Ask us where any number came from.
Every published validation figure traces to a named locked source record — file, sheet and row. If you want to see the working, ask.
View Technical Validation Details
Provenance — where each published number comes from
| Published number | Where it comes from | How it is calculated |
|---|---|---|
| 58 | Blind-Output Lock Registry and phase records | 14 + 24 + 7 + 9 + 4 locked runs |
| 13 | Validation closeout, distinct-athlete count | 7 + 2 + 4, with no double counting |
| 6 of 6 → 0 of 6 | Prohibited-output rate, unseen formal cohort | Athletes with at least one violation, out of 6 |
| 9 / 9 | Final confirmation suite | Cases passed, out of 9 frozen cases |
| 0 misses | Frozen production baseline | Per-athlete rows, Lead Leg Braking + Rotational Output |
| Retired44 of 99 | Phase-3 pooled view, pre-revision system — superseded, not current performance | Exact matches, out of scoreable feature-level decisions |
| Retired2 of 5 | Phase-3 archetype accuracy, pre-revision system — superseded, not current performance | Confirmed, out of tested instances |
Production configuration locked 2026-08-14 with cryptographic hash verification. Any future analytical change requires a new version, a verified diff and a fresh confirmation suite before it can be authorised.
The percentages we can publish, individually
The frozen production system has no publishable percentage, and that is a deliberate feature of the record rather than an omission. Its scored pass produced counts, not a rate; the counts that reconcile exactly are all zeros; and our own validation record prohibits building a blended cross-round figure. A percentage here would be manufactured, so there isn't one. Every publishable percentage below describes the system we replaced.
| Value | Fraction | What it measures | Scope and denominator |
|---|---|---|---|
| Retired44.4% | 44 of 99 | Exact-match rate at the individual-finding level | Pre-revision system, ungoverned, 7-athlete controlled retest, Full-7 cohort view · 99 scoreable feature-level decisions |
| Retired42.3% | 33 of 78 | Same measure, same-era subset | Excludes the two athletes whose benchmark data is years older than their assessment · 78 scoreable feature-level decisions |
| 47.5% | 28 of 59 | Same measure, clean-input subset | Excludes the two athletes with documented blind-test contamination exposure · 59 scoreable feature-level decisions |
| 40% | 2 of 5 | Accuracy of the most frequently selected primary finding | Pre-revision system, the 5 athletes on whom that archetype was tested · 5 tested instances |
| 25% | 1 of 4 | Same finding, ungoverned, on the next unseen cohort | 4 resolvable athletes; 2 excluded because their benchmark format could not grade it · 4 resolvable athletes of 6 |
These are separate measurements with separate denominators. They are never averaged together, and no cross-round global accuracy percentage exists.
Development phases, reported separately
Phases are reported separately by design — each used its own scoring method and denominator. The progression itself is the evidence; blending it into one figure would destroy it.
-
2026-05 → 2026-07
Early development
Low accuracy under an informal rubric. Superseded — these figures are not publishable.
-
2026-07-19 → 2026-07-20
Phase 3 — controlled retest
Ungoverned system matched the benchmark exactly on 44 of 99 scoreable decisions. Four systemic failure modes documented.
-
2026-07-20 → 2026-07-21
Revision design
12 amendments drafted, frozen and hashed. Not authorised for operational or commercial use during the round.
-
2026-07-21 → 2026-08-14
Phase 6 — unseen four-condition round
Unsupported claims: 6 of 6 athletes ungoverned, 0 of 6 governed. Two new failure modes found.
-
2026-08-15
Amendment regression
8 fixes tested against pre-registered criteria: 4 passed, 4 failed or were confounded. All published.
-
2026-08-17 → production lock
Targeted revision + frozen final confirmation
9 of 9 confirmation cases passed. Candidate frozen and hashed before the first case ran.
-
2026-08-14T17:43:25Z
Production lock
Prompt and governance locked with SHA-256 hashes. Substantive diff vs. the tested artifacts: zero.
-
2026-08-15
Apex Core v1 frozen baseline
6 direct hits, 0 misses, 0 false positives — Lead Leg Braking and Rotational Output only, across 13 athletes.
Documented misses, in full
- Amplitude is not velocityA large, visually dramatic torso rotation was read as fast rotation; the measurement said it was one of the athlete's worst features. Every configuration willing to state a direction got it backwards, in 3 of 4 scored athletes. This is the failure the final revision was built to fix, and it did not recur in the confirmation suite.
- A real limiter went unconveyedThe one held-out athlete had a genuine lead-leg braking limiter ranked fourth in their benchmark report. The locked system withheld rather than calling it. Separately, a hip-mobility limiter flagged as the single largest deviation in one report was underweighted or withheld by all four configurations tested.
- A good composite can hide a bad featureA correctly-called above-average build-up composite contained a feature scoring 6 out of 100. A 32-degree trunk lateral-tilt outlier was surfaced by none of the four configurations.
- Half of our own fixes failed on first testOf eight proposed amendments, four passed and four failed or could not be cleanly tested. Two were confounded by an unrelated pitch-count gate; one feature turned out never to have been implemented at all; one was declined for production. All four are on the record as failures.
Full limitation register
- Sample size13 distinct athletes; the formal unseen cohort was N = 6; one realised held-out athlete. All rates are directional evidence, not statistically tight estimates.
- No overall accuracy figure existsScoring methods and denominators differ across rounds by design. The locked source states a cross-round global accuracy percentage “DOES NOT EXIST” and “must never be created”.
- Category scopeThe frozen validation pass scored 2 of the 8 framework categories at full category level. The other 6 have finding-level evidence only.
- Benchmark format heterogeneityGround truth came from several independent 3D motion-capture and biomechanics providers plus narrative laboratory reports, which differ in metrics, normalisation and whether population reference bands are printed at all.
- Some benchmarks cannot grade a claimOne formal-cohort export carried no population reference bands, so above/below-average claims were logged Not Determinable rather than scored. One pilot athlete's benchmark images were cropped.
- Era gapsOne athlete's benchmark predates their assessment by roughly five years. Another has two sessions six weeks apart with real test-retest variability.
- Validation footage is not production footageLegacy validation footage is explicitly not representative of the standardised submissions production expects. Video quality was logged as methodological context and no locked result was re-weighted because of it.
- Older intakes lack current fieldsOlder intake packets lack fields the current Master Intake Form captures — one documented miss (Hip Internal Rotation) traces directly to unstructured intake wording rather than to the assessment logic.
- Documentation gapsSeveral historical and condition files lack input-hash manifests or precise UTC lock timestamps. Only 2 of 9 final confirmation reports self-document input hashes.
- Procedural deviations, logged not curedThe two pilot athletes ran in fixed rather than randomised condition order, and were opened before any held-out subset was designated. Both were permanently excluded from the formal held-out subset rather than retroactively reclassified.
- Amendment failures4 of 8 amendments failed or were confounded on first regression. Two were confounded by an unrelated pitch-count gate; one feature was never implemented; one was declined for v1.0. All four are published as failures.
- Held-out subset was under-filledTwo held-out slots were designated; only one was filled, because the cohort closed at six athletes when eligible athletes with usable benchmark data ran out.
- Video observation is not laboratory measurementApex reads observational evidence from video plus intake. It does not measure joint angles or segment velocities. Claims that require direct measurement — exact hip-shoulder separation in degrees, rotational velocity in deg/s, projected velocity gain — are categorically blocked, not estimated.
- Computer vision is not production-approved for pitchingZero pitching CV measurements are production-approved. The most promising metric (throwing elbow flexion at ball release, mean absolute difference 5.44 deg) remains RESEARCH_ONLY for insufficient N and replication. Mobility CV is authorised only as a lower-authority supporting layer that cannot change or create a grade.
Source figures, as generated from the validation package
These are the charts produced directly from the locked validation package. They are retained here unchanged so that every figure on this page can be checked against its source rendering.
Cohort and run composition
Results, governance and confirmation
Category coverage, amendments and phase progression
