# Frozen one-shot pose robustness

Final v3 December study: frozen MotionBERT, three training recipes, three paired seeds, 20 epochs, and nine query-only corruption conditions. Three engineering pilots precede 27 selection runs, nine fresh final heads, and 180 locked evaluation cells. Implementation, measured capacity, and test-opening gates remain prerequisites; null or negative results are valid outcomes.

**Research question:** Does the full contextual training recipe reduce one-shot retrieval degradation relative to a standard contrastive recipe when only novel pose queries are corrupted?

**Schedule:** 2026-09-15 to 2026-12-18
**Required-task progress:** 22/60 (37%)
**Data source:** supabase

## Protocol

Complete the three-recipe study faithfully: nine verified final heads, all 180 declared result cells, reproducible conditional analysis and descriptive results, and a report with costs, failures and limitations. A positive, null or negative contextual result is successful. A failed capacity or validity gate requires an honest blocked-study record and result-blind amendment; it is not a completed empirical comparison.

**Binding source:** plan/research-plan.v3.md and decision 0004 define the scientific v3 scope; decision 0006 (October 9) amended execution, verification and reporting. Decision 0007 (October 12) records researcher authorization for command-local validation reuse and separate CPU/GPU time accounting in execution-r2, with full fresh successor regeneration and no cross-release artifact adoption. Scientific settings remain fixed. The independently reviewed execution-r2 software passed the stable full local suite and source release f101f0e9d9bd64275d2088deab1b0bdcfcc53620 is frozen. Separate SCC installation and the allocated 128-test CPU rehearsal passed independent verification. Successor preparation 8045152 passed independent completion review of its accounting, commands, resources, seals, all eight original manifest identities and original → r1 → r2 scientific equality. Successor supplementary checks 8048540 passed independent completion review of 86 executed fixture checks, commands, resources, seals and reproduced metadata-only dependence tables. Fresh scientific regeneration, full admission and measured CPU/GPU capacity remain pending before selection. Original locks, prior releases, evidence and closed history remain unchanged. Planning documents are not executable authorization or final-test locks.

### Locked decisions

- **Final scope:** One NTU RGB+D 120 one-shot study comparing Contrastive, full Contextual, and single-view supervised contrastive (SupCon). The confirmatory comparison is the full contextual recipe versus the standard contrastive recipe; it does not isolate the contextual term from margins or regularization.
- **Frozen representation:** Freeze the generic pretrained MotionBERT encoder in evaluation mode. Pool time and people to 8,704 features; train a biased linear 8,704-to-2,048 head with L2-normalized output and zero dropout (17,827,840 trainable parameters).
- **Fixed duration and pairing:** Use exactly 20 epochs for every pilot, selection run and final run, with paired seeds 7, 17, 29 and physical P=8, K=4 batches of 32. Pair initial head states and exact batch plans across recipes and learning-rate candidates within each seed. No automatic ten-epoch fallback, early stopping or best-epoch choice.
- **Required campaign:** Three valid execution-r2 engineering pilots + 27 clean-development selection runs + nine fresh final heads = 39 admitted campaign runs and 1,960,020 updates. Retain the original Week 4 and execution-r1 pilot sets separately: 45 complete runs and 2,245,140 updates across generations before other retries/fixtures. The prior 42-run estimate remains historical. Final evaluation remains 180 cells.
- **Current planning status:** Week 5 remains active. The separate SCC installation of frozen source release f101f0e9d9bd64275d2088deab1b0bdcfcc53620 passed verification. CPU rehearsal job 8044010 completed October 12 at 21:54:03.440725 UTC on scc-ke1: 128 tests passed with no skips, failures or errors; accounting failed=0/exit_status=0 and 194 scheduler seconds. Independent completion review accepted all three monitored commands, nine resource samples and both real dataset seals. Its nine heads and 180 cells are synthetic fixtures only, not eligible scientific heads or results. Receipt: study-v3/operations/week5-r2-rehearsal-verification/20261013T002413Z/verification.json; SHA-256 963babe02eba70d34f3f9c9aa47af21145de387ca4d5a1ad6ccf9c226fb43381. Successor preparation job 8045152 completed October 13 at 00:31:55 UTC on scc-ke1 with four CPU slots, accounting failed=0/exit_status=0 and 44 scheduler seconds. Independent completion review wrote its passed receipt at 06:39:11.971297 UTC after verifying 73 metadata bindings, 616 source bindings, three commands, seven resource samples, both seals, all eight byte-identical original manifests and original → r1 → r2 scientific equality. Receipt: study-v3/operations/week5-r2-preparation-verification/20261013T063819Z/verification.json; SHA-256 0a3823e66a97192c5665185edbb908bf5b44c6dd9b451328d95d96476a84d785. Successor protocol SHA-256: 18f7d17fa81e7ae3a6b1e5deddf42d9eda589d96e5a13968188a8ce275bfb95f. Supplementary CPU checks job 8048540 completed October 13 at 08:31:43.006302 UTC on apolkovnikov-pub@scc-h30 with four CPU slots, accounting failed=0/exit_status=0 and 87 scheduler seconds. All 86 fixture checks executed and passed without skips. Independent review ran from 20:42:57.604956 to 20:43:00.762966 UTC, verifying 108 metadata bindings, 616 source bindings, three commands, seven resource records and both seals. The actual dependence tables reproduced from locked IDs: 18,929 development queries and 18,884 primary novel query IDs, each across 20 actions. No novel poses or outcomes were accessed. Receipt: study-v3/operations/week5-r2-additional-verification/20261013T204221Z/verification.json; SHA-256 7703859e50361123de9054fba446c8384d423b8c549b69eab92720314c46c3d2. The submitted launch at 07:58:52.315989 UTC and initial queued observation at 08:17:32.236172 UTC remain retained. Launch: study-v3/operations/week5-r2-launches/20261013T075848Z/launch.json; SHA-256 0e77ecb106c977dd9922f473b80f29fd7475d8b4f5b243d9db9b34cc92368fd8. Successor geometry job 8070391 was submitted October 13 at 20:46:55.956765 UTC and observed queued (qw), four CPU slots, at 20:49:22.424937 UTC. No geometry operation, log or resource records existed at that observation; no outcome is claimed. Launch: study-v3/operations/week5-r2-launches/20261013T204651Z/launch.json; SHA-256 a8c55f5a6054996d49cb2abab304a8b8925914a4cf0495a2f0368f84e4177452. The geometry-only 1-GiB bound passed the fresh quota/filesystem storage reserve; measured campaign capacity remains pending. Next manual status check: October 13, 2026 at 5:55 PM EDT (21:55 UTC); hourly polling remains in force. Task 02 remains in progress; task 04 is blocked by incomplete measured capacity. Tasks 01 and 03 retain their verified execution-r1 history; all scientific gates remain pending. Full fresh execution-r2 regeneration, full A and measured CPU/GPU capacity are required before selection. Preserve the prior measured-profile no-go, original evidence and every failed attempt. Actual researcher time remains unreported by request. The scientific matrix, fixed seeds/epochs/physical batches, all checks and the shared dataset seal are unchanged.
- **Measured capacity prerequisite:** The October 9 120-decimal-GB planning envelope covered the original and execution-r1 generations; the historical 100 GB envelope is also retained. Execution-r2 requires a new actual retained/remaining inventory including all regeneration, corruption, timing and temporary artifacts; no larger numeric envelope or capacity pass is invented. Current min(filesystem free, quota remaining) must be at least 200 GiB + 1.25 times remaining peak bytes; existing files count once. Schema-2 CPU/GPU maps and dated classified windows must fit complete commands with slack and queue/setup; CPU/I/O within a GPU job remains part of its GPU wall time.
- **October 18 go/no-go:** Before selection require the authorized execution-r2 completion release and full fixture-to-v3-report rehearsal, fresh successor parity/five caches/three pilots, fixed geometry panel and nine complete development corruption paths, fixed analysis/report template, resolved seals and a measured dependency/schedule/storage/availability ledger. Use classified CPU/GPU job/remaining-time maps and nonoverlapping dated windows; preserve all CPU/I/O overhead within its whole allocated command. Apply at least 25% remaining-work slack plus queue/setup delays; otherwise stop before selection.
- **Test seal:** Preserve the original design/locks, Week 4 closure and immutable execution-r1 evidence. Prepare execution-r2 beneath the same canonical root with explicit predecessor/successor bindings, unchanged scientific fields and fresh caches/pilots; no cross-release adoption or sidecar relabeling. The successor digest may change for provenance paths and stays fixed within its admitted campaign. One shared dataset seal spans all releases. Lock all nine final heads and atomically authorize before novel loading; forbid subsequent training/selection across versions.
- **Historical access audit:** Register and audit all historical artifact roots and releases. Historical aggregate loading deserialized the shared annotation container; the seal claim concerns absence of novel preprocessing, representations, retrieval, visualization, scoring or tuning. Unknown or unreadable roots block the gate; a new namespace cannot restore a compromised test.
- **Governance:** The researcher records decisions and result-blind amendments. No advisor approval is required. Preserve the recorded BU determination and reconcile dataset-terms acceptance with private evidence. The unlicensed contextual repository is reference-only. Retain permitted upstream attribution; do not invent approvals, actual hours, results or completed gates.
- **December delivery:** Finish final heads and locks by November 29, complete novel evaluation by December 6, and preserve December 7-18 for analysis and delivery. Regularizer-only control is omitted before new scores; the 26-method v2 benchmark, fine-tuning and Multi-Similarity are future studies.
- **October 9 execution amendment:** Complete the currently disabled final-training, novel-extraction, selection/final authorization, corrupted-development evaluation and actual v3 report interfaces before selection. A new immutable completion release and fresh admission evidence are chosen over bypass wrappers or relabeled artifacts. No scientific settings, accuracy-based acceptance rules or new evaluation arms are added. Release 731938a implements these interfaces and passed the isolated SCC command rehearsal; fresh scientific admission evidence remains pending.
- **October 12 validation/capacity amendment:** The researcher explicitly authorized ADR0007 after the measured-profile no-go. A new execution-r2 release may reuse only successful deterministic numerical proof markers within one guarded command, rehashing complete transitive contents on every hit. Dynamic seal/authorization/time/attempt/resource checks stay fresh; no persistent or cross-command proof reuse. Independent recertification starts empty. Separate schema-2 CPU/GPU accounting preserves whole-job overhead, all checks, budgets, deadlines and the storage reserve. Full fresh parity/caches/pilots/corruption/report/timing/admission regeneration is required. The independently reviewed implementation passed 1,373 local tests with two historical AVSL skips; source release f101f0e9d9bd64275d2088deab1b0bdcfcc53620 is frozen. Separate SCC installation and the allocated 128-test CPU rehearsal passed independent verification. Successor preparation 8045152 passed independent completion review of its accounting, commands, resources, seals, all eight original manifest identities and original → r1 → r2 scientific equality. Successor supplementary checks 8048540 passed independent completion review of 86 executed fixture checks, commands, resources, seals and reproduced metadata-only dependence tables. Fresh scientific regeneration, full A, measured capacity and all scientific gates remain pending.

### Experimental design

- **Dataset and episode:** Bind 113,945 usable HRNet annotations plus 535 official missing-skeleton exclusions (114,480 nominal captures), source hashes and the frozen checkpoint. Preserve the existing 80-action development-training / 20-action validation split, official 20 novel actions, all official exemplars and the exact historical development anchor/query identities.
- **Selection and final data:** Use 76,013 training rows, 20 clean gallery references and 18,929 primary queries for development. Final training uses all 100 auxiliary actions and 95,001 rows. No training-subset contingency or novel outcome access is allowed.
- **Two query definitions:** Primary queries exclude each clean gallery sample and every synchronized view of its underlying performance, keyed by setup, performer, repetition and action. The exact-official secondary set excludes only the exemplar. Verify and lock the exact ordered lists: expected novel counts are 18,884 primary and 18,924 exact-official queries.
- **Deterministic inputs:** Use the adopted float32 tensor with two people, 100 frames, 17 H36M joints and x/y/confidence channels. Keep camera normalization, tracking, joint conversion, uniform temporal sampling, padding and spatial normalization fixed. Random movement is off; compare encoder parity on the same locally corrected confidence tensor.
- **Contrastive recipe:** Standard cosine-margin contrastive loss with positive margin 0.9 and negative margin 0.6. Average positive and negative strictly active hinges separately; empty active sets contribute zero.
- **Contextual recipe:** 0.4 L_context + 0.6 L_contrast + 0.1 L_reg, positive/negative margins 0.75/0.6, neighborhood 4, epsilon 0.05, STE alpha 10 and target mean similarity 0.25. Implement the same-label target in Equation 5 and independently verify labeled forward/backward fixtures. These fixed transfer settings are pose-study adaptations, not proven optima.
- **SupCon recipe:** Single-view supervised positive log-softmax at temperature 0.07 with no extra temperature multiplier. It is a required descriptive comparator, not a second confirmatory superiority test.
- **Optimizer and sampling:** Float32 features and parameters; constant-rate AdamW, betas (0.9, 0.999), epsilon 1e-8, amsgrad false, weight decay 1e-4 on both weight and bias. No scheduler, mixed precision, clipping or augmentation. Epochs contain ceil(N/32) physical batches under archived paired sample plans; this is a balanced sampling budget, not one visit per sample.
- **Engineering pilots:** One full 20-epoch pilot per arm, seed 7 and learning rate 3e-4. Retain immutable checkpoints and complete clean development scores at epochs 5, 10, 15 and 20, plus epoch-zero diagnostics. Require mathematical/learnable-fixture checks, finite updates, nonzero vectors and no total numerical collapse. Pilots are ineligible for selection; real-data accuracy has no acceptance floor.
- **Learning-rate selection:** For each arm attempt rates 1e-4, 3e-4 and 1e-3 with all three seeds: 27 fresh 20-epoch development runs. Rank by unrounded mean epoch-20 clean development top-1, then mean MRR, then lower rate. A candidate needs all three valid seeds. Preserve failures and disclose unequal eligibility; an arm with no valid candidate blocks the campaign.
- **Fresh final heads:** Lock the selected rate for each arm and selection evidence, then train nine fresh heads on all 95,001 auxiliary rows for 20 epochs. Keep epoch-20 weights; do not reuse pilot or selection heads. Infrastructure retries restart identical settings in new immutable attempt directories. Corrections rerun every affected comparison under the amendment policy.
- **Fresh caches and evaluation:** Create v3-bound manifests with unchanged identities and fresh clean caches, including parity and two repeat extractions on the same GPU. Encode the union of novel queries once per condition, then export verified immutable primary and exact-official cache files. All nine heads share pre-head caches: the encoder workload is one clean gallery plus ten query conditions, not 180 encoder passes.
- **Metrics:** Use cosine retrieval with a fixed gallery order for ties and query-weighted top-1, MRR and R@5. With one relevant exemplar, mAP equals MRR. Report all methods, seeds, clean/corrupted cells and both query definitions, alongside actual resource costs and retained failures.
- **Failure adjudication:** Freeze infrastructure, numerical/health, implementation/provenance and unresolved failure categories before ranking. Infrastructure needs logs/accounting, cause resolution and identical fresh retries; allow one automatic retry, then record cause/remedy and renewed capacity. Numerical failures cannot be rerun to shop for success. Block unresolved eligibility; correct implementation defects across every affected comparison. Report all 27 slots, retries, updates, hashes, unrounded metrics, eligibility and cost. An endpoint winner does not expand the grid.
- **Measured execution and reuse:** Reuse nine development corruption encodings across three pilot heads and one novel gallery plus ten query-condition encodings across nine final heads. Retain 21 final pre-head files, nine gallery projections and 180 query projections. Time hashing/loading, preprocessing, encoding, CPU projection/revalidation, retrieval and serialization separately. Start with one scientific subprocess at a time on the validated L40S; preserve float32 arithmetic and validation. Group fresh runs only after measured savings and whole-allocation capacity checks.
- **Provisional remaining compute:** Pilot extrapolation estimates 6.23 allocated GPU hours for selection/final head jobs including 25% slack; fresh completion-release caches/pilots add 2.46 hours with slack. Combined 8.69 hours excludes CPU validation, development/novel corruption extraction, projection/report work, queues, retries and engineering/researcher time. Replace with successor and full-size measurements before the gate; do not apply slack twice.

### Corruptions

Apply each family independently to clean, fully preprocessed query tensors before the frozen encoder; keep the gallery clean. Derive a per-sample family seed from SHA-256 of protocol-v3, sample ID and family, excluding severity. Share the realized corruptions across methods/seeds and nest severities. These are synthetic tensor stress tests, not measured camera occlusion or pose-estimator shifts.

- **Coordinate jitter:** Reuse one seeded CPU float32 Gaussian x/y field at 1%, 2.5% and 5% of clean torso scale. Add noise only where original confidence is positive; leave confidence and missing/padded observations unchanged. Do not clip or renormalize.
- **Joint or limb masking:** Shuffle the fixed ordered limb groups once per sample; take the first 3, 6 or 8 joint trajectories from that same ordering. Zero x/y/confidence across all people and frames, with exact nested masks. Already absent joints are not replaced by visible ones.
- **Consecutive-frame masking:** For L = 10, 25 or 40, use start = floor(u × (101 − L) / 2^64), where u is the first eight hash bytes as an unsigned big-endian integer. Zero the consecutive block across people, joints and channels. Blocks are nested; no frames are deleted, resampled or interpolated.
- **Scale and fallback:** Use the lower median positive finite shoulder-to-hip midpoint distance among clean observations with four positive source confidences. If a sample lacks a scale, use the fixed lower median of valid sample-level scales from the 80-action training set only. Freeze its value, contributing IDs/count/hash before pilots; never recompute it from validation or novel data.
- **Identity and geometry checks:** Freeze 40 base rows: two smallest SHA256(protocol-v3|geometry-panel-v1| + sample_id) per development action, with sample-ID tie-breaks. Add up to eight deterministic support/second-person/scale/fallback/all-missing sentinels, deduplicate and log absent categories. Record sorted IDs/hashes outside Git before corrupted scores; use constructed fixtures for absent cases. Repeat all panel conditions in two fresh processes and require exact output agreement plus identity/nesting/missingness checks. This is engineering coverage, not a statistical sample.
- **Preselection verification:** By October 18 complete all nine full-size development query paths. Time a maximum-severity pass per family through source checks, preprocessing, encoding, storage, projection of three pilot heads and full retrieval. Corrupted development scores may validate the evaluator but cannot choose settings, severities, rates or duration.
- **Full-query damage telemetry:** During all nine full 18,929-query development paths, record clean observed support (confidence > 0), newly masked counts/fractions, jitter per-coordinate RMS sqrt(sum_observed(dx^2 + dy^2)/(2*N_observed))/own-or-fallback scale (undefined for zero observed locations), fallback fraction, zero-damage and undefined-denominator counts. Report min/median/95th percentile/max over finite per-row values, using NumPy linear quantiles for median/p95. No outcome-based severity changes, query exclusions or redraws.

### Primary analysis

`drop(method, seed, cell) = clean_top1(method, seed) − corrupted_top1(method, seed, cell)`

`Delta = mean over the three fixed seeds and nine corruption cells of [drop(Contextual, seed, cell) − drop(Contrastive, seed, cell)]. Negative Delta means less degradation for the full contextual recipe.`

- The primary estimand uses the primary query set. Queries have equal weight within a cell; all three seeds and all nine cells have equal weight. The nine-cell mixture is prespecified synthetic stress, not real-world corruption prevalence.
- Average the paired correctness-drop difference over the fixed seeds and cells for each query. Cluster by setup, performer, repetition and action. Independently within each of the 20 fixed actions, resample its original number of performance groups with replacement, retaining all camera views and multiplicities.
- Apply the same cluster draw to every method, seed and condition; divide the summed effects by the sampled query count. Do not equally weight unequal-view groups or force equal action weights. Do not resample seeds or actions.
- Use 10,000 bootstrap replicates from numpy Generator(PCG64(2026)), with a two-sided 95% percentile interval and linear 0.025/0.975 quantiles. Freeze and test the analysis before selection; rehash it at opening.
- The interval is conditional on the trained heads, three seeds, fixed official actions/anchors, query definitions and realized corruptions. It assumes exchangeable performance groups within each action; it does not quantify arbitrary seed/action/participant/anchor or real-camera variation.
- Support the single confirmatory statement of smaller clean-to-corrupt top-1 degradation only if the interval upper bound is below zero. Otherwise report a null or contrary result. Always show clean and absolute corrupted accuracy and the descriptive Contextual-minus-Contrastive corrupted-accuracy difference; smaller degradation need not mean better corrupted retrieval.
- Report all three seed effects, nine cell effects/curves and 20 leave-one-action-out effects. For each sensitivity, remove only that action’s query observations while keeping all 20 gallery anchors and original ranks. Report range and sign changes descriptively, without new tests.
- SupCon, exact-official queries, individual cells, seed variation and sensitivities are descriptive. No extra superiority tests, selected winning cells or unique causal attribution to contextual neighborhoods. Report the effects of margins and regularization as part of the full recipe.
- Missing, invalid or unbound cells prevent a complete report. Keep every attempt and preserve the test-opening record after a crash. Post-opening scientific or implementation corrections must be disclosed as deviations; they cannot silently replace the confirmatory pipeline.
- Before selection, test the actual v3 report entry point and its exact 180-cell fixture matrix. Show clean accuracy, mean corrupted accuracy, drop, contextual-minus-contrastive gaps and Delta/interval in percentage points together; verify Delta = clean gap - mean corrupted gap. Keep every seed/cell and both query definitions.
- Prewrite claim cases: upper Delta bound below zero supports smaller degradation; a nonpositive absolute corrupted gap prevents describing higher corrupted accuracy. An interval containing zero is inconclusive, not equivalence/no effect; lower bound above zero favors greater contextual degradation. No invented useful-effect cutoff or extra test. Explain percentage points also as descriptive correct retrievals per 1,000 queries.
- Before selection report development metadata counts by action: queries, performance groups, views/group, distinct performers and largest performer share of groups. Produce the novel counterpart from permitted metadata before opening. These tables expose residual dependence without changing the bootstrap or implying participant-generalization.

### Manual actions

- **By 2026-10-18, Complete the dated CPU/GPU and researcher schedule:** The researcher confirmed up to 5 hours weekly normally and 10 if necessary, with limited Thanksgiving time; budget no guaranteed hours for November 23 to 29. Under decision 0007, build a dated ordinary-queue CPU/GPU calendar with separate resource classes, complete-job limits, queue/setup allowance and explicit contingency weeks. Count all CPU and I/O overhead inside a GPU allocation in its GPU job time. Apply at least 25% slack to remaining compute and researcher work; prior access is not a reservation.
- **By 2026-10-18, Verify current remaining capacity:** Remeasure quota and filesystem space for the October 18 decision. Use 200 GiB reserve plus 1.25 times remaining peak bytes, including fresh execution-r2 regeneration, projections, failures and staging; existing files count once. The October 9 estimate was 120 decimal GB including slack, before the authorized execution-r2 regeneration. A fresh inventory of all retained evidence and execution-r2 additional peak storage is pending; the historical estimate does not establish successor capacity. October 8 observed 1,008.48 GB quota free; neither that nor September shortfalls establish current capacity. Request additional authorized allocation only if the current calculation fails.

## Week 1: Protocol and access

**Dates:** Sep 15-20  
**Phase:** Protocol  
**Budget:** 8 hours planned, no time recorded  
**State:** closed  
**Ready to close:** yes

Turn the prospectus into a documented, hash-bound, testable protocol and prove that the required data and compute are reachable.

**Deliverable:** A documented, hash-bound protocol-v1, verified data and checkpoint inventory, and measured GPU profile.

### Tasks

- [x] Finalize protocol-v1: Documented, hash-bound protocol document.
- [x] Verify licensed inputs: Input inventory and checksum report.
- [x] Profile the GPU path: Compute profile with a justified physical batch-size decision.
- [x] Record governance decisions: Decision evidence or a visible blocking record.
- [x] Close the weekly record: A complete Week 1 progress record.

### Advancement gates

- **met:** Protocol-v1 is documented under researcher control and its hash is recorded.
- **met:** Verified input inventory matches the recorded protocol input contract.
- **met:** A representative GPU profile confirms a feasible physical batch size.
- **met:** Required BU governance determination is recorded.

### Risks

- NTU or checkpoint access may require manual license acceptance.
- GPU availability may not match the planned memory or runtime assumptions.
- Protocol amendments or the BU determination may remain unresolved after technical setup begins.

**Review prompt:** Are the research question, test-opening rule, primary estimand, synchronized-view exclusion, and claim language documented as the binding protocol?

## Week 2: Immutable data manifests

**Dates:** Sep 21-27  
**Phase:** Data  
**Budget:** 8 hours planned, no time recorded  
**State:** closed  
**Ready to close:** yes

Create deterministic sample identities and split manifests that make class, exemplar, and synchronized-view leakage impossible.

**Deliverable:** Checksummed manifests for development, final training, anchors, and both novel-query definitions, plus a leakage audit.

### Tasks

- [x] Parse NTU sample identities: Tested parser and normalized sample inventory.
- [x] Generate split manifests: Seven deterministic, checksummed manifest files.
- [x] Prove leakage invariants: Passing leakage and duplication test suite.
- [x] Publish the manifest audit: Human-readable manifest audit report.

### Advancement gates

- **met:** All seven manifests have frozen counts and checksums.
- **met:** Automated checks find zero forbidden overlap or duplicate identities.
- **met:** Every synchronized-view exclusion is reproducible from the manifest audit.

### Risks

- NTU filename parsing mistakes can silently merge performances.
- Synchronized cameras can leak the anchor performance into queries.
- Upstream one-shot files may use a different action index convention.

**Review prompt:** Does the primary deduplicated query definition answer the intended question while the exact-official query set preserves comparability?

## Week 3: Verified v3 implementation and sealed preparation

**Dates:** Sep 28-Oct 4  
**Phase:** V3 implementation and design freeze  
**Budget:** 8 hours planned, no time recorded  
**State:** closed  
**Ready to close:** yes

Complete the frozen one-shot v3 protocol and pilot specification, corruption operators and fallback, versioned validators, shared dataset seals, source access separation and pilot instrumentation; freeze cache-affecting code and configuration before fresh extraction.

**Deliverable:** Verified v3 software with historical evidence preserved, plus auxiliary-only SCC preparation, measured fallback evidence, a complete root audit and an immutable design lock before Week 4 caches and pilots.

### Tasks

- [x] Port minimal MotionBERT inference: Device-agnostic frozen encoder with attribution.
- [x] Build provenance-aware feature caching: Repeatable cache plus machine-readable provenance manifest.
- [x] Implement retrieval evaluation: Metric module with exact fixture tests.
- [x] Measure extraction reproducibility: Determinism report and remaining-compute forecast.
- [x] Validate v2 retrieval and protocol transition: New v2 correctness fixtures and sealed-test validation; no reuse of v1 completion as v2 completion.
- [x] Implement and verify the v3 frozen study path: Reviewed v3 implementation and passing applicable software checks; no real-data preparation or GPU result inferred.
- [x] Prepare auxiliary inputs and freeze the real v3 design on SCC: Immutable auxiliary preparation and fallback evidence, canonical root audit and a validated v3 design lock from the real SCC inputs.

### Advancement gates

- **met:** Two clean feature extractions are identical within the declared tolerance.
- **met:** Hand-calculated top-1, MRR, and R@5 fixtures pass.
- **met:** Every cache contains complete provenance and incompatible reuse fails closed.
- **met:** Measured extraction time and storage fit the Week 3 limits.
- **met:** V2 multi-positive retrieval and full-roster test-seal checks pass independently of the historical one-shot metric gate.
- **met:** V3 protocol, corruption, shared-seal, source-separation and pilot instrumentation checks pass while historical behavior is preserved.
- **met:** Real auxiliary-only preparation, measured fallback and canonical root audit are bound to an immutable v3 design lock before fresh feature extraction.

### Risks

- Historical v1/v2 caches remain preserved; Week 4 must generate fresh v3 parity and caches under the frozen design.
- Week 4 bounded storage passed at the recorded observation; the October 18 full-campaign capacity and availability gate still needs fresh evidence.
- Actual researcher minutes remain unreported and are displayed as No time recorded; the zero placeholder is not a claim of zero work.

**Review prompt:** Do auxiliary-only preparation, the measured fallback, the shared seal audit and the immutable design lock authorize fresh Week 4 caches and three engineering pilots?

## Week 4: Verified recipes and full pilots

**Dates:** Oct 5-11  
**Phase:** Implementation and engineering  
**Budget:** 8 hours planned, no time recorded  
**State:** closed  
**Ready to close:** yes

Verify the three fixed recipes and the frozen-head path, then complete three full 20-epoch pilots under the immutable v3 design.

**Deliverable:** Loss fixtures, paired initialization/batch evidence, fresh parity and clean caches, and three complete pilot records with timing and diagnostics.

### Tasks

- [x] Freeze and validate the v3 path: An immutable v3 design/pilot specification and passing version, provenance and seal checks.
- [x] Verify losses and paired training: Verified three-arm numerical behavior and paired initialization/batch hashes.
- [x] Generate fresh frozen features: Fresh v3 parity/repeatability evidence and five immutable clean feature caches.
- [x] Run three complete engineering pilots: Three valid full pilots, ineligible for learning-rate selection, with complete diagnostic evidence.

### Advancement gates

- **met:** The complete v3 design and cache-affecting release are frozen; source identities are preserved and historical seal audits are resolved.
- **met:** All three losses pass independent numerical and learnable-fixture checks; initialization and physical batch plans pair exactly.
- **met:** Fresh parity/repeatability checks and all three 20-epoch pilots pass with four immutable checkpoint/score records each.

### Risks

- The v3 protocol, shared seal checks and all cache-affecting code must be frozen before fresh extraction.
- A mathematical, nonfinite or collapse failure blocks selection; pilot accuracy cannot justify tuning recipes or duration.

**Review prompt:** Do the mathematics, immutable inputs and complete pilot records demonstrate a valid fixed-budget comparison?

## Week 5: Complete the execution path and preselection evidence

**Dates:** Oct 12-18  
**Phase:** Preselection go/no-go  
**Budget:** 5 hours planned, no time recorded  
**State:** active  
**Ready to close:** no

Complete one successor release, revalidate its caches/pilots and meet every scientific/resource prerequisite before selection.

**Deliverable:** A reviewed completion release, full fixture-to-report rehearsal, fresh bound admission evidence, nine verified development conditions and immutable go/no-go report.

### Tasks

- [x] Audit the fixed geometry panel and fallback: Immutable panel IDs/hashes, invariant/repeatability receipts and full-query damage denominators for all nine conditions.
- [ ] Revalidate the release and development paths: Fresh successor admission receipts, nine complete corruption paths, 27 head projections and measured stage-level runtime/memory/bytes.
- [x] Complete and rehearse the execution/report path: Reviewed successor release and regeneration matrix; complete fixture-to-v3-report receipts and fixed analysis/interpretation template before real revalidation.
- [ ] Record the October 18 go/no-go: Measured dependency, capacity and availability ledger plus immutable selection authorization or explicit blocker; no date-based waiver.

### Advancement gates

- **pending:** The fixed geometry panel and all nine full development conditions pass exact nesting, fallback, identity and fresh-process checks, with full-query damage telemetry and reusable head projections.
- **pending:** Completion-release scientific equivalence, full command/report rehearsal, fresh parity/five caches/three pilots, analysis/interpretation fixtures and shared seal audits all pass with successor bindings.
- **pending:** The measured dependency/schedule/storage/availability ledger passes, including revalidation and all remaining work with slack, before immutable successor selection authorization; otherwise no sweep starts.

### Risks

- Frozen final-training/extraction guards require a new release and fresh evidence; preserve original locks and Week 4 closure.
- Inherited five-hour tracker estimates are not a measured forecast: re-estimate implementation, revalidation, corruptions, reporting and resource planning before October 18.
- Current quota and remaining peak bytes must pass; dated capacity readings and prior GPU access are not reservations.

**Review prompt:** Do all October 18 validity, seal, storage, runtime and researcher-availability criteria pass without changing the declared study?

## Week 6: Launch the paired selection matrix

**Dates:** Oct 19-25  
**Phase:** Clean development selection  
**Budget:** 8 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Start the equal attempted budget of 27 fresh development runs after the October 18 authorization.

**Deliverable:** Immutable attempts for three arms, three learning rates and three seeds, with clean epoch-20 development metrics.

### Tasks

- [ ] Verify selection authorization: A verified immutable selection authorization and exact 27-run attempt matrix.
- [ ] Run the first selection allocations: Completed authorized selection attempts with paired hashes and epoch-20 checkpoints.
- [ ] Score clean development endpoints: Unrounded clean endpoint metrics with checkpoint and input provenance.
- [ ] Maintain the attempt and capacity ledger: A current immutable attempt ledger with all failures and current capacity checks.

### Advancement gates

- **pending:** All launched runs match the declared matrix and use paired seeds, initialization and physical batch plans.
- **pending:** Selection scores use only complete clean epoch-20 development evaluations; pilots and corrupted scores cannot select configurations.
- **pending:** Every attempt and failure is retained, and current allocation/memory/storage limits pass.

### Risks

- A candidate requires three valid seeds; partial-seed averages cannot compete.
- Infrastructure retries preserve identical settings in new attempt directories and consume forecast capacity.

**Review prompt:** Is every selection attempt authorized, paired, complete at 20 epochs and evaluated only on clean development data?

## Week 7: Complete selection and lock rates

**Dates:** Oct 26-Nov 1  
**Phase:** Clean development selection  
**Budget:** 8 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Finish all 27 declared selection attempts and lock one valid rate per arm using the exact ranking rule.

**Deliverable:** A complete selection ledger and timestamped three-recipe selection decision by November 1.

### Tasks

- [ ] Finish the declared selection matrix: A complete 27-setting attempt ledger with valid-run and failure status.
- [ ] Reproduce candidate ranking: An independently reproducible candidate table with tie and eligibility decisions.
- [ ] Lock the selected configurations: A timestamped selection lock for all three recipes, or an explicit blocker.
- [ ] Prepare all-auxiliary final training: Nine fresh final-run specifications and updated capacity evidence.

### Advancement gates

- **pending:** All 27 declared selection settings have an accounted attempt; every selected candidate has three valid seeds.
- **pending:** The three selected rates reproduce under the exact unrounded ranking rule and are immutably recorded before final training.
- **pending:** Final training inputs, paired plans and allocation/capacity checks pass for all nine fresh heads.

### Risks

- A numerical failure makes a candidate ineligible until resolved; each selected candidate needs all three valid seeds.
- An arm with no valid candidate requires a result-blind amendment, not an alternate recipe or reduced seed count.

**Review prompt:** Can an independent replay reproduce every selected rate from the complete unrounded clean development records?

## Week 8: Train the nine final heads

**Dates:** Nov 2-8  
**Phase:** Fresh final training  
**Budget:** 8 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Train the selected recipes from fresh paired initialization on all 100 auxiliary actions.

**Deliverable:** Authorized final training attempts toward nine 20-epoch heads with immutable provenance and epoch-20 weights.

### Tasks

- [ ] Validate all-auxiliary inputs: A verified final-input and configuration manifest.
- [ ] Train paired final recipes: Immutable final attempts with exact 59,380 updates per complete head.
- [ ] Record final health and cost: Per-head health, pairing, cost and provenance records.
- [ ] Review remaining final workload: An updated remaining-work forecast that preserves the final-lock deadline.

### Advancement gates

- **pending:** Every final attempt begins fresh and matches its selected rate, all-100-action inputs, paired seed and 20-epoch budget.
- **pending:** All completed heads have immutable epoch-20 checkpoints and valid health, pairing and provenance evidence.
- **pending:** Novel retrieval inputs and outcomes remain sealed while final training and any identical retries continue.

### Risks

- Development or pilot heads cannot be reused as final heads.
- Complete runs must fit approved allocations; partial attempts remain retained and cannot become resumable selection shortcuts.

**Review prompt:** Are the final heads fresh, paired and bound to the selected clean-development recipes and all-auxiliary data?

## Week 9: Complete and verify final heads

**Dates:** Nov 9-15  
**Phase:** Fresh final training  
**Budget:** 8 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Finish and independently verify all nine final heads by November 15.

**Deliverable:** Nine valid final checkpoints and complete final-run evidence ready for independent lock audit.

### Tasks

- [ ] Complete remaining final attempts: All nine complete 20-epoch final heads or a visible blocker.
- [ ] Audit checkpoint and manifest integrity: A complete nine-head integrity and provenance report.
- [ ] Build the final evaluation manifest: A fixed 180-cell evaluation matrix with exact ordered identities and role contracts.
- [ ] Confirm the audit schedule: A feasible final-audit and evaluation calendar with recorded capacity evidence.

### Advancement gates

- **pending:** All nine final heads are complete and valid at epoch 20, with no selected checkpoint or configuration dependent on novel outcomes.
- **pending:** The exact query/gallery identities and 180 unique evaluation cells are fixed and cross-bound to the complete final-run set.
- **pending:** The independent audit and evaluation schedule preserves November 29 locks, December 6 results and December 7-18 delivery work.

### Risks

- One missing or invalid required head blocks final opening.
- Corrections before opening require a dated result-blind decision and rerunning all affected comparisons.

**Review prompt:** Are all nine heads complete, correctly paired and independently verifiable without any novel-data access?

## Week 10: Audit the evaluator and final run set

**Dates:** Nov 16-22  
**Phase:** Independent audit and final locks  
**Budget:** 8 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Independently validate final provenance, authorization and result completeness using development evidence.

**Deliverable:** A reviewed final-run set, evaluator/analysis verification and report/figure templates prepared without novel outcomes.

### Tasks

- [ ] Audit the evaluator end to end: An independent evaluator and completeness-verifier report.
- [ ] Verify dataset-wide opening controls: Passing v1/v2/v3 seal and first-opening safety checks.
- [ ] Audit and assemble final locks: An independently checked complete final-run set and staged final-lock evidence.
- [ ] Independently verify the existing report templates: Reproducible report/figure templates ready for locked final results.

### Advancement gates

- **pending:** The evaluator, role-specific caches, metrics and 180-cell verifier pass independent development/fixture checks.
- **pending:** Shared seal enforcement and authorization-before-loading checks pass across all scientific entrypoints.
- **pending:** The final-run set and lock dependencies cross-bind all nine heads, exact identities and unchanged scientific settings.

### Risks

- Novel loader authorization must precede deserialization; shared dataset seals cannot be bypassed by legacy commands or namespaces.
- A changed protocol digest or cache-affecting defect invalidates affected artifacts and requires documented rework.

**Review prompt:** Can the complete final campaign be authorized and reproduced without bypassing seal, identity or provenance checks?

## Week 11: Finish locks and preserve recovery time

**Dates:** Nov 23-29  
**Phase:** Independent audit and final locks  
**Budget:** 4 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Complete all final-head and lock audits by November 29 while preserving Thanksgiving as recovery time.

**Deliverable:** A complete final authorization package and current capacity/seal evidence for the following week’s single novel campaign.

### Tasks

- [ ] Resolve pre-opening audit findings: An audit ledger with no unresolved final-opening blocker.
- [ ] Complete final locks by November 29: The complete immutable final lock and final-run-set record.
- [ ] Repeat seal and capacity audits: Current resolved seal and capacity reports for the authorized source.
- [ ] Preserve recovery and handoff readiness: A ready or explicitly blocked evaluation handoff with no invented completion.

### Advancement gates

- **pending:** All nine final heads and the complete final lock are verified by November 29; no scientific training is active.
- **pending:** Current historical-root, source, allocation and storage checks pass for the entire final evaluation campaign.
- **pending:** The novel test remains unopened until the single authorized campaign; unresolved requirements are visible blockers.

### Risks

- Reaching the date does not authorize opening; every final dependency and current resource check must pass.
- Thanksgiving is recovery capacity, not permission to compress or skip independent verification.

**Review prompt:** Are all nine heads and every opening prerequisite verified, with sufficient current capacity for the complete campaign?

## Week 12: Open once and evaluate all 180 cells

**Dates:** Nov 30-Dec 6  
**Phase:** Authorized novel evaluation  
**Budget:** 8 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Run one fully authorized novel extraction and evaluation campaign by December 6.

**Deliverable:** All 180 unique finite provenance-valid result cells and locked per-query ranks, with every operational attempt retained.

### Tasks

- [ ] Validate and atomically authorize opening: One immutable dataset-opening event bound to the complete v3 final lock.
- [ ] Extract the final novel conditions: Authorized deterministic caches for both query definitions with complete corruption and source provenance.
- [ ] Evaluate all heads and cells: 180 unique finite result cells and their checkpoint-bound per-query evidence.
- [ ] Verify campaign completeness: A complete final evaluation audit or an explicit incomplete/deviation record.

### Advancement gates

- **pending:** The immutable dataset opening precedes all novel loading and binds the complete final authorization.
- **pending:** All 180 required result cells are finite, unique, provenance-valid and reproducible from locked per-query evidence.
- **pending:** No scientific training/selection occurs after opening; all operational attempts and any deviations remain visible.

### Risks

- Opening is irreversible: a failed process does not reseal data or permit new training and selection.
- Missing or invalid cells block a complete report; post-opening corrections require explicit deviation reporting.

**Review prompt:** Does every required final cell reproduce from the same authorized source, opening, protocol, caches and checkpoint set?

## Week 13: Analyze and report the declared study

**Dates:** Dec 7-13  
**Phase:** Analysis and communication  
**Budget:** 8 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Produce the conditional primary analysis and complete descriptive report with accurate recipe-level claims.

**Deliverable:** Reproducible conditional interval, absolute accuracy/degradation curves, seed/cell/action sensitivities, resource tables and report/poster draft.

### Tasks

- [ ] Compute the locked primary analysis: A reproducible conditional primary estimate/interval and positive, null or contrary conclusion.
- [ ] Publish complete descriptive results: All prespecified tables and curves, including sign changes and conditional limits.
- [ ] Write the report and poster: A report/poster draft with claims traceable to locked evidence and honest scope limits.
- [ ] Regenerate figures and resolve review findings: A reviewed reproduction log and permitted compact final evidence package.

### Advancement gates

- **pending:** The conditional primary interval and all prespecified descriptive outputs regenerate from the locked records.
- **pending:** The prespecified interval rule and interpretation cases are followed; the Delta identity, absolute accuracy, full-recipe attribution and conditional/dependence limitations accompany every primary claim.
- **pending:** The report includes all methods, seeds, cells, costs and failures without outcome-driven omissions or new superiority tests.

### Risks

- Less degradation can coexist with lower absolute corrupted accuracy; both must be reported.
- The interval is conditional on fixed heads, seeds, actions, anchors and corruptions, and cannot support universal robustness or a contextual-term-only causal claim.

**Review prompt:** Do the claims follow the single prespecified interval rule while showing absolute performance, uncertainty limits and all required descriptive results?

## Week 14: Deliver and archive the December study

**Dates:** Dec 14-18  
**Phase:** Delivery and archive  
**Budget:** 4 hours planned, no time recorded  
**State:** planned  
**Ready to close:** no

Deliver the final report/poster and recoverable permitted evidence by December 18.

**Deliverable:** Final report/poster, independent reproduction smoke test, public-safe archive and reconciled tracker; if blocked, an explicit incomplete-study record.

### Tasks

- [ ] Run the independent reproduction smoke test: An independent reproduction record for the final deliverables.
- [ ] Finalize the report and poster: Final report/poster or an honest dated blocked-study report.
- [ ] Archive permitted material: A recoverable license-safe release and evidence index.
- [ ] Reconcile and export the tracker: A verified public tracker/export and researcher-supplied retrospective/time record where available.

### Advancement gates

- **pending:** The final report/poster and required outputs are delivered by December 18, or the incomplete study and dated amendment are explicitly recorded.
- **pending:** All nine final heads, 180 cells and declared analyses are independently evidenced before claiming empirical completion.
- **pending:** Source, permitted evidence and reproduction instructions are recoverable and exclude licensed/private material.
- **pending:** The live tracker/export matches v3 while preserving historical progress and actual hours without invented completions.

### Risks

- Licensed data, checkpoints, caches, full artifacts and private governance records must remain outside Git and public exports.
- A blocked or incomplete campaign cannot be presented as a completed empirical comparison.

**Review prompt:** Can a reader reproduce every permitted result and distinguish completed evidence, limitations and any unresolved work?

## Literature

- [NTU RGB+D 120 dataset and one-shot protocol](https://github.com/shahroudy/NTURGB-D/blob/master/README.md) (Liu et al., 2019): Canonical dataset access, action indexing, and one-shot split reference.
- [Show Your Work: Improved Reporting of Experimental Results](https://aclanthology.org/D19-1224/) (Dodge et al., 2019): Supports equal search-budget disclosure and reporting all development trials, eligibility and compute.
- [Supervised Contrastive Learning](https://papers.nips.cc/paper_files/paper/2020/hash/d89a66c7c80a29b1bdbab0f2a1a94af8-Abstract.html) (Khosla et al., 2020): Published source for supervised contrastive learning; v3 declares the single-view temperature-0.07 pose adaptation as a descriptive comparator.
- [Measuring Robustness to Natural Distribution Shifts in Image Classification](https://proceedings.neurips.cc/paper/2020/hash/d8330f857a17c53d217014ee776bfd50-Abstract.html) (Taori et al., 2020): Image evidence motivates limiting synthetic-stress claims; it does not establish pose robustness.
- [Accounting for Variance in Machine Learning Benchmarks](https://proceedings.mlsys.org/paper_files/paper/2021/file/0184b0cd3cfb185989f858a1d9f5c1eb-Paper.pdf) (Bouthillier et al., 2021): Motivates explicit sources of variability and conditional uncertainty; does not validate this exact bootstrap assumption.
- [Skeleton-DML: Deep Metric Learning for Skeleton-Based One-Shot Action Recognition](https://openaccess.thecvf.com/content/WACV2022/html/Memmesheimer_Skeleton-DML_Deep_Metric_Learning_for_Skeleton-Based_One-Shot_Action_Recognition_WACV_2022_paper.html) (Memmesheimer et al., 2022): Provides precedent for class-disjoint development validation within the NTU one-shot setup.
- [Supervised Metric Learning to Rank for Retrieval via Contextual Similarity Optimization](https://proceedings.mlr.press/v202/liao23b.html) (Liao et al., 2023): Defines the contextual objective, controlled metric-learning comparisons, and the limits of the original robustness evidence.
- [MotionBERT: A Unified Perspective on Learning Human Motion Representations](https://openaccess.thecvf.com/content/ICCV2023/html/Zhu_MotionBERT_A_Unified_Perspective_on_Learning_Human_Motion_Representations_ICCV_2023_paper.html) (Zhu et al., 2023): Specifies the pretrained pose encoder and its NTU RGB+D 120 one-shot evaluation context.
- [MaskCLR: Attention-Guided Contrastive Learning for Robust Action Representation Learning](https://openaccess.thecvf.com/content/CVPR2024/html/Abdelfattah_MaskCLR_Attention-Guided_Contrastive_Learning_for_Robust_Action_Representation_Learning_CVPR_2024_paper.html) (Abdelfattah et al., 2024): Grounds the use of Gaussian noise and joint or frame masking as pose-robustness stress tests.
- [Week 5 execution record](https://github.com/JJCAPPE/pose-embedding/blob/codex/pose-week5-execution/docs/protocol/week-5-execution.md) (Pose Embed research project, 2026): Public-safe progress, verified corrections, SCC status and remaining Week 5 admission requirements.
- [Robust Human-Motion Retrieval from Noisy Pose Sequences](https://github.com/JJCAPPE/pose-embedding) (Giacomo Cappelletto, 2026): Binding local prospectus and scope authority for the project.
- [Boston University semester dates](https://www.bu.edu/reg/calendars/semester/) (Boston University, 2026): Defines holiday, class-end, study-period, and final-exam constraints on the schedule.
- [Motion retrieval v2 method registry and protocol](https://github.com/JJCAPPE/pose-embedding/blob/main/configs/benchmark-methods.v2.json) (Pose Embed independent research, 2026): Historical 26-configuration v2 benchmark, preserved for provenance and future work. It is outside the December v3 study and does not authorize v3 execution.
- [Final v3 plan: frozen one-shot pose robustness](https://github.com/JJCAPPE/pose-embedding/blob/main/plan/research-plan.v3.md) (Pose Embed independent research, 2026): Final December execution design: three fixed recipes, 20 epochs, three seeds, nested corruption, conditional analysis and explicit capacity/test-opening gates; decision 0004 records the scope change.