a sampled group, with multiplicity.
3. Apply that same draw to every seed, method, clean/corrupted condition, and cell. Divide the summed per-query effects by the sampled number of queries. Do not average group means equally: groups can have unequal view counts. Do not force equal action weights.
4. Use 10,000 replicates from numpy.random.Generator(PCG64(2026)) and a two-sided 95% percentile interval, quantiles 0.025/0.975 with method="linear". Do not resample seeds or actions.
5. Also report all three seed effects, all nine cell effects/curves, and all 20 leave-one-action-out effects. For the last, remove that action's query observations only; retain all 20 gallery anchors and the original ranks. Report the range and sign changes descriptively, without new tests.

This interval is conditional on the trained heads, three seeds, official actions/anchors, query definitions, and realized synthetic corruptions. Its resampling assumption is exchangeability of performance groups within a fixed action. It does not provide uncertainty over arbitrary seeds, actions, participants, anchors, or real camera errors; repeated performances from a person may still be dependent. Ten thousand replicates do not create more independent training seeds.

Support the one confirmatory statement, smaller clean-to-corrupt top-1 degradation under the declared mixture, only if the interval's upper bound is below zero. Otherwise report a null or contrary result. Always accompany it with clean accuracy, absolute corrupted accuracy, and the descriptive contextual-minus-contrastive corrupted accuracy difference. A smaller drop does not establish better corrupted retrieval. SupCon, exact-official queries, per-cell differences, seed variation, and sensitivities are descriptive: no additional superiority tests, cherry-picked winning cells, or causal attribution uniquely to contextual neighborhoods.

Before selection produce a metadata-only development table, by action, of query count, performance-group count, views per group, distinct performers and the largest performer's share of groups. Produce its novel counterpart from the already-permitted metadata manifests before opening, without loading poses. This documents residual participant dependence; it does not make groups independent or replace the declared resampling unit. Keep the conditioning and exchangeability caveat beside the primary interval, not only in an appendix. Benchmark variance research supports disclosing sources of variation; three seeds do not justify population inference over arbitrary training runs.

Build and test the report template before selection. In percentage points, show clean accuracy, mean corrupted accuracy, drop, their contextual-minus- contrastive differences and Delta with its interval in one table. Check the identity Delta = clean accuracy gap - mean corrupted accuracy gap. Report all cells/seeds separately as already required. Use these fixed interpretations:

Do not invent a smallest meaningful effect from these papers. Give effect sizes in percentage points and their descriptive per-1,000-query scale (1 point = 10 correct retrievals per 1,000 under the declared mixture), without a new pass threshold, equivalence test or deployment recommendation. The complete recipe comparison cannot isolate contextual neighborhoods from its margins/regularizer. Synthetic stress does not establish natural-camera robustness; that caution is consistent with Taori et al., whose evidence concerns image benchmarks, not these pose recipes.

7. Concrete architecture work

The committed architecture already retains the frozen pipeline next to the separate v2 benchmark. Extend that path rather than route this study through v2 fine-tuning. The table below retains the September 29 implementation inventory; several capabilities are now complete. Its gap descriptions refer to that original base. The completion-release subsection identifies the remaining work; do not rewrite verified v3 operators or analysis merely to repeat this table.

Completion release before selection

The following completion-release requirements record the October 9 amendment for execution-r1. The October 12 subsection below defines its authorized successor; neither changes or relabels the original release.

Adopt one coordinated completion release. The current final-training guard in training/runner.py, novel-role guard in motionbert_features.py, final-lock guard in protocol.py, selection authorizer, corrupted-development evaluator and report dispatch all need completion and testing. The training/extractor files are already in the design's frozen hash set. Do not add bypass wrappers, remove validation or edit the original locks to keep old artifacts eligible.

Use the immutable successor prefix study-v3/revisions/execution-r1/ for its locks, manifests and attempts, under the same canonical artifact root. Keep study-v3/locks/ and all prior attempts unchanged. Bind predecessor protocol and design hashes, successor full protocol/config hashes, release commit, changed-file list and regeneration matrix in a dated amendment receipt. Retain one shared dataset opening ledger and access lock across every revision/version; a new prefix must not create a new test seal. No mutable current-release alias or caller-selected ledger may authorize a scientific operation.

The successor full digest may change for provenance/path bindings. A reviewed field comparison must show identical scientific choices, including IDs/order, recipes, duration/grid, corruption seed prefix and operators, fallback value, metrics, bootstrap and claim rule. Preserve original fallback measurement and identity evidence; regenerate protocol-bound manifests without recomputing or reselecting them. Any scientific difference requires a separate explicit amendment before running, not classification as a path-only change.

Before freezing this release, rehearse its real command path on fixture data: selection authorization and ranking, fresh final training, complete final-run lock, atomic opening, novel-role extraction, projection, retrieval and v3 report. Synthetic final roles must remain scientifically ineligible. Opening, concurrency and crash fixtures use a temporary synthetic dataset registry and artifact root, never production roots or the real opening ledger. Assert that real opening records remain absent or unchanged after rehearsal. Require exactly 180 unique cells, missing/duplicate rejection, ordered primary/official subset equality, unequal-view bootstrap fixtures, the interpretation identity, tamper/cross-release rejection, authorization before loader invocation, concurrent/partial-opening recovery and denial of old/new training after opening. A unit test of an unused analysis function is insufficient: exercise report.py through to the v3 analysis. Retain the independent November audit.

Complete this dependency work first in Week 5, then revalidate caches/pilots, then run development corruptions and capacity measurements. October 18 remains a gate, not permission to compress checks if implementation takes longer.

Finish and commit the extractor, preprocessing, corruption operators, config, dependency, and artifact validation interfaces before fresh scientific cache creation. Current cache provenance binds those code paths and the dependency lock; a later correction may require fresh parity/extraction. Do not weaken a digest check to keep using an invalidated cache. The October 18 gate adds evidence and selection authorization against the unchanged successor protocol digest after its fresh evidence passes; it does not edit the protocol under existing pilot artifacts. A correction to that digest or cache-affecting code requires regenerating affected caches and rerunning affected pilots before the gate passes. Schedule this work and retain superseded artifacts in the capacity ledger.

Fresh v3 auxiliary caches are chosen over a cross-protocol adoption wrapper. Regenerate matching v3-bound manifests without changing sample identities; run encoder parity and two fresh-process final-training extractions on the same GPU. Retain both for repeatability, plus development train/gallery/query caches. Bind the original source/checkpoint hashes in the new sidecars.

For final evaluation, encode the union of eligible novel queries once per condition, then write two immutable role-specific pre-head cache files from those rows: exact-official and primary. The primary file is an exact ordered subset selected by its locked manifest, with matching feature rows verified during export. This deliberately retains duplicate rows so existing head application and role validation need no shared-view abstraction. No additional encoding is needed to change query definition. All nine heads reuse these pre-head caches. Apply each checkpoint to clean gallery and all query conditions, preserving existing byte/provenance validation and checkpoint-bound normalized embeddings. The 180 result cells are not 180 encoder passes: the semantic encoder workload is one clean gallery plus ten query conditions, with batching/sharding as recorded.

The current projector and artifact revalidation include CPU projection and repeated hashing. Include their wall time and temporary memory in measurements; do not assume the GPU extraction rate predicts end-to-end evaluation. Record hashing, loading, preprocessing, encoding, projection, artifact revalidation, retrieval and serialization separately, with host/GPU peaks and saved bytes.

Enforce reuse counts in the campaign manifest: nine development corruption encodings shared by three pilot heads; one novel gallery plus ten query-condition encodings; 21 final pre-head files (ten official, ten primary, one gallery); nine reusable gallery projections and 180 role-specific query projections. Keep immutable role-specific files and existing validation. The October 9 amendment prohibited memoized validation; that prohibition remains part of the immutable 731938a/execution-r1 evidence. The October 12 exception below authorizes only bounded numerical proof reuse in a new successor. CPU projection arithmetic, precision and all content hash checks remain unchanged. Optimize measured orchestration/I/O reuse first. Start with one scientific subprocess at a time on the validated L40S. Group fresh subprocess runs only when measured queue/setup savings justify it and the complete group fits the allocation; preserve independent seeds, attempts, manifests and failure handling. No unmeasured concurrent GPU jobs.

October 12 validation and capacity successor

The researcher authorized ADR0007 after measured P=778.274100575014 seconds made the prior profile's repeated report component with slack 122.145796 hours against the shared 48-hour opening limit. Preserve that no-go; it does not prove all hardware infeasible.

Within one guarded command, a new execution-r2 release may reuse only complete successful deterministic numerical proofs. Each hit must rehash every transitive bound file and match validator/options, role, protocol/design/source/runtime identities. Reuse never substitutes for fresh seal, authorization, timestamp, attempt inventory, resource or calendar decisions. No stat-only trust, persisted proof cache, failed/partial result reuse or cross-command reuse is permitted. Store successful numerical proof markers only, not mutable result objects. Independent recertification starts fresh, and changed content/options fail closed or force full validation.

Use the same canonical artifact root and shared seal, new predecessor-bound locks and explicit scientific-field equality. Follow ADR0007's regeneration inventory: retain original source/identity/fallback evidence through validated bindings and byte-identical manifest rebinding; regenerate parity, five caches, three complete pilots, panel/repeat evidence, nine development paths/projections, report rehearsal and successor timing/admission receipts. Preserve original and execution-r1 artifacts without a cross-release adoption wrapper. Freeze and verify the complete successor before scientific execution; passing software checks alone cannot authorize selection or opening.

8. Capacity gates

The fixed campaign has the following head-update count under the existing sampler. Counts are not GPU-hour estimates. The main table counts the accepted campaign. The original and execution-r1 pilot sets add six retained historical runs and 285,120 updates. With execution-r2 engineering regeneration, 45 complete training runs would be retained before other retries/fixtures, totaling 2,245,140 updates across releases. The earlier two-generation estimate was 42 runs and 2,102,580 updates; preserve it as historical planning evidence.

Target one allocated L40S, the measured extraction hardware; validate the exact head-training and corruption workload there. A different accelerator requires repeatability, memory, and runtime checks before inclusion in the paired run schedule. Keep complete jobs within their approved wall limits. The current launcher requests four CPU slots and 32 GiB aggregate host memory; confirm that is sufficient for full arrays, serialization, and projection rather than assuming from GPU memory alone.

Before selection, create an immutable capacity report that includes:

• All three complete pilot times, checkpoint/validation/I/O time, CPU and GPU memory peaks, scheduler queue/setup observations, and actual saved bytes.
• The complete corrupted development passes, projection for all pilot heads, source/provenance checking, and one-shot scoring. Measure at least one timing segment with all 95,001 auxiliary feature rows resident for final-training memory/extrapolation. Update only fixed 80-action training batches, keeping held-out labels out of optimization. This is separate timing-only engineering outside the 39 scientific runs; record its extra steps/bytes, and make its weights ineligible for selection or final results. No novel inputs or scores are needed.
• A dated CPU/GPU allocation calendar specifying resource-class windows, complete-job limits and queue/setup allowance. Under ADR0007, schema-2 capacity evidence separates CPU/GPU maximum-job, remaining pre-opening and remaining evaluation time; every window declares its class. CPU/I/O within a GPU job stays in that complete GPU wall time. Preserve global non-overlap and one scientific subprocess; an empty queue or prior access is not a reservation.
• Researcher-confirmed available hours against the work packages and final analysis. The inherited 77 planned hours in Weeks 4–14 and zero recorded actual minutes are not evidence of remaining availability or no past work.

Forecast remaining training per arm from its measured update cost and fixed validation/checkpoint overhead; use the slower pilot/full-size timing for that arm and include initialization, hashing, scoring, and report regeneration. Count every fresh extraction, integrity scan, retained attempt, and measured CPU stage. Multiply remaining compute/wall-work estimates by at least 1.25, then add the allocation calendar's queue/setup delays. Apply the same minimum 25% slack to remaining researcher work. The resulting schedule must finish final heads and locks by November 29, full novel evaluation by December 6, and preserve December 7–18 for analysis/delivery. Do not launch overlapping jobs beyond the demonstrated GPU, host-memory, or storage budget.

At opening, compare actual GPU allocation time remaining with the complete GPU-job bound, while still requiring sufficient CPU and GPU evaluation windows and researcher delivery time. Reject scalar/unclassified predecessor receipts as successor evidence. Independent report/recertification compute remains due by December 6; separate accounting does not create later compute capacity.

October 8 pilot scheduler seconds were 502/525/415 and training-stage seconds 286.739518/323.617977/254.941551 (Contrastive/Contextual/SupCon). A provisional head-job estimate is sum(9*S + 3*(S + T*(59380/47520 - 1))), where S is a pilot's scheduler time and T its training time: 4.987 allocated hours, or 6.23 with 25% slack. Completion-release regeneration adds a measured baseline of 5,644 cache-job seconds plus 1,442 pilot-job seconds: another 2.46 hours with slack. The combined 8.69 allocated hours is not a campaign capacity pass. CPU validation, nine development corruptions, novel extraction, projection/report work, queues, retries and software/researcher time are extra. Replace these estimates with successor measurements and the slower required full-size timing before authorization. Do not multiply the slack twice.

The Week 5 work ledger must separate implementation/review, release revalidation, corruption audit, analysis/report verification and resource/availability planning. Existing tracker minutes are inherited estimates, not measured forecasts or researcher availability. Re-estimate each package before the go/no-go decision; retain the deadline only if their dependencies and 25% slack fit. Set scheduler wall limits from complete measured jobs plus slack and scheduler granularity, rather than copying the pilot launcher's 12-hour request by default.

Storage uses bytes, not ambiguous GB labels. Let

available = min(filesystem_free_bytes, project_quota_remaining_bytes)
required = 200 * 2**30 + 1.25 * forecast_remaining_peak_bytes
pass iff available >= required.

The September 29 planning envelope was 100 GB incremental including slack. The October 9 campaign-wide planning envelope was 120 decimal GB including slack, while retaining the 200 GiB reserve. It includes retained predecessor evidence and the planned completion-release regeneration; it is not 120 GB of new work on top of files already present. It deliberately keeps the role-specific files expected by the existing architecture. The following uncompressed payload estimates include float32 features, int64 labels, and 20-character Unicode sample IDs; compression is not credited.

This table is the pre-execution-r2 estimate. ADR0007 adds a retained fresh engineering generation; inventory all actual retained bytes, successor outputs and remaining temporary peaks before admission. No new numeric envelope or capacity pass is inferred here. The current-space/reserve equation above remains mandatory, and no historical evidence may be deleted to fit the plan.

The original roughly 4.14 million development/final per-query result rows, 511,083 development-corruption rows and successor pilot records are covered by the allowance, not assumed free. Pilot epoch-zero diagnostics do not retain another full embedding cache; compute their summaries in memory. At epoch 20, retain one checkpoint artifact per pilot and reference it from both the epoch record and completion record rather than copying it. The 10 GB allowance is bounded, not permission for unlimited failed attempts. Replace estimates with actual bytes at the October gate, including all new artifacts already created.

The September 29 reading failed the original 100 GB envelope by about 41.95 GB; retain it as historical evidence. The October 8 observation of 1,008.48 GB quota free is newer but also not a reservation. Remeasure both quota and filesystem space and apply the remaining-peak formula; request additional authorized capacity only if that current calculation fails. If a new canonical root is necessary, register and verify transfers without relabeling or deleting old evidence. Filesystem space beyond quota is not usable capacity. The peak forecast includes all yet-to-be-created retained checkpoints, intermediate pilot weights, batch plans, clean/corrupted pre-head caches, primary/official duplicates, projected embeddings, ranks, logs, failed attempts, immutable output staging, and reports. Already present files are reflected in the current free-space reading, not counted twice. After work begins, use the remaining measured forecast rather than demanding the 120 GB envelope again. Retain original v1/v2 artifacts; deleting evidence is not a capacity remedy. Recheck before each large job and before opening, and stop new work if the reserve would be crossed. The September 29 observation permits at most 46.44 decimal GB of additional unmargined peak work under this formula; it does not reserve that amount.

October 18 go/no-go: all three objectives mathematically verified; three complete, valid 20-epoch pilots; unchanged fixed development episode; nested corruptions and fallback validated; completion release and end-to-end fixture rehearsal passed; successor cache/pilot validation complete; full development extraction/projection/scoring verified; analysis/report template tested; every historical seal audit resolved; and the complete measured schedule/storage/availability forecast passes. Otherwise stop before the sweep and record a result-blind amendment. Do not reduce a single method's budget, discard seeds/cells, or claim a completed comparison.

9. Locks, opening, and failures

Keep POSE_EMBED_DATA_ROOT for licensed inputs and a stable absolute POSE_EMBED_ARTIFACT_ROOT for all scientific records. Resolve all v3 outputs under its study-v3/ namespace and the explicit immutable revision described in Section 7. Retain predecessor lock paths without overwriting them. The shared dataset-level opening record remains at the canonical root and binds the source inventory/physical-source hashes and the historical-root audit. Neither a caller-selected ledger nor a different study namespace bypasses it.

Use a staged lock sequence to avoid circular prerequisites:

1. Before fresh parity/caches and pilots: complete immutable v3 protocol and design/pilot specification with fixed recipes, 20 epochs, identities, numerical/corruption/analysis conventions, fallback binding, finite LR grid, hashes, and the omitted control decision. It authorizes auxiliary engineering only. Freeze all cache-affecting code in the release.
2. By October 18, before selection: append an immutable selection authorization binding the successor protocol digest used by its fresh pilots to corruption/analysis verification, pilot records, the complete run matrix/manifests, and the capacity decision. Scientific settings are unchanged; the new provenance digest is explicit. There are no final checkpoint hashes yet; this is not a test-opening lock.
3. After selection/final training: selected configs and timestamps; nine final checkpoint/run-manifest hashes, seeds, initializations/batches, source/checkpoint/input-contract bindings, code/dependency hashes, exact primary/official/gallery IDs, evaluation plan and complete final-run-set. Record the researcher-controlled final lock only after its dependencies.
4. Immediately before opening: validate the complete cross-binding and all prior ledgers, verify no scientific training remains active, recheck capacity, and atomically create the immutable dataset opening event and its v3 authorization. Coordinate creation so a partial/crashed transaction fails closed and can only complete the identical lock, never a new one. Only then invoke the novel annotation loader/preprocessor/encoder.
5. After opening: every extraction shard, head application, evaluation cell, and report must match the same source, protocol, run-set, opening, cache, code, checkpoint, and corruption bindings. All scientific training and selection for this study/source are forbidden across v1/v2/v3 paths.

A crash after the opening event does not reseal the data. Preserve the attempt and retry only the identical authorized operation and inputs. Missing/malformed required cells remain visible and prevent a complete report. A scientific or implementation defect discovered after opening cannot be repaired by silently changing the confirmatory pipeline: preserve results, explain the limitation, and label any necessary corrected analysis as a post-opening deviation.

10. Execution calendar and verification

Test the implementation in layers. Python changes require locked dependency sync, python scripts/verify_workspace.py, Ruff lint/format checks and the relevant unit/integration suite, followed by the complete Python suite before the scientific release. Required fixtures cover loss values/backward rules, sampler pairing, pilot exclusion, failed-candidate selection, corruption nesting, fallbacks, unequal-view bootstrap weighting, fixed gallery sensitivity, cross-version seal denial, first-open ordering/concurrency, provenance mutation, and report matrix completeness. Run scheduler shell syntax checks and GPU parity/repeatability on the allocated hardware.

For tracker changes run its lint, typecheck, tests, build, local database reset, database tests and security advisors. Migrate the public v3 plan with expected row versions, returned-row conflict checks, trigger-managed version increments, RLS and private ownership preserved. Preserve hosted task evidence and actual minutes. Closed-week changes require an explicit audited reopen reason; future-week plan adoption does not silently rewrite closed history. Verify hosted Auth sign-up/redirect restrictions before enabling editing. Publish the matching fallback and protocol bundle; do not merely redeploy an older app.

No tracker migration/deployment, experiment, or scientific code modification is performed by this planning document. The independent-research governance and recorded BU determination remain in force; no advisor approval is required. Reconcile dataset-terms acceptance with private evidence, and do not invent researcher time, approvals, or completed gates.

11. Completion criterion

The study is complete when the three selected recipes have nine verified final heads, all 180 declared result cells reproduce from locked per-query evidence, the conditional primary analysis and all descriptive tables/curves regenerate, and the December report/poster explains absolute performance, seed variation, cost, failures, adaptations, and limits. Deliver permitted code, configs, manifests, compact summaries, source/license notices, and reproduction instructions. Keep licensed poses, checkpoints, caches, full artifacts, and private governance records outside Git.

The final answer can be positive, null, or negative. If a capacity or validity gate prevents the declared campaign, deliver an honest blocked-study record and amendment; that is not a completed empirical comparison. This is a final execution design with explicit go/no-go rules, not evidence that unmeasured GPU work or researcher availability has already been secured.