Research lab book · measurement before interpretation

Question. Build. Question again.

This is the working record of an instrument being made: questions become constructions; constructions are attacked; failures become boundaries. The computational phase closes with no demonstrated differentiated biomechanical capability. The only remaining product hypothesis requires a prospective physical experiment.

COMPUTATIONAL PHASE CLOSED29 LINKED QUESTIONSPHYSICAL CLAIM NOT EARNED 2 EVIDENCE BLOCKERS552 / 560 TRIALS

Real corpus

552

included trials · 28 participants

Measurement table

357

auditable output columns per trial

Implementation verification

102 / 102

automated tests passing

Known-answer AUC

1.000

persistent plurality vs collapse

Orthogonal perturbation ΔAUC

−0.001

nonlinear nuisance · linear −0.018 · both fail

Structured-loss simulation

WITHIN CORPUS

AUC .912 · FPR 7.4% · no physical claim

01 · The instrument is the dialectic

Can organizational measurements reveal something consequential about the biomechanical system that produced them, beyond the complete incumbent read?

1 · QUESTION
one consequential unknown, named before the coordinate

→

2 · BUILD
the smallest instrument capable of answering it

→

3 · ATTACK
independence · null · resolution · ablation · incumbent

→

4 · CONSEQUENCE
held · killed · blocked · reframed

↺

What is now executable. config/research_program.yaml is not a roadmap paragraph. It is an audited graph of 29 questions. A build fails audit if it has no adversary, claims a nonexistent artifact, skips its operating envelope, or ends without a next question. Run it with ztoy-biomech research-audit.

Built / tested

mode competence now passes clean gate

Blocked honestly

independent evidence depth · synchronized EMG

Open frontier

route persistence · external/device transport · precedence · cross-domain transfer

The quiver is plural

Instrument	Question it answers	It must not be collapsed into	State
Evidence depth	How many independent families must agree before a demand site exists?	correlation or a composite confidence score	built / DVJ blocked
Persistence	What survives changing sensors, timing, scale, and assumptions?	strength of evidence	built
Dependency	What other structures rely on a node?	causal importance	built
Dynamic Ω / σ(Ω)	How much do independent reads disagree, and is disagreement stable or intermittent?	distance from normal	built / provenance blocked
Anatomical rotation	Do strong local anchors migrate while the whole landing remains globally unsettled?	peak timing or joint excursion	built / testing
Resolution-floor morphology	When does a coordinate become true, and how does truth arrive with resolution?	a one-time robustness check	built / unstable on DVJ
Structure-null protocol	Which destruction removes organization while preserving marginals?	label permutation	built / direction diagnostic
Reference basin	How far has this system moved from a declared memory, and along which axis?	constraint Ω	built
Reference Laboratory	Is viable movement one basin or several, and is this observation known, ambiguous, or outside all of them?	one universal normal or an outcome-trained class	built / DVJ one basin
Directional deviation field	Where, when, and by what anatomical route does the system leave its memory?	a scalar abnormality score	built / consequence open
Backward claim integrity	Can the headline be reconstructed from every untouched prediction and fitted object?	a passing test count or mean fold score	built / corrected headline
Direct consequence laboratory	Does the quiver reduce direct internal-load waveform residuals beyond the complete incumbent?	condition classification or tissue-stress language	evaluator only / no qualifying data
Orthogonal incremental value	What survives after incumbent-predictable outcome and instrument structure are removed?	ordinary augmented-model gain	built / DVJ gate fails
Registered shift abstention	Does the instrument refuse and localize degradation better than calibrated conventional comparators?	beating a classifier with no abstention	built / competitive gate fails
Plate-native two-family read	How do independently acquired marker and plate organizations disagree over landing phase?	anatomical Ω or a third witness	tested / reliability mixed
Frozen-device transport	Does a competent load estimator fail safely on a physically unseen device without adapting to it?	internal CV or arbitrary-shift coverage	evaluator only / no qualifying data
Acquisition decision laboratory	Which corpus is fit for which scientific job after access, license, sensing, target, and device gates?	one mythical best dataset	tested / staged path selected
Registered capture QA	Can structured hardware loss be detected and localized at a subject-held-out operating point?	arbitrary shift detection or a commercial device claim	single-corpus simulation only
Lag-free periodic plurality	Do IMU, EMG-envelope, and EEG-electrical families share recording-specific frequency organization?	pointwise synchronization, neural coupling, or one scale-free scalar	within-corpus relation / scalar gate fails
Frozen local-atlas transport	Does local relational morphology survive unseen people and an entirely held walking speed?	external apparatus transport or biological consequence	internal gate passes
External local-atlas transport	Does the minimal PhysioNet IMU–EMG atlas survive a new apparatus and protocol?	locomotion competence, mechanism, load, or clinical value	external gate killed

02 · Live questions, not a terminal hypothesis

After conditioning on the complete conventional kinematic, kinetic, force, coordination, and symmetry baseline, do relational measurements of disagreement and anatomical redistribution add held-out information about biomechanical constraint?

ID	Hypothesis	Required observation	Current state
H1	Independent observable families form a measurable phase×anatomy plurality.	Nonzero Ω; plurality persists across scale; source-overlap audit passes.	synthetic only
H2	Constraint is intermittency as well as magnitude.	σ(Ω), rolling sigma, and burst phase distinguish planted intermittent structure.	held on controls
H3	Added structure improves an incumbent-union model.	Positive held-out ΔAUC and ΔBrier; bootstrap excludes zero; subject-aware permutation p≤.05.	not held
H4	The result is not a proxy for sensor reuse or reference definition.	Provenance, every-read ablation, and block-ablation gates all pass.	not held on DVJ
H5	Anatomical localization is repeatable enough to interpret.	Condition-specific ICC/CV/SEM/MDC and stable phase/node attribution.	mixed

Question frontier

ID	Question now opened	Construction or next build	Evidence state
BQ-001	Do independent physical observables support the same demand sites, and at what epistemic depth?	`evidence_depth.py`	blocked until ≥3 independent synchronized families
BQ-002	Can a landing have strong local anatomical anchors without settling on one global anchor?	`rotation.py`	built; planted route controls pass
BQ-003	At what phase and anatomical resolution do Ω coordinates become stable enough to be true?	`resolution_floor.py`	built; DVJ floor is unstable
BQ-004	Which null destroys anatomical organization without changing marginal demand?	`null_protocol.py`	built; null direction diagnosed by arm
BQ-007	Does the morphology of the resolution curve identify the kind of organization being measured?	`resolution_morphology.py`	tested; controls, segments, and whole-subject source null pass
BQ-006	Is compensation a stable route through anatomy or merely unrelated peaks?	route survival + rotation	open
BQ-008	Does the instrument improve direct internal-load estimation rather than condition labels?	`consequence_lab.py`	evaluator built; planted control only; no qualifying data
BQ-009	Does anatomical rotation precede conventional performance loss or follow it?	fatigue / longitudinal precedence test	new data needed
BQ-010	Which structures are deep, persistent, and load-bearing—and which are merely visible?	joint Evidence–Persistence–Dependency read	open synthesis
BQ-011	Where else does a field collapse from many viable organizations to one consequential path?	new-domain transfer charter	open
BQ-012	Does viable movement occupy one normative center or multiple stable organizational basins?	`reference_lab.py`	built; DVJ supports one basin and refuses a taxonomy
BQ-013	When a trial leaves a viable basin, where, when, and toward what organization does it travel?	`deviation_field.py`	built; planted controls pass, biological consequence open
BQ-014	Can every headline be reconstructed backward through folds, memories, provenance, and raw contracts?	`evaluation.py`	built; fold-mean headline corrected to pooled OOF
BQ-015	Does internal-load improvement survive an external cohort and unseen device?	external device/site transport test	open; required before physical-load differentiation
BQ-016	Does the instrument retain information after removing every component predictable from the incumbent?	`orthogonal.py`	built; planted residual passes, both DVJ gates fail
BQ-017	Is independence a corpus limitation or a sensing-topology boundary?	`INDEPENDENCE_BOUNDARY.md`	reframed; kinematics-only can never clear plurality
BQ-018	Can refusal and localization beat identically calibrated conventional comparators under degradation?	`shift_lab.py`	built; current competitive gate fails
BQ-019	Can DVJ support an acquisition-independent two-family read without promoting it to anatomical evidence?	`source_native.py`	two-source descriptive read computed; reliability mixed; no independent-plurality claim
BQ-020	Does a competent direct-load model transport frozen to an unseen device and defer high-error trials?	`transport_lab.py`	evaluator and planted control only; no qualifying physical corpus
BQ-021	Must one corpus satisfy plurality, direct load, device transport, and commercial rights simultaneously?	`acquisition.py`	tested; no—staged falsification dominates the single-corpus bet
BQ-022	Can the audit detect and localize structured capture failure at a usable held-out operating point?	`capture_qa.py`	registered injected-fault result; no physical or transport evidence
BQ-023	Does the released PhysioNet corpus preserve enough synchronization to test plurality without manufacturing agreement?	`physionet_stage1.py`	recording-level admitted; pointwise phase and force closed
BQ-024	Does lag-free three-family periodic organization survive source, subject, clock, artifact, segment, and resolution attacks?	`physionet_plurality.py`	within-corpus separation measured; registered scalar resolution gate fails
BQ-025	Does scalar and cohort aggregation hide stable local organizations that reach the same total through different mechanisms?	`local_plurality_atlas.py`	internal topology measured; aggregate hiding demonstrated; taxonomy refused
BQ-026	Does the frozen local atlas survive simultaneous unseen-subject and held-speed transport beyond scalar and global-curve summaries?	`atlas_transport.py`	tested; internal gate passes, IMU–EMG strongest, external claim refused
BQ-027	Does the frozen PhysioNet IMU–EMG atlas retain recording identity under an external ENABL3S apparatus and protocol?	`enabl3s_transport.py`	killed; AUC .513, no hidden subject/route/frequency island, target-native AUC .493
BQ-028	Does local IMU–EMG organization add held-subject locomotion-mode information beyond the matched full incumbent and fail safely under channel loss?	`enabl3s_raw_mode_lab.py` + `enabl3s_fault_lab.py`	tested; confidence refinement narrow, classification and fault gates fail
BQ-029	Can physical capture faults be detected and localized after a complete unseen-subject/device/site freeze?	`stage3_capture.py`	admission and freeze package built; physical corpus not collected

Binding language. This instrument measures registered external demand patterns. It does not estimate tissue stress, internal joint contact force, injury risk, or causality without direct validation against those targets.

03 · The trail — what changed and what it asked next

The lab book preserves wrong turns because they define the current gates.

2026-08-19 · orientation
Started in the wrong generation. The duplicate Ztoy folder was found to be stale. Work was redirected to mature documents/russell/ZTOYBOX; existing instruments were reviewed read-only.

opportunity map
Proposed a product hypothesis: not another force or symmetry scalar, but audited relational information—where independent observations disagree, whether it persists, and whether it adds information beyond incumbent coordinates. The computational phase did not demonstrate that hypothesis.

pilot v0 · insufficient reuse
The first implementation was too thin. It used conceptual vocabulary—Ω, dependency, persistence—but not the mature 50-coordinate field machinery, multiscale collapse, dynamic σ(Ω), full read ablation, configuration locks, or validation storm. This criticism was accepted and became the expansion brief.

structural port
Ported phase×anatomy geometry, persistent regions, islands, spectra, multiscale curves, dynamic Ω/σ, bursts, lags, correlations, plurality, and every-read ablation. Added moving-block uncertainty to preserve local phase dependence.

biomechanical reconstruction
Replaced generic marker speed with bilateral landmark-derived laboratory sagittal angles and continuous relative phase. Added center-of-pressure side assignment and side-aware external moment-magnitude projections. These remain external-demand proxies.

corpus pass #1 · 551/560
One extra exclusion was traced to a missing heel marker—not a missing trial. The adapter now uses a declared ankle fallback and records fallback count. Final yield returned to 552/560; eight source files are genuinely absent.

falsification found an audit defect
Two observable families produced no valid leave-one-family-out test but were being labeled stable. The gate was changed: fewer than three fused families is now explicitly not ablation-auditable, never silently stable.

first validation turn found outcome encoding
REAL30-reference features initially drove apparent perturbation AUC from 0.524 to 0.862. Because the outcome was REAL30 versus VR, distance-to-REAL30 encodes the definition. The block was removed from the primary benchmark and retained only as a labeled diagnostic. The later fold-mean rerun was 0.558 honest versus 0.841 diagnostic; the backward audit below subsequently replaced both headline estimators with pooled OOF values.

pilot turn closed, program opened
The first thirty tests passed and the public DVJ corpus did not earn differentiation. That result was not the end. It asked what the scalar Ω was hiding, when its coordinates became valid, and how many evidence families actually supported each demand site.

question → build: evidence depth
Saltimbanques’ assumption-depth instrument was brought into biomechanics. Each phase×node site now records how many families support it, the depth ladder, origin family, and all-family dropout. On DVJ, the coordinate is repeatable but cannot be called independent evidence because the sources overlap.

question → build: anatomical rotation
The art-domain local-anchor/global-field rotation question became a biomechanical one: can every moment be locally anchored while the whole landing refuses one global anchor? The new route instrument records anchor migration, travel, reversals, coverage, entropy, and local-minus-global anchoring.

question → build: resolution floor
Materials asked “when does the coordinate become true?” A phase×anatomy resolution ladder now measures that floor and its curve morphology. On DVJ the floor size has poor repeatability, so it is a calibration warning, not a new feature to celebrate.

Pathology interrogation · reference defects become gates
The pathology studies contributed the frame–memory distinction, directional drift, and the finding that distributions can beat mean deviation. Their notebooks also exposed dangerous failures: empty references that still save, self-cohort memories, silent mean imputation, and ambiguous Ω naming. Each became an executable refusal in the biomechanical Reference Laboratory.

question → build: Reference Laboratory
Multiple covariance-aware viable basins, leave-one-subject-out envelopes, bootstrap center uncertainty, destination ambiguity, unknown-regime output, frozen-memory serialization, and separate frame/memory fingerprints are now implemented. Known-answer two-regime controls pass.

real memory refuses invented phenotypes
The 27-subject REAL30 memory supports one stable basin. Twenty-four of 28 subject-excluded discovery runs select one; four select two. That instability is not converted into a movement taxonomy. Under the corrected 95% finite-sample conformal rule, 2.0% are unknown and 2.5% exceed the observed calibration range.

nested memory validation
Every memory was rebuilt entirely inside each outer subject-training fold. Pooled OOF clinical AUC moves from 0.421 to 0.426 (Δ +0.005); the outcome-anchored perturbation diagnostic moves from 0.531 to 0.510. The memory adds no current differentiating evidence.

backward claim audit · fold means fail reconstruction
The published headline used the unweighted mean of fold AUCs. Recombining every untouched prediction produced different primary values: clinical full-instrument ΔAUC changed from −0.185 to −0.122; honest perturbation changed from +0.033 to +0.021. The conclusion remains no differentiation, but the estimator was wrong for the headline. Pooled OOF metrics are now primary, every row must appear once, and permutation nulls reuse fixed observed folds.

tail-resolution adversary
A 97.5% empirical cutoff with 27 subjects claimed precision the calibration cohort cannot supply. The minimum conformal tail is 1/(n+1). The memory now uses a resolvable 95% envelope; undersupported basins return unresolved_outside_calibration. Frozen schema v2 memories carry an integrity fingerprint and reject tampering.

question → build: direct consequence laboratory
The next instrument is now executable: nested subject-held-out direct-load waveforms, complete incumbent versus augmented, target-sensor provenance refusal, and RMSE/R² plus peak, impulse, and timing residuals. The planted control improves RMSE from 0.3080 to 0.0397 and refuses target-source contamination. No physical-load claim exists until a qualifying external/device corpus is run.

question → build: orthogonal incremental value
The ordinary +0.021 perturbation delta was challenged directly. Both the outcome and all 200 added coordinates were residualized against the 127-coordinate incumbent inside subject-held-out training folds. The delta becomes −0.018 with linear nuisances and −0.001 with nonlinear nuisances; Brier calibration worsens. The observed gains also fail dimension-matched noise controls. The planted residual-signal arm passes, so this is a measured failure of the current DVJ block—not an evaluator that cannot detect signal.

independence boundary · corpus question reframed
Kinematics-only sensing cannot clear independent plurality regardless of sample size. Force plates are physically independent at acquisition, but the present anatomical load field reuses marker-derived moment arms. Acquisition independence, registration dependence, and target dependence are now separate contracts. A third synchronized physical source is a design requirement, not a desirable extra.

question → build: registered shift abstention
All 106 available REAL30 repeats were attacked with phase jitter, shared marker dropout, node substitution, force gain, and plate loss. The audit envelope produced median shift AUC 0.550 versus 0.594 conventional and abstained on only 0.7% of corruptions. Its competitive gate fails. The perturbation classifier is not a useful foil: it is already wrong on 92.6% of clean REAL30 subjects, and its split-conformal wrapper defers 85.2% of clean cases.

failure split · instrument versus substrate
The shift result contains two distinct failures. The audit coordinates are insensitive to three of five registered attacks; that is an instrument problem. The clean perturbation model is already unusable; that is a target/corpus problem. Selective prediction cannot be judged as graceful degradation until a clean competence gate passes.

question → adversary → rebuild: strict two-witness
Different transducers were not enough: net GRF is coupled by Newton's equations to marker-derived COM acceleration. It was removed from the witness score. COP speed and unordered inter-plate sharing remain. Across 552 trials, median JS is 0.261–0.284 bits and JS ICC only 0.124–0.577. Information-nonredundant two-witness measurement survives; reliability and anatomical plurality do not.

question → build: frozen-device consequence + refusal
A whole-device protocol now excludes the target device from fitting, tuning, normalization, residual calibration, and OOD thresholds. It reports zero-shot waveform error, empirical coverage, selective risk, and error enrichment. Base-task competence is a binding pre-gate. The shifted-device control passes; no qualifying physical multi-device load corpus has been run.

acquisition prescription adversaried
Scientific direct-load truth and a sellable route are disjoint. PhysioNet remains Stage 1 plurality/reliability; CAMS-Knee remains a six-subject noncommercial scientific kill test; rights-clean multi-device capture QA is Stage 3. Neither Stage 1 nor Stage 2 tests unseen-device transport.

question → build → run: injected-fault operating point
The two encouraging fault rows were forced through nested subject-held-out threshold calibration. The audit linear model reaches AUC 0.912, 7.4% clean false alarms, 74.1% fault sensitivity, and 90.7% localization versus conventional linear AUC 0.691 and sensitivity 42.6%. This is a registered result on injected faults in one corpus, not physical capture QA, transport, or commercial validation.

2026-08-20 · PhysioNet admitted under a narrower contract
All 713 publisher checksums verify. The extracted corpus contains 177 EEG, 169 EMG, 170 IMU, and 173 force recording keys; 164 contain the IMU–EMG–EEG trio and 162 contain all four modalities. EEG CSV and EDF are counted once, five subject-summary rows are excluded, and DOB is flagged for minimization.

released-clock adversary · synchronized does not mean traceable
Only seven of 164 EEG files retain a visible trigger transition; EMG is a 16.05 Hz MAV envelope, not raw EMG; three S47 IMU exports are empty; four more recordings exceed the 0.5 s duration-agreement gate. The surviving state is 157 recordings / 54 subjects for recording- or cycle-aggregate three-family work. Pointwise phase plurality and force-target work remain closed.

question → build → attack: lag-free plurality
IMU, released EMG envelope, and common-average EEG were converted independently to normalized gait-band spectra; no lag was estimated. Across 156 recordings / 54 subjects, observed mean JS is 0.203 versus 0.265 after speed-matched source permutation (p=.002). Every sensor pair and the within-subject cross-speed attack pass. One corrupt EEG channel map is refused.

frozen resolution gate refuses the scalar
Segment, fixed ±0.5 s clock-trim, planted, and common-mode attacks pass, but median resolution sensitivity is 0.093 bits against the frozen 0.05 limit. The separation survives 2× and 4× coarsening, while magnitude falls from median 0.184 to 0.160 to 0.091 bits. The phenomenon is multiscale; a single calibrated magnitude is not earned.

failure → new coordinate: resolution morphology
The scalar was not retuned. Four nested rungs now measure the curve itself. Fine, hierarchical, and coarse controls order correctly; empirical curves are monotone; median segment curve-AUC range is 0.023 bits. Observed curve AUC is 0.091 versus 0.111 under source permutation, and the 49-subject block null also passes (both p=.002).

aggregate adversary · the scalar hides mechanism
Cohort-mean spectra retain only 21.8–36.4% of mean recording-level disagreement across four lower-band edges. There are 197 pairs within 0.002 scalar bits; 69 have local-shape L1 above 1.0. One global number can therefore be reached through sharply different frequency-family organizations.

question → build → attack: local plurality atlas
Thirty-four channels remain nested inside three physical families. Median topology contains six islands over 18.65 effective bins. Matched segment-shape L1 is 0.516 versus 0.943 for speed-matched donors; the 49-subject repeated-measures null is 0.512 versus 0.943 (both p=.002). The best cluster silhouette is only 0.182, so no phenotype taxonomy is manufactured.

question → freeze → transport → adversary
The atlas was tested with both subjects and one complete walking speed held out. It reaches AUC .929 versus scalar .701 and global curve .728; combined .934. The same-subject/different-speed attack reaches .970. Pairwise ablation finds the strongest internal read in IMU–EMG topology: .945 pooled and .986 hard-adversary AUC, so EEG is not required. External apparatus remained the next burden.

external apparatus transport · recording identity killed
Ten ENABL3S archives were verified against publisher byte counts and MD5 values. AB156 was quarantined after informing the adapter. On nine untouched subjects and 792 same-person/same-route pairs, the frozen atlas reaches AUC .513 [.494, .533] versus scalar .519. No subject, route, or frequency island survives; minimum max-|t| familywise p=.610. A target-native held-subject diagnostic also fails at AUC .493 [.478, .508]. A complete hostile 500 Hz clock rerun remains null (atlas .484). The internal result was an apparatus fingerprint, not a portable identity coordinate.

physiological substrate recovered
ENABL3S five-mode classification passes the clean gate on nine untouched subjects. Equal-subject IMU balanced accuracy is .868 with macro AUC .978 and worst-subject accuracy .742; conventional EMG raises accuracy to .904. The feature table's missing circuit identity blocks relational attachment, so raw-circuit reconstruction is now mandatory.

raw relational increment · aggregate split
Exact two-second crops restore circuit and row provenance for 1,497 windows. Against the matched full conventional union, the atlas moves balanced accuracy from .928 to .933, but the paired subject interval crosses zero [−.012, .022] and only two of nine subjects improve. Log loss improves from .236 to .191 in seven of nine subjects (p=.037), although one subject supplies 74.2% of positive magnitude. Confidence refinement is retained; broad discrimination is not.

registered raw faults · localization without detection moat
Five faults are injected before feature extraction and thresholds are nested inside training subjects. Audit AUC is .958 versus .951 conventional, clean FPR 2.2%, sensitivity 44.0%, and localization 93.1%. The atlas adds only .006 AUC and misses most lag and left/right-substitution faults. The gate fails; refusal helps gain and swap but harms or does not help the other mechanisms.

failure → physical freeze package
BQ-029 now fails closed before collection: seven physical fault families at two severities per device; raw, synchronization, and independent fault-log hashes; hardware/firmware/software and ethics/consent provenance; rights; clean pairing; subject/device/site separation. Two devices can ask only named A→B transport; three or more can rotate the held device. Admission can never award a performance or commercial claim.

current state · not a terminus
One hundred two tests pass; the measurement frame remains fee2924dbd60ca54, Reference Laboratory configuration is a34d559cae069fa3, and frozen library integrity is 32697b2a9102d84b. The 29-question program has eleven built instruments, nine tested constructions, one killed external target, one reframed boundary, two blockers, and five open questions.

04 · Biomechanical Reference Laboratory

Does viable movement occupy one normative center or multiple stable organizations—and when a trial leaves, where does it go?

FIXED FRAME
declared coordinates
fee2924dbd60ca54

→

INSERTED MEMORY
config a34d559cae069fa3
integrity 32697b2a9102d84b

→

DEVIATION
distance · direction · route · unknown

→

CONSEQUENCE
held-out external criterion

Usable reference subjects

physical 30 cm landings

Stable viable basins

no supported phenotype taxonomy

Conformal unknown

2.0%

95% cutoff · 2.5% outside observed range

Center uncertainty

0.284

mean bootstrap standardized units

Evaluation scope	Baseline AUC	With memory	ΔAUC	Interpretation
Clinical · memory fitted inside outer subject folds	0.421	0.426	+0.005	pooled OOF; internal test only
Physical vs VR · same leakage control	0.531	0.510	−0.021	pooled OOF diagnostic; REAL30 anchors outcome

What the memory refuses. Subject self-reference; fewer than six subjects per basin; unstable or structure-null-equivalent partitions; any missing predeclared axis; candidate-basin mean imputation; corrupted frozen memory; covariance collapse; phase-grid, anatomical-node, observable-family, or configuration drift; false tail precision; and forced assignment outside an undersupported envelope.

Full contract and outputs: docs/REFERENCE_LAB.md and outputs/reference_lab_observation.json.

04b · Direct Consequence Laboratory

After the complete incumbent predicts a direct internal-load waveform, does the organizational quiver reduce untouched-subject residual error?

DIRECT TARGET
implant · tissue · validated simulation

→

PROVENANCE GATE
target sensor excluded from predictors

→

NESTED HOLDOUT
inner tuning · outer subjects

→

WAVEFORM RESIDUAL
RMSE · R² · peak · impulse · timing

Planted incumbent RMSE

0.3080

pooled untouched-subject waveform

With planted field read

0.0397

ΔRMSE +0.2683

Planted ΔR²

+0.630

localized residual recovered

Physical corpus

qualifying datasets evaluated

The line is explicit. The planted result validates evaluator behavior, not physiology. A physical-load claim remains false until a preregistered instrumented cohort and, where applicable, unseen device/site transport pass without target-driven recalibration.

Executable contract: src/ztoy_biomech/consequence_lab.py · protocol: docs/CONSEQUENCE_LAB.md.

04c · Registered Shift & Abstention Laboratory

Can the instrument win on trustworthy refusal under degradation when it does not win on classification accuracy?

REGISTERED ATTACK
jitter · shared dropout · substitution · gain · plate loss

→

UNTOUCHED SUBJECT
27 subject-held-out memories

→

STRONG COMPARATORS
conventional envelope · split-conformal classifier

→

RISK + REFUSAL
detection · abstention · localization

Conventional shift AUC

0.594

median across five attacks

Audit shift AUC

0.550

median · delta −0.003

Audit shift abstention

0.7%

too low to be protective

Localization

61.5%

mean registered-family accuracy

Registered attack	Conventional AUC	Audit AUC	Audit Δ	Audit abstention	Localization
Phase jitter 2%	0.501	0.498	−0.003	0%	51.9%
Shared marker knee dropout	0.594	0.671	+0.077	0%	66.7%
Shared ankle/knee substitution	0.503	0.501	−0.003	0%	37.0%
Force gain 20%	0.605	0.550	−0.055	3.7%	59.3%
Right plate loss	0.748	0.797	+0.049	0%	92.6%

Headline refused. The trust-under-shift axis is not yet a demonstrated advantage. The perturbation classifier cannot rescue the story: it misclassifies 92.6% of clean REAL30 subjects, with 70.4% high-confidence errors. Its subject-calibrated split-conformal wrapper defers 85.2% of clean cases and produces no correct singleton decisions. That is a weak target, not safe selective prediction.

Do not bundle the failures. Shift blindness on phase jitter, node substitution, and force gain is an audit-sensitivity defect. The 92.6% clean error is a base-task/corpus defect. The first requires a sharper instrument; the second requires a learnable target. Neither is evidence that selective prediction is impossible.

Coverage language. Standard conformal coverage requires exchangeability. These registered corruptions are intentionally shifted from the clean calibration set. The results are empirical stress tests, not guaranteed coverage under arbitrary distribution shift. Weighted or nonexchangeable conformal methods require additional assumptions and data.

Protocol: docs/SHIFT_ABSTENTION.md · canonical outputs: outputs/shift_lab/.

04d · Registered capture-quality laboratory

Can the apparatus detect and localize structured capture failure at a usable operating point on an untouched subject?

REGISTERED FAULTS
knee-marker dropout · plate loss

→

NESTED SUBJECT CV
threshold learned inside training only

→

OPERATING POINT
≤10% training FPR

→

LOCALIZE
clean · marker · plate

Audit AUC

.912

linear · conventional .691

Clean FPR

7.4%

conventional 14.8%

Fault sensitivity

74.1%

conventional 42.6%

Localization

90.7%

conventional 77.8%

Feature family	Model	AUC	Clean FPR	Sensitivity	Localization
Audit	linear	0.912	7.4%	74.1%	90.7%
Audit	nonlinear	0.931	7.4%	70.4%	81.5%
Conventional	linear	0.691	14.8%	42.6%	77.8%
Conventional	nonlinear	0.888	14.8%	66.7%	85.2%

Registered within-corpus result. Thresholds are cross-fitted inside training subjects, and this injected-fault operating point improves false alarms, sensitivity, and localization over the matched linear comparator. It does not establish that the detector works on a physically degraded sensor.

Commercial claim refused. These are injected faults in one captured corpus. No physical fault recollection, unseen device, external site, or commercial-use prospective cohort has been tested. Stage 3 is designed to cross that boundary.

Protocol: docs/CAPTURE_QA.md · canonical outputs: outputs/capture_qa/.

04e · Plate-native strict two-witness laboratory

What remains when two reads must be independent in information—not merely measured by different transducers?

MARKER ONLY
mean motion · change · spread

↔

PLATE NATIVE
COP speed · unordered sharing · change

→

GRF RESULTANT
coupled diagnostic only

⊣

HARD STOP
no joint · tissue · plurality claim

Trials

552

28 subjects · 8 source exclusions

Nonredundant witnesses

marker organization + COP/sharing

Median JS range

.261–.284

bits across five conditions

JS ICC range

.124–.577

poor to mixed repeatability

Condition	Median JS bits	Median phase transport	Mean phase correlation	JS ICC(2,1)
REAL30	0.284	0.088	0.353	0.322
VR0	0.261	0.080	0.354	0.577
VR10	0.277	0.081	0.334	0.124
VR30	0.274	0.084	0.331	0.307
VR50	0.281	0.089	0.359	0.402

What survived. COP motion and unordered inter-plate load sharing provide a plate-native witness not determined by whole-body COM acceleration. Amplitude, planted-shift, shared-source, information-coupling, and leave-one-read-out adversaries are explicit.

What did not. Different sensor physics alone was insufficient; net GRF was a coupled echo and has been removed from the score. The strict read still cannot localize demand, supply a third family, or clear evidence-depth. Median leave-one-read-out JS range is 0.112 bits (p95 0.366); no dose coordinate survives Holm.

Protocol: docs/NATIVE_TWO_FAMILY.md · canonical outputs: outputs/native_two_family/.

04f · Frozen-device consequence + refusal

On a learnable direct-load task, can the instrument remain useful—or know it is not—when the physical device changes?

SOURCE DEVICES
subject OOF competence

→

FREEZE
fit · tune · normalize · calibrate

→

UNSEEN DEVICE
zero-shot waveform

→

SELECTIVE RISK
defer high-error trials

Clean competence

PRE-GATE

no refusal claim if source R² fails

Target adaptation

NONE

including normalization and calibration

Shift guarantee

empirical whole-device evidence only

Physical corpus

OPEN

known-structure control passes

The benchmark is now one question, not two roadmaps. Zero-shot internal-load error and audit-ranked refusal are scored on the same untouched device. A strong temporal model must be registered as the real incumbent; grouped ridge is only the executable reference control.

Protocol: docs/FROZEN_DEVICE_TRANSPORT.md · executable contract: src/ztoy_biomech/transport_lab.py.

04g · Acquisition prescription under attack

Should we buy the final corpus now—or force the instrument through cheaper, differently decisive corpora first?

Corpus	Subjects	Independent sources	Direct load	Commercial path	Decision
DVJ-CAI	28	2	No	CC BY	exhausted for current claims
PhysioNet multimodal gait	59	4	No	CC BY 4.0	Stage 1 · aggregate plurality admitted
CAMS-Knee v1.1	6	4+	implant forces + moments	scientific-only	Stage 2 · direct-load falsification
Knee Load Grand Challenge	6 reported	4	implant contact force	file terms unresolved	refused pending terms + provenance
Prospective study	to register	2 = named A→B; ≥3 = rotating holdout	physical fault labels; direct load optional	rights secured	Stage 3 · commercial capture QA

STAGE 1 · SCIENCE
Does plurality exist and repeat?

↗

STAGE 2 · LOAD
Separate noncommercial consequence test

↘

STAGE 3 · QA
Independent rights-clean product route

Checksums

713/713

publisher SHA-256 manifest

Triggered trio

164

57 subjects before clock gates

Admitted

157

54 subjects · aggregate/cycle scope

Pointwise phase

CLOSED

released trigger traceability insufficient

Assertive decision. PhysioNet Stage 1 proceeds, but at recording/cycle level with clock-shift and EEG-artifact attacks. Use CAMS-Knee strictly as a noncommercial scientific direct-load kill test. The only commercially testable hypothesis is a rights-clean multi-device physical capture-QA study.

Released-data boundary. A common trigger existed during acquisition, but the processed exports do not retain equivalent trigger evidence. Fine phase alignment may not be reconstructed by maximizing cross-modal agreement. Force was manually co-started and its released timestamp/sync columns are non-informative; force-target analysis remains closed.

Admission: docs/PHYSIONET_ADMISSION.md · outputs: outputs/acquisition/physionet_admission.json and outputs/physionet/clock_report.json.

04h · PhysioNet lag-free plurality laboratory

Do independently acquired IMU, EMG-envelope, and EEG-electrical families share recording-specific periodic organization without manufacturing agreement through alignment?

SOURCE-NATIVE
IMU 60 Hz · EMG MAV 16.05 Hz · EEG 300 Hz

→

LAG-FREE
normalized Welch spectra · 0.5–6 Hz

→

THREE WITNESSES
JS · pairwise ablation · depth sweep

→

ATTACK
source · subject · segment · clock · artifact · scale

Measured recordings

156

54 subjects · one corrupt EEG map refused

Observed mean JS

.203

permuted .265 · Δ −.062 bits

Source permutation

p=.002

all three pairwise tests also pass

Frozen headline gate

FAIL

resolution Δ .093 > .05 bits

Nested resolution morphologyobserved recording match vs speed-matched EEG source permutation

The 45× rung is the registered one-bin collapse and is zero by construction. At the three informative rungs, observed-versus-source-permuted separation survives (each one-sided p=.002). The failed scalar-invariance gate is retained.

Adversary	Observation	Gate
Planted frequency + amplitude	Shifted JS .875 vs aligned .003; amplitude delta 1.1×10⁻¹⁶	pass
All pairwise speed-matched permutations	IMU–EMG Δ −.122; IMU–EEG −.086; EMG–EEG −.048 bits; each p=.002	pass
Within-subject cross-speed EEG	Observed .203 vs .331 bits; Δ −.128; p=.002	pass
Three independent temporal segments	Median JS range .050 bits against ≤.10	pass
Fixed ±0.5 s family trims	Median maximum delta .002 bits against ≤.05	pass
EEG common-mode comparator	Median CAR minus common-mode JS −.021 bits	pass
2× / 4× frequency coarsening	Median maximum delta .093 bits against ≤.05	fail

What is now supported. Three separately acquired families contain recording-specific shared periodic organization beyond speed-matched source substitution and subject/session identity. Separation is present at every registered scale, while the scalar resolution gate fails.

What remains refused. A scale-invariant scalar; pointwise phase; brain or muscle mechanism; anatomy; tissue stress; internal load; and clinical value. Common-average referencing is an artifact adversary, not proof that gait-locked electrical structure is neural.

Resolution morphology earned. The failed gate changed the object. The four-rung curve passes known fine/hierarchical/coarse controls, has no nested-monotonicity violations, and is stable across three segments (median curve-AUC range .023 bits). Mean curve AUC is .091 observed versus .111 permuted (p=.002); a whole-subject donor permutation across 49 complete subjects gives the same −.020-bit separation (p=.002). Median half-collapse occurs at 9×.

Next burden. Transport the frozen curve to an external task or apparatus and ask whether its morphology carries consequence. The original single-resolution invariance gate remains failed; morphology does not rescue a scale-free scalar.

Protocol: docs/PHYSIONET_PLURALITY.md · canonical outputs: outputs/physionet_plurality/.

04i · Channel-aware local plurality atlas

When global totals look alike, are they governed by the same local organization—or does aggregation hide different frequency islands, family ownership, and channel structure?

3 FAMILIES
IMU · EMG envelope · EEG electrical

→

34 NESTED CHANNELS
substructure, never extra witnesses

→

LOCAL ATLAS
frequency density · ownership · islands

→

REFUSALS
band edge · donor segments · taxonomy

Median local islands

q75 support · 18.65 effective bins

Matched segment L1

.516

donor .943 · p=.002

Same-scalar pairs

197

69 have shape L1 >1

Taxonomy

REFUSED

best silhouette .182

How much disagreement survives cohort aggregation?family spectra renormalized symmetrically inside each band

The aggregate retains 21.8–36.4% of mean recording-level disagreement. Moving the lower edge does not remove the hiding effect.

Layer or adversary	Observation	Decision
Same mass, planted topology	One contiguous island versus three separated islands at identical aggregate mass	control passes
Recording segments	Mean L1 .516 matched versus .943 speed-matched donors; p=.002	stable local state
Whole-subject donor null	49 complete subjects: .512 matched versus .943 donors; p=.002	repeats respected
Lower edge .5–1.25 Hz	Within-recording median L1 .494–.519 versus .948–.987 between recordings	topology persists
Channel heterogeneity	Median within-family JS: IMU .658 · EMG .215 · EEG .173	deeper aggregate layer
Boundary peak	111/156 recordings peak at .5 Hz	calibration warning
Unsupervised types	Best k=2–6 silhouette .182	no phenotype classes

Internal measurement result. The global scalar and cohort mean hide local differences within this corpus. Local frequency-family topology is internally stable and distinguishes many recordings whose global JS is essentially identical; later external transport fails.

Mechanism still refused. Stable recording-state topology is not a subject phenotype, neural signature, anatomical route, internal load, or clinical value. Channel differences remain nested within their physical family.

Next step preserved and sharpened. Freeze the four-rung resolution curve and the local atlas, then transport both unchanged. Transporting only the aggregate would validate the hider rather than the instrument.

Protocol: docs/LOCAL_PLURALITY_ATLAS.md · canonical outputs: outputs/physionet_local_atlas/.

04j · Frozen local-atlas internal transport

If the aggregate is the hider, does the local organization still identify a recording when both the person and the walking speed are unseen?

TRAIN
other subjects · other speeds

→

FREEZE
scaling · coefficients · thresholds

→

TEST
unseen subject + held speed

→

ATTACK
same subject · other speed · channel pairs

Scalar AUC

.701

aggregate local JS mass

Global curve AUC

.728

four resolution rungs

Local atlas AUC

.929

held person + held speed

IMU–EMG atlas

.945

strongest pair · EEG excluded

What survives the double holdout?same-recording segment matching · pooled out-of-fold AUC

The combined model reaches .934 but does not beat the atlas alone; accumulation is not the result. IMU–EMG local morphology is the strongest read.

Attack	Observation	Decision
Zero target-speed fitting	Each 0.5, 0.75, or 1.0 m/s test excludes that complete speed and every test subject from fitting	pass
Held-speed floor	Combined AUC .933, .970, .933 across the three speeds	pass
Scalar incumbent	Combined ΔAUC .233; left-subject-block bootstrap 95% CI .195–.277	local structure retained
Global-curve incumbent	Combined ΔAUC .205; 95% CI .172–.243	locality matters
Same subject, another speed	Atlas AUC .970; IMU–EMG AUC .986	not explained by person identity
Pairwise family ablation	IMU–EMG .945; IMU–EEG .875; EMG–EEG .875	EEG not required
External apparatus	All recordings originate in one PhysioNet apparatus and task	not tested

Internal result only. Inside this apparatus, frequency-local IMU–EMG relational morphology retains recording-specific structure across unseen subjects and a held walking speed that scalar and global-curve summaries discard. The subsequent ENABL3S test shows this is not externally portable.

What is not earned. This can still reflect task-, session-, or apparatus-specific spectral consistency. It is not a muscle mechanism, neurological coupling, anatomical constraint, internal load, phenotype, or clinical value result.

Subsequent decision. The frozen IMU–EMG atlas was moved unchanged to ENABL3S and collapsed to chance. The internal result is retained as apparatus-specific recording discrimination, not transport.

Protocol: docs/ATLAS_TRANSPORT.md · canonical outputs: outputs/physionet_atlas_transport/ · contract bbd73c72ff2394e9.

04k · Frozen external local-atlas transport

Does the PhysioNet IMU–EMG atlas survive when the apparatus, protocol, participants, EMG representation, and recording structure all change?

SOURCE
PhysioNet only · AUC .909

→

FREEZE
channels · grid · model

→

QUARANTINE
AB156 adapter development

→

TEST
9 untouched ENABL3S subjects

Verified archives

10 / 10

publisher bytes + MD5

Evaluation pairs

792

same person · same route parity

Frozen atlas AUC

.513

95% CI .494–.533

Target-native AUC

.493

held-subject diagnostic

Attack	Observation	Decision
Development leakage	AB156 informed the 2.5-second crop contract and is excluded from the gate	quarantined
Scalar comparator	Scalar AUC .519 [.508, .530]; atlas-minus-scalar −.006 [−.034, .020]	no local increment
Subject decomposition	Atlas AUCs .451–.569 across nine untouched people	no stable subgroup
Route mechanism	Even .533; odd .489	no repeatable route split
Frequency islands	12 bin AUCs .487–.520; minimum max-\|t\| familywise p=.610	aggregate not hiding a local win
Target-native substrate	12-bin ENABL3S-trained model, held subjects: AUC .493 [.478, .508]	recording identity unusable
Multirate clock ambiguity	Hostile 500 Hz rerun: atlas .484 [.459, .508], scalar .499, target-native .479	null invariant to clock contract

BQ-027 is killed. The PhysioNet atlas is a strong internal recording fingerprint, but it does not transport to ENABL3S. The null is not a hidden-island cancellation, and a target-fitted held-subject diagnostic finds no usable recording-identity substrate either.

Dialectical consequence. Stop tuning identity. Use ENABL3S for its physiological labels: locomotion modes and transitions. The next build compares an IMU-only incumbent with local IMU–EMG organization under subject holdout, then attacks both with registered channel loss and confident-error/refusal tests.

Protocol: docs/ENABL3S_EXTERNAL_TRANSPORT.md · canonical outputs: outputs/enabl3s_atlas_transport/ · contract 9fa77dc21ea54719.

04l · ENABL3S physiological competence gate

Before asking whether an instrument refuses degradation gracefully, is there a clean held-subject task on which it is genuinely competent?

Evaluation subjects

AB156 development-only

Stable windows

15,410

five locomotion modes

IMU balanced accuracy

.868

equal-subject · worst .742

IMU + EMG

.904

macro AUC .992

Representation	Balanced accuracy	Macro AUC	Worst subject
Published IMU features	.868	.978	.742
IMU + conventional EMG	.904	.992	.776
IMU + goniometer + EMG	.965	.999	.899

Clean competence passes. Unlike DVJ perturbation and ENABL3S recording identity, this target supplies a working held-subject baseline from which degradation can be measured.

Backward refusal. The publisher feature table lacks circuit identity. It may establish subject-held competence, but it cannot receive a guessed raw-waveform atlas or support circuit-block uncertainty. Relational and channel-loss arms must be rebuilt from circuit CSVs with archive/member/row provenance.

Protocol: docs/ENABL3S_MODE_LAB.md · canonical outputs: outputs/enabl3s_mode_lab/.

04m · Circuit-traceable mode and raw-fault laboratory

When the aggregate moves, is it broad information, confidence refinement, or one subject and one fault mechanism carrying the result?

Exact raw windows

1,497

1,323 held-subject · two seconds

Matched accuracy Δ

+.005

95% CI −.012 to .022

Matched log-loss Δ

−.045

7/9 improve · p=.037

Fault AUC increment

+.006

audit .958 · conventional .951

Clean relational increment

Representation	Balanced accuracy	Log loss	Adversarial reading
IMU only	.920	.257	strongest sparse sensor read
IMU + EMG	.892	.303	EMG harms this fixed model
IMU + EMG + atlas	.913	.248	recovers damage; does not beat IMU
Full conventional	.928	.236	matched primary incumbent
Full + atlas	.933	.191	confidence improves; accuracy CI crosses zero

The aggregate was hiding a different mechanism. It was not hiding broad accuracy. Only two subjects improve on balanced accuracy. Log loss improves in seven, but 74.2% of positive magnitude comes from one subject. The atlas currently adjusts confidence more credibly than decisions.

Fault-by-fault detection: nonlinear audit arm

Registered raw fault	Flag rate	Localization	Interpretation
Waist IMU dropout	1.000	1.000	hard loss found
Right TA electrode dropout	1.000	1.000	hard loss found
EMG bank gain ×2	.638	.967	partial detection
Left/right IMU substitution	.241	.819	mostly missed
EMG lag 100 ms	.186	.675	mostly missed

BQ-028 decision: substrate yes; differentiator no

The competent five-mode task survives, and the relational read shows a bounded confidence-refinement signal. But it does not earn broad incremental discrimination, and the audit detector does not separate materially from the conventional detector. Strong localization is conditional on finding the fault. Further ENABL3S tuning would optimize known simulations; the evidentiary move is rights-clean physical faults. Two devices support one named A→B transfer; three or more permit rotating held-device tests.

Protocol: docs/ENABL3S_RAW_MODE_AND_FAULT_LAB.md · clean artifacts: outputs/enabl3s_raw_mode_lab/ · fault artifacts: outputs/enabl3s_fault_lab/ · contract 3b974e5dbcd79e44.

04n · Stage-3 physical capture freeze

What must be frozen before physical data exist so that a friendly dropout demonstration cannot masquerade as device-general capture QA?

Physical fault families

each device · ≥2 severities

Two devices

A→B

named transport only

Rotating holdout

≥3

still not device-general proof

Claims earned by admission

evaluation permission only

Admission surface	Executable requirement	Overclaim prevented
Fault truth	Physical realization plus independently hashed fault log	self-declared or feature-masked “physical” fault
Synchronization	event ID, timestamp source, and hashed sync log	outcome-optimized alignment
Device identity	manufacturer, model, serial hash, hardware, firmware, software	different IDs for the same acquisition surface
Fault coverage	dropout, intermittency, gain, substitution, offset, drift, attachment/occlusion	easy-dropout benchmark sold as general QA
Transport unit	subject-disjoint devices; site scope tracked separately	pooled trials or subjects counted as device replication
Rights and governance	commercial development/validation, derivatives, audit retention, ethics and consent versions	scientific access converted into a product asset

Safety boundary. Faults may not alter participant-facing control or safety-critical feedback. On assistive or robotic platforms, inject them into a shadow acquisition path unless a separate safety protocol explicitly authorizes otherwise.

Study-size boundary. Three subjects per device is only the executable minimum. The planning target begins at four platforms across at least three sites and 20 subject-disjoint participants per platform, then increases through clustered precision simulation until clean-refusal and fault-sensitivity confidence bounds can resolve their gates.

What is ready. The frozen YAML protocol, empty manifest template, checksum/provenance auditor, CLI, and known-answer attacks are built. What remains external is SOP rehearsal, precision simulation from rehearsal variance, ethics/rights execution, hardware/site selection, registration, and physical collection.

Protocol: docs/STAGE3_PHYSICAL_CAPTURE_PROTOCOL.md · freeze: config/stage3_capture_protocol.yaml · manifest: templates/STAGE3_CAPTURE_MANIFEST_TEMPLATE.csv · YAML fingerprint 6579c6b6bba41966 · executable contract 0cdccb54993650cb.

04o · Computational phase synthesis — paper wrap

After every backward audit and external adversary, what actually survives—and what experiment can still change the claim?

Tier earned

deterministic measurement

Domain result

NO WIN

no differentiated biomechanical capability

Remaining hypothesis

CAPTURE QA

physical performance untested

Next evidence unit

DEVICE

not pooled trials

Headline result

No differentiated biomechanical capability was demonstrated. The software deterministically measures, reconstructs, plants, attacks, and refuses. DVJ incremental discrimination, dose, evidence depth, and broad shift claims fail or remain blocked. PhysioNet local organization survives internal transport but dies on the external ENABL3S apparatus. ENABL3S is a competent substrate, yet relational accuracy and broad fault advantages fail; confidence refinement is narrow and magnitude-concentrated. Physical capture QA is the only remaining product hypothesis, and it has not been tested physically.

Finding	Status	What may be said
Measurement, provenance, backward reconstruction, planted controls	survives	Tier-1 apparatus is functioning and auditable
ENABL3S log-loss refinement	bounded	0.045 mean improvement; interval 0.004–0.114; 7/9 subjects, but 74.2% of positive magnitude from one subject
Conditional fault localization	bounded	93.1% among faults the detector finds; not unconditional detection
DVJ structured-loss simulation	within corpus	AUC .912 vs .691; clean FPR 7.4% vs 14.8%; sensitivity 74.1% vs 42.6%; no physical QA established
DVJ clinical/perturbation increment, dose, stable coordinates	not earned	No classifier, mechanism, or dose claim
External atlas transport	killed here	PhysioNet-to-ENABL3S AUC .513; internal identity did not transport
Broad ENABL3S fault advantage	fails	AUC advantage .006 and sensitivity 44.0%; gain, lag, and substitution remain weak
Physical/device-general/clinical/commercial performance	untested	No statement is authorized

The physical experiment, in one page

Layer	Requirement before the claim can move
Governance	Ethics, versioned consent, commercial development/validation and derivative rights, retention, sponsor/data-controller, site and hardware agreements
Topology	Two platforms for one named A→B test; preferably four across ≥3 sites for an initial rotating architecture; more device units for device-general inference
Participants	Subject-disjoint device cohorts, representative capture difficulty, ≥2 clean repeats per subject-task; final n set by clustered rehearsal precision, not “20” by convention
Tasks	Standardized walking, sit/stand, step-up/down, and a safety-approved higher-transient task with scripted events and recovery
Fault matrix	All seven physical families on every device, ≥2 calibrated severities: dropout, intermittent loss, gain, substitution, offset, drift, attachment/occlusion
Truth	Independent reference clock/packet/calibration logger, second-operator record, randomization, blinding, and SHA-256 raw/fault/sync evidence
Freeze	Features, model, normalization, thresholds, taxonomy, tasks, severity, estimands, multiplicity, missingness, sample size, software and target-data embargo
Primary gates	Clean false refusal ≤10%; sensitivity ≥70%; localization ≥70%; audit AUC advantage ≥.05; task competence first; worst fault and worst device headline
Inference	Subject clusters for named transport; devices as units for multi-device claims; site separate; no pooled-trial pseudo-replication
Kill rule	If the frozen audit ties the complete incumbent, remains blind to informational faults, or wins only on one easy dropout/device/operator, close the current product wedge

Readiness verdict. The study is computationally prepared but not collection-ready. The immediate external milestone is governance plus a bench rehearsal on one development platform. Participant rehearsal then supplies cluster variance and score correlation for final precision planning. Only after that freeze should untouched target devices be opened.

Operational handoff built. In addition to the executable capture manifest, three empty prospective templates now bind the platform/site register, physical fault and safety SOP matrix, and development-rehearsal metrics needed for clustered precision. They do not invent sample size; they ensure the rehearsal captures what the final calculation requires.

Device-count boundary. Four platforms are an operational starting architecture, not a magic sample size. They enable rotating held-device attacks but may still be too few for a stable device-population interval. If device-level precision is inadequate, acquire more devices or narrow the wording to named transports.

Full synthesis: docs/COMPUTATIONAL_PHASE_SYNTHESIS.md · physical design and resources: docs/PHYSICAL_EXPERIMENT_REQUIREMENTS.md.

05 · What the instrument quiver reads

L0Registration201 phase points × seven anatomical nodes; disclosed 33-point anatomical regularization for topology.

L1ConventionalForce, impulse, loading rate, bilateral angles, velocity, asymmetry, relative phase, waveform summaries.

L2RelationalΩ mean/sigma, burst, plurality, lag, correlation, source overlap, every-family ablation.

L3StructuralGeometry, void/cohesion, persistent regions, islands, contour, gradient, spectral bands.

L4ReferenceWithin-person REAL30 envelope, node/phase exceedance, L1 and Jensen–Shannon distance.

L5ValidationReliability, dose, block ablation, held-out models, bootstrap, permutations.

L6EpistemicEvidence depth, family origin, depth ladder, source overlap, all-read loss.

L7DialecticalAnatomical rotation, structure nulls, resolution-floor curve, next-question audit.

L8Frame + MemorySupported viable basins, covariance-aware direction, uncertainty, destination, unknown regime, phase×anatomy deviation route.

L9Orthogonal valueCross-fitted outcome/instrument residuals, nuisance-family agreement, duplicate and matched-noise controls.

Markers
bilateral landmarks

→

Joint motion field
|angular velocity|

↘

Dynamic Ω
centroid · spread · entropy

→

Audit gates
plurality · provenance · ablation

Force plates + COP
side assignment

→

External load field
|r × F| / body scale

↗

Field kernel
34 coords per field + aggregate

→

Held-out tests
incumbent vs augmented

Why the real Ω gate fails. External-load moment arms use the same marker coordinates as joint kinematics. This source overlap means the two fields are not independent. There are also only two fused families, so a complete leave-one-family-out audit is impossible.

06 · What came from existing ZTOYBOX instruments

Source	Mature mechanism reused	Biomechanical adaptation
`Material/vtl_materials_kernel.py`	Weighted field geometry, anisotropy, void/cohesion, persistent regions, islands, contour/gradient, spectra.	Applied to phase×anatomical demand fields; 34 coordinates per family and aggregate.
`Saltimbanques/.../persistence.py`	Plurality, multiscale survival, collapse, separate Ω and σ(Ω).	Phase/anatomy scale ladder with “never collapses” retained as an explicit state.
`persistence_regions.py`	h-maxima region survival and monotonic count audit.	Persistent demand regions, anchor emergence, collapse depth, count curves.
`parallax_portrait/omega.py`	Robust diagonal basin, axis attribution, configuration refusal.	Subject-mean median/MAD basin plus hard config-fingerprint mismatch error.
`Markets/starter_independence_audit.py`	Correlations, Ω curve, rolling sigma, lag, all-read sensitivity.	Centroid/spread/entropy reads at every phase; pairwise and aggregate curves.
`seismology/bootstrap_stage1.py`	Block bootstrap respecting autocorrelation.	Lag-one phase autocorrelation sets bounded block length for Ω confidence intervals.
Material validation briefs	Repeat aggregation, incumbent union, nonlinear residual, grouped validation.	Subject×condition aggregation; 127 incumbent vs 200 instrument coordinates; complete subjects held out.
Pathology frame–memory and drift studies	Invariant measurement frame, replaceable contextual memory, directional failure, distribution over mean.	Multiple supported basins, separate fingerprints, unknown-regime refusal, directional phase×anatomy routes, and memory fitting inside outer folds.
Pathology delta logic + backward audit	Normative expectation, directional residual, and reconstruction from untouched predictions.	Cross-fitted removal of incumbent-predictable outcome and instrument structure before an added-information claim.

Full implementation mapping: docs/ZTOYBOX_LINEAGE.md.

07 · Corpus, provenance, and accounting

Trial accounting by conditionincluded n=552

Expected: 112 trials per condition. REAL30 has six missing source trials; VR0 and VR50 have one each. Every manifest row is retained in the accounting table.

Registered archive

Zenodo 18503500

CC BY 4.0
MD5 a3b7e2c0ed48d727f434dfa8bfae380d
3,097 files · 3.95 GB expanded

Cohort

16 / 12

control / chronic ankle instability

Integrity item	Observation	Action
Missing source files	8 `.mot`/`.trc` paths absent	Excluded by explicit reason; not imputed.
Marker alternatives	30 trials use ≥1 declared fallback; 6 use two	Fallback count included in measurement table.
EMG	Present for a subset, but no sample rate, time, or sync event	Not phase-fused; blocker stated in summary.
Configuration drift	Settings could otherwise change under same metric names	Every row stamped `fee2924dbd60ca54`; basin mismatch raises.
Runtime	101.25 s for manifest; 0.181 s/trial	Recorded after adding evidence, rotation, and resolution instruments.

08 · Known-answer falsification arms

Ω magnitude and temporal dispersion100 repetitions / arm

Ω meanσ(Ω)

The dependent-source arm is intentionally identical in geometry to persistent plurality. It is rejected only by provenance—demonstrating why structural signal alone is insufficient.

Plurality vs collapse

AUC 1.0

complete separation

Dependent source refusal

100%

despite high, persistent Ω

Burst sigma success

100%

burst σ exceeded collapse

Arm definitions

Collapsed: independent reads share one anatomical trajectory. Persistent plurality: nine independent redundant reads occupy three separated anatomical trajectories and survive scale/ablation. Dependent plurality: identical geometry but all reads declare a shared source. Intermittent burst: one read departs only near phase 0.58.

09 · Industry comparison: did it add information?

Subject-held-out AUC: incumbent versus augmentedtechnical repeats aggregated first

127-feature incumbentaugmented primaryoutcome-anchored diagnostic

Pooled predictions from every untouched row are primary; fold means are diagnostic. Complete subjects are held out. For perturbation, REAL30-reference coordinates are excluded from the primary model because they encode the outcome definition. The violet diagnostic shows the misleading result if they are included.

Clinical block additionsΔAUC vs incumbent

Perturbation block additionsdiagnostic reference shown

Outcome / model	Baseline AUC	Augmented AUC	ΔAUC	95% bootstrap CI	Permutation
CAI · logistic	0.421	0.298	−0.122	[−0.272, 0.019]	p=.925
CAI · nonlinear	0.549	0.517	−0.032	[−0.153, 0.073]	—
Physical vs VR · logistic, honest primary	0.531	0.552	+0.021	[−0.073, 0.110]	p=.323
Physical vs VR · nonlinear, honest primary	0.592	0.566	−0.027	[−0.139, 0.080]	—
Physical vs VR · reference-inclusive diagnostic	0.531	0.821	+0.290	[0.141, 0.428]	not eligible

Orthogonal incremental-value attack

The ordinary comparison asks whether refitting with a wide new block helps. The stronger test removes everything in the outcome and added block predictable from the incumbent, learns only from residuals inside training subjects, and tests untouched subjects once.

Outcome / nuisance	Baseline AUC	Orthogonal AUC	ΔAUC	ΔBrier	Clipped
CAI · linear	0.421	0.381	−0.040	−0.084	65.5%
CAI · nonlinear	0.549	0.381	−0.167	−0.128	28.8%
Physical vs VR · linear	0.531	0.513	−0.018	−0.144	38.1%
Physical vs VR · nonlinear	0.592	0.592	−0.001	−0.008	24.5%

The current gate is not cleared. The observed linear deltas do not beat matched-noise maxima (+0.133 clinical, +0.060 perturbation); calibration worsens; nuisance families do not yield a positive residual gain. Exact-duplicate incumbent deltas stay inside the declared ±0.05 sanity band. The planted control does pass: AUC 0.635 → 0.989, Δ +0.353, bootstrap lower bound +0.207, maximum matched-noise delta +0.001.

Decision at this turn: no residual information demonstrated on this target and corpus

The current DVJ quiver measures additional organizational coordinates, but this 28-subject study has not demonstrated outcome information orthogonal to the incumbent. The result does not establish absence of information in another target, sensing topology, or adequately powered cohort. The ordinary +0.021 perturbation result is compatible with redundant re-expression, finite-sample refitting opportunity, or unresolved small-sample variation. The program now carries a harder admission gate and moves toward direct internal-load consequence, genuinely independent synchronized evidence, and external device/site transport.

10 · Reliability and dose behavior

ICC(2,1) by conditionselected focal coordinates

<.40 poor.40–.59.60–.74≥.75

Reliability is a property of each coordinate under each condition, not of the instrument as a whole. The canonical table contains 100 condition×metric rows with CV, SEM, and MDC95 in addition to ICC.

Ordered VR-height responsemean within-subject Spearman ρ and 95% bootstrap CI

Multiplicity result.

Peak force and loading rate show unadjusted negative trends, but no focal coordinate survives Holm correction across the declared family. Dynamic Ω mean and sigma show no ordered dose behavior.

The physical 30 cm condition is not part of the ordered VR-height series.

11 · Claim ladder and present boundary

Tier	What may be said	Evidence required	Status
1 · Measurement	Software deterministically computes the named registered quantities.	Tests, fingerprints, accounting, known-answer arms.	earned
2 · Association	A coordinate differs by condition/group in this corpus.	Grouped estimates, uncertainty, multiplicity.	mixed/exploratory
3 · Incremental value	Added coordinates retain information beyond the complete incumbent union.	Positive orthogonal AUC and Brier deltas across nuisance families, bootstrap, fixed-fold null, duplicate and matched-noise controls.	DVJ orthogonal gate fails
4 · Mechanism	A stable compensatory path is localized.	Independent synchronized families and stable phase/anatomy attribution.	not earned
5 · Physical load	Internal force or tissue loading is better estimated.	Independent direct target, target-source exclusion, nested subject holdout, peak/impulse/timing residuals, and external/device transport.	evaluator built; data not tested
6 · Clinical	Patient decisions or harm prediction improve.	Prospective external clinical validation.	not tested

12 · Where we are going

Stage 1d closed · physiological target, no registered-fault moat.
ENABL3S supplies a competent five-mode substrate. The atlas refines confidence but does not add broad accuracy, and its registered-fault detector essentially ties the conventional union. Preserve the result; stop tuning these simulations.

RESULT: bounded signal + failed gates

Stage 2 · CAMS-Knee scientific kill test.
Request access under its noncommercial terms. Freeze the construction before opening implant targets; test direct-load residuals across six subjects and multiple activities. Treat low power as a bound, not transport evidence.

GATE: direct-load residual information

Register the strongest temporal incumbent.
Use raw-waveform LSTM/CNN/Transformer predictions or an immutable adapter under the exact whole-device freeze. Ridge remains a control, never the adversary we claim to beat.

GATE: non-strawman incumbent

Stage 3a · governance and bench rehearsal.
Name the sponsor, rights owner, platforms, sites, safety lead, and independent truth system. Realize all seven faults on a shadow/bench path with calibrated physical severities, start/end clock truth, recovery checks, and hashed evidence before participant recruitment.

GATE: safe independently logged faults

Stage 3b · quarantined participant rehearsal and precision.
Use development hardware only. Estimate clean false refusal, task competence, every-fault sensitivity/localization, device/subject variance, paired detector correlation, attrition, and operating burden. Freeze final n and platform count from confidence-bound resolution; 20 per platform remains only a seed.

GATE: resolvable clustered intervals

Stage 3c · freeze once; open untouched hardware once.
Lock tasks, fault SOP, severities, models, memories, thresholds, multiplicity, missingness, target embargo, and fingerprints. Collect subject-disjoint target devices/sites and report every device, task, severity, and fault. A target-phase repair becomes a new prospective version and cannot erase the primary failure.

GATE: physical + named/rotating transport

Frozen external normative memory.
Build a larger task- and apparatus-declared memory, freeze it before evaluation, and test whether multiple viable basins survive resampling and external subjects. Unknown observations must remain unknown rather than being forced into the nearest basin.

GATE: stable regimes + external calibration

Physical degradation recollection.
Keep the registered simulations for diagnosis, but recollect timing, dropout, substitution, plate loss, and device-change faults physically. Improve the three shift-blind audit surfaces without tuning against the held-out device.

GATE: sensitivity + clean false-deferral

Prospective clinical question—only after tiers 1–5.
Define a decision and outcome first; evaluate incremental utility, calibration, subgroup stability, and consequences. Do not retrofit “injury risk” to CAI classification.

GATE: clinical utility

Strategic boundary. DVJ records one structured injected-fault result; ENABL3S shows that the broader trust story fails: hard loss is detectable, informational faults are mostly missed, and the atlas scarcely exceeds conventional features. More single-corpus simulation tuning is no longer the best path. CAMS answers a separate noncommercial direct-load question; a rights-clean, physically faulted, multi-device Stage 3 determines whether any physical capture-QA capability exists.

13 · Reproduce and inspect

Canonical artifacts

outputs/dvj/measurements.csv
outputs/dvj/trial_accounting.csv
outputs/validation/model_benchmark.csv
outputs/validation/orthogonal_validation.json
outputs/shift_lab/report.json
outputs/capture_qa/report.json
outputs/capture_qa/models.csv
outputs/capture_qa/predictions.csv
outputs/native_two_family/report.json
outputs/native_two_family/trial_metrics.csv
outputs/native_two_family/repeatability.csv
outputs/native_two_family/dose_response.csv
outputs/acquisition/candidate_matrix.csv
outputs/acquisition/decision.json
outputs/acquisition/physionet_admission.json
outputs/physionet/recording_manifest.csv
outputs/physionet/released_clock_audit.csv
outputs/physionet/clock_report.json
outputs/physionet_plurality/report.json
outputs/physionet_plurality/recording_metrics.csv
outputs/physionet_plurality/spectra.csv
outputs/physionet_plurality/speed_matched_permutations.csv
outputs/physionet_local_atlas/report.json
outputs/physionet_atlas_transport/report.json
outputs/physionet_atlas_transport/predictions.csv
outputs/physionet_atlas_transport/same_subject_cross_speed_predictions.csv
outputs/enabl3s_atlas_transport/report.json
outputs/enabl3s_atlas_transport/external_predictions.csv
outputs/enabl3s_atlas_transport/frequency_diagnostics.csv
outputs/enabl3s_atlas_transport/target_native_predictions.csv
outputs/enabl3s_raw_mode_lab/report.json
outputs/enabl3s_raw_mode_lab/stratified_diagnostics.csv
outputs/enabl3s_fault_lab/report.json
outputs/enabl3s_fault_lab/detector_scenarios.csv
outputs/enabl3s_fault_lab/selective_summary.csv
outputs/stage3_capture_protocol/protocol_freeze.json
outputs/instrument_benchmark.json
outputs/research_program_audit.json
outputs/reference_lab_observation.json

Research documents

config/research_program.yaml
config/stage3_capture_protocol.yaml
templates/STAGE3_CAPTURE_MANIFEST_TEMPLATE.csv
templates/STAGE3_PLATFORM_SITE_REGISTER_TEMPLATE.csv
templates/STAGE3_FAULT_SOP_MATRIX_TEMPLATE.csv
templates/STAGE3_REHEARSAL_METRICS_TEMPLATE.csv
docs/COMPUTATIONAL_PHASE_SYNTHESIS.md
docs/PHYSICAL_EXPERIMENT_REQUIREMENTS.md
docs/STAGE3_PHYSICAL_CAPTURE_PROTOCOL.md
docs/REFERENCE_LAB.md
docs/CONSEQUENCE_LAB.md
docs/CAPTURE_QA.md
docs/PHYSIONET_ADMISSION.md
docs/PHYSIONET_PLURALITY.md
docs/LOCAL_PLURALITY_ATLAS.md
docs/ATLAS_TRANSPORT.md
docs/ENABL3S_EXTERNAL_TRANSPORT.md
docs/ENABL3S_RAW_MODE_AND_FAULT_LAB.md
docs/NATIVE_TWO_FAMILY.md
docs/FROZEN_DEVICE_TRANSPORT.md
docs/ACQUISITION_DECISION.md
docs/ORTHOGONAL_INCREMENT.md
docs/INDEPENDENCE_BOUNDARY.md
docs/SHIFT_ABSTENTION.md
docs/COMPETITIVE_POSITION.md
docs/ZTOYBOX_LINEAGE.md
docs/PILOT_RESULTS.md
docs/PREREGISTRATION.md
docs/CLAIMS.md
docs/INDUSTRY_SCORECARD.md
docs/DATA_DICTIONARY.md

Commands

cd /Users/russellparrish/Documents/russell/ZTOYBOX/Biomechanical
PYTHONPATH=src python3 -m pytest tests -q
PYTHONPATH=src python3 -m ztoy_biomech.cli research-audit \
  --program config/research_program.yaml \
  --output outputs/research_program_audit.json
PYTHONPATH=src python3 -m ztoy_biomech.cli instrument-bench \
  --output outputs/instrument_benchmark.json --repetitions 100
PYTHONPATH=src python3 -m ztoy_biomech.cli shift-lab \
  data/extracted/dvj-cai outputs/dvj/measurements.csv \
  --output outputs/shift_lab
PYTHONPATH=src python3 -m ztoy_biomech.cli native-two-family \
  data/extracted/dvj-cai --output outputs/native_two_family
PYTHONPATH=src python3 -m ztoy_biomech.cli capture-qa \
  outputs/shift_lab/shifted_measurements.csv --output outputs/capture_qa
PYTHONPATH=src python3 -m ztoy_biomech.cli acquisition-decision \
  --output outputs/acquisition
PYTHONPATH=src python3 -m ztoy_biomech.cli physionet-audit \
  a-multimodal-gait-dataset-of-brain-activity-muscle-activity-kinematics-and-ground-forces-in-young-adults-1.0.0 \
  --verify-checksums --output outputs/acquisition/physionet_admission.json
PYTHONPATH=src python3 -m ztoy_biomech.cli physionet-clock-audit \
  a-multimodal-gait-dataset-of-brain-activity-muscle-activity-kinematics-and-ground-forces-in-young-adults-1.0.0 \
  --output outputs/physionet
PYTHONPATH=src python3 -m ztoy_biomech.cli physionet-plurality \
  a-multimodal-gait-dataset-of-brain-activity-muscle-activity-kinematics-and-ground-forces-in-young-adults-1.0.0 \
  --output outputs/physionet_plurality --permutations 500
PYTHONPATH=src python3 -m ztoy_biomech.cli physionet-resolution-morphology \
  outputs/physionet_plurality/spectra.csv \
  outputs/physionet_plurality/segment_spectra.csv \
  --output outputs/physionet_resolution_morphology --permutations 500
PYTHONPATH=src python3 -m ztoy_biomech.cli physionet-local-atlas \
  a-multimodal-gait-dataset-of-brain-activity-muscle-activity-kinematics-and-ground-forces-in-young-adults-1.0.0 \
  --output outputs/physionet_local_atlas --permutations 500
PYTHONPATH=src python3 -m ztoy_biomech.cli physionet-atlas-transport \
  outputs/physionet_local_atlas/local_atlas.csv \
  outputs/physionet_resolution_morphology/segment_curves.csv \
  outputs/physionet_local_atlas/segment_features.csv \
  --output outputs/physionet_atlas_transport --bootstrap 1000
PYTHONPATH=src python3 -m ztoy_biomech.cli enabl3s-atlas-transport \
  data/raw/enabl3s outputs/physionet_local_atlas/channel_spectra.csv \
  --output outputs/enabl3s_atlas_transport
PYTHONPATH=src python3 -m ztoy_biomech.cli enabl3s-atlas-transport \
  data/raw/enabl3s outputs/physionet_local_atlas/channel_spectra.csv \
  --sample-rate-hz 500 --output outputs/enabl3s_atlas_transport_clock500
PYTHONPATH=src python3 -m ztoy_biomech.cli enabl3s-raw-mode-lab \
  data/raw/enabl3s --output outputs/enabl3s_raw_mode_lab
PYTHONPATH=src python3 -m ztoy_biomech.cli enabl3s-fault-lab \
  data/raw/enabl3s --output outputs/enabl3s_fault_lab
PYTHONPATH=src python3 -m ztoy_biomech.cli stage3-capture-audit \
  stage3_manifest.csv --data-root /path/to/capture/files \
  --output outputs/stage3_capture_audit
PYTHONPATH=src python3 -m ztoy_biomech.cli stage3-protocol-freeze \
  --protocol config/stage3_capture_protocol.yaml \
  --output outputs/stage3_capture_protocol/protocol_freeze.json
PYTHONPATH=src python3 -m ztoy_biomech.cli reference-lab \
  outputs/dvj/measurements.csv --output outputs/reference-lab
PYTHONPATH=src python3 -m ztoy_biomech.cli consequence-lab \
  waveforms.csv --output outputs/consequence --target internal_load \
  --outcome-kind instrumented_implant --outcome-source implant_sensor \
  --feature-provenance provenance.json --baseline phase force moment \
  --instrument rotation route persistence
PYTHONPATH=src python3 -m ztoy_biomech.cli transport-lab \
  waveforms.csv --output outputs/transport --target internal_load \
  --outcome-kind instrumented_implant --outcome-source implant_sensor \
  --feature-provenance provenance.json --baseline temporal_incumbent \
  --instrument rotation route persistence --device-column device_id
PYTHONPATH=src python3 -m ztoy_biomech.cli dvj-measure \
  data/extracted/dvj-cai --output outputs/dvj
PYTHONPATH=src python3 -m ztoy_biomech.cli dvj-validate \
  outputs/dvj/measurements.csv --output outputs/validation \
  --permutations 100 --bootstrap 1000

This lab book is a living research record, not a medical device report or a terminal pilot summary. The invariant measurement frame is fee2924dbd60ca54; Reference Laboratory configuration is a34d559cae069fa3; frozen library integrity is 32697b2a9102d84b. The executable 29-question graph is config/research_program.yaml; its audit is outputs/research_program_audit.json.