Skip to content

Phase 5 exit review — 2026-09-10

Decision: Phase 5's exit gate is CLOSED as a conditional pass. Every exit-gate criterion is met. The Phase 4 gate's HOLD stands as recorded and is resolved as a construction defect with a known fix, not as a defect in the delta/angle/shape estimators. The magnitude specificity claim is withheld pending re-measurement of the corrected construction. Phase 6 is authorised to start. Phase 7 stays deferred.

Inputs. Paper-grade findings report (run results/phase5-2026-09-10/, 19 cells × 500 replicates × 999 permutations = 9,500 units, job 880599) and the magnitude-construction diagnostic (run results/magnitude-construction-2026-09-10/). No new measurement was made for this review; it decides on what those two produced.

1. The exit-gate criteria are met

The roadmap's Phase 5 exit gate has three criteria. Each is satisfied independently of the Phase 4 gate's verdict, and the distinction matters: the Phase 4 gate asks whether the statistics met predeclared operating targets, while the exit gate asks whether Phase 5 produced publishable evidence.

criterion status evidence
Conclusions are stable at paper-grade Monte Carlo precision met Every quantity agrees with two independent matched-seed pilot families within Monte Carlo error, including both failures (magnitude/angle 0.276 vs the ladder's 0.24; magnitude/shape 0.742 vs 0.74). The 22-fold increase in Monte Carlo work changed no conclusion.
All deviations from preregistered targets are explained met Four deviations, each with a mechanism and a named revision — §3 below. None is a Monte Carlo sample-size question, and none was addressed by raising the replicate count (which the Phase 4 exit gate forbids).
The report distinguishes statistical operating characteristics from biological construction cross-talk met This is precisely what the diagnostic settled: the magnitude off-target response is construction geometry (exactly size-pure within every omic block, angle 0.0000 at population_native), not estimator behaviour.

2. What Phase 5 establishes, and what it withholds

Claimable now, at paper-grade precision:

  • Type I error control on all three statistics: 0.040–0.062 across both baseline cells against α = 0.05, every one within 2·SE (report/acceptance_report.csv, 6/6 targets met), and all 21 gate inflation checks met.
  • Translation invariance: a constant observed-space offset moves none of the three statistics at any effect (0.028–0.056 across the whole sweep). This is the estimators' invariance contract, confirmed.
  • Power on all three diagonals against the 0.80 floor: magnitude/delta 1.000, orientation/angle 0.840 ± 0.016, shape/shape 0.996.
  • The measurement architecture: features → PLS latent space → trajectory geometry, with the stage-supervised double-CV rank rule selecting rank 3 = n_stages − 1 in all 19 cells.
  • Operational reproducibility: 9,500/9,500 units present once, 0 signature mismatches, 0 failures, 0 censored surgeries, and a committed report/ that regenerates byte-identically from the merged records.

Withheld pending re-measurement:

  • The specificity claim for angle and shape against a pure size change. The construction used to test it was not a pure size change (§3), so the test was invalid rather than failed. The claim is neither asserted nor abandoned; it awaits the corrected construction. The paper's specificity section cannot be written until then.

Not claimable, on independent grounds:

  • Cross-replicate driver stability — no design holds the driver set fixed, so it is descriptive only.
  • A strictly-monotone orientation power claim — see §3.

3. Resolution of each deviation

D1 — mandatory_control[magnitude,angle] 0.276 > 0.0695 (10.3 SE). Method revision. magnitude_kind='all' scales delta_methyl alone while the measurement space standardizes and concatenates all three omic blocks, so group B grows in one block and not the others and the pooled trajectory rotates by construction. Within each omic block the surgery is exactly size-pure (angle 0.0000 and shape ~1e-17 at population_native, every effect); the response appears only in the joint scope (angle 6.70° → 20.33°) and is confirmed shape-free at two stages (19.05° vs 18.91° at four). A uniform-δ construction that scales every omic's δ together is exactly size-pure in the joint space after per-block standardization (joint angle 8.5e-07 → 2.6e-06, joint shape ~1e-15 — the anchor's own floor — while joint delta grows 14.96 → 44.80). The angle estimator is not implicated. Remedy: adopt the corrected construction, calibrate its effect axis, re-measure. Tracked as its own change (§5).

D2 — mandatory_control[magnitude,shape] 0.742 > 0.0695 (34.4 SE). Method revision, same fix. Same mechanism. The paper-grade report named a diagnostic prerequisite — whether the 0.05 normalized materiality threshold was calibrated for the shape statistic's scale — and it is discharged: it was not. localize_off_diagonal normalized delta by path length and angle by 180° but passed shape through raw, so one absolute cut could never classify a statistic whose entire response range is ~1e-2. The recalibrated null-dispersion rule classifies magnitude/shape as construction_present at every effect while leaving translation not_material on all eight rows. The response was real all along and the instrument was hiding it.

D3 — power[orientation,angle] acceptance target not met (monotonicity). Claim revision. The 0.80 floor is cleared (0.840). The strictly-monotone form is not supported: the curve saturates from e = 0.25 (0.842, 0.878, 0.872, 0.840) because the angle null widens with the observed statistic — median q95 5.5° at the anchor to 11.3° at e = 1.00, with the IQR growing faster (4.9° → 13.9°). This is the pivotality property already established in 2026-09-01, showing its consequence at paper grade. Adopted claim: a saturating orientation response that clears the floor, stated with the eigengap stratification (0.754 / 0.886 / 0.880 by tercile), never as a monotone power curve. No replicate count or rank change addresses it — the ladder already returned keep_cv over five fixed ranks; a different pivot would.

D4 — orientation/shape predeclared "projection-associated". Claim revision. Refuted. The label does not survive. Orientation's response is present within each omic block (max block angle 90.08° against joint 89.93°; max block shape 0.108 against joint 0.059), and under the recalibrated materiality rule no pair anywhere in the study localizes as projection-associated — the paper-grade run's single such classification, orientation/delta at e = 1.00, becomes construction_present once the population checkpoint is judged against its own null. Adopted claim: orientation → shape is construction-present cross-talk of a global per-omic feature permutation. No gate impact — both orientation off-diagonals are descriptive roles. The projection-associated framing must not be repeated.

4. Phase 6 is authorised, with one binding condition

Phase 6 (real-data case study) may start. The HOLD concerns a synthetic construction; real cohorts have no surgery modes, so nothing about D1/D2 blocks applying the validated configuration to real data.

Binding condition — report orientation power against the cohort's own recorded eigengap, never the pooled figure. Pooled orientation/angle power at e = 1.00 is 0.840, but stratified by eigengap tercile it is 0.754 / 0.886 / 0.880: the lowest tercile sits below the predeclared 0.80 floor. The eigengap, not n and not ρ, is the observable that transfers. A cohort landing in that tercile does not inherit 0.840.

Two stated limitations travel with any Phase 6 orientation result:

  • The claim is n-conditional: established at n = 1200 (300 per group-stage cell) at the isotropic endpoint ρ = 0. Along n at fixed ρ the eigengap is constant while the angle null width contracts.
  • The orientation surgery's realized latent contrast is not ρ-invariant (median observed angle 28° at ρ = 0.5 vs 66° at ρ = 0), so baseline continuity is not a lever that buys power.

5. Phase 7 stays deferred — question closed

SNF-native trajectory statistics remain deferred pending graph-native metrics. Nothing in the paper-grade run or the diagnostic bears on SNF: the study measured the PLS latent space throughout, and concat appeared only as a baseline. The deferral is therefore not an open question awaiting this review's evidence — it awaits a decision to define graph-native path-length and angle statistics, which is independent work. Recording it as closed here so it stops reading as pending.

6. Consequent work, in order

  1. Adopt the size-pure magnitude construction (its own change). Promote the uniform-δ probe from diagnostic-only to a production magnitude_kind; it is currently unreachable from configuration and refused by the loader by design. Calibrate its effect axis first — at e = 1.00 it reaches joint delta 44.80 against production's 18.14, so the two effect axes are not comparable and its power curve cannot be read beside the existing one without recalibration. Then re-measure the magnitude controls and lift the withheld specificity claim or report it as failed on its own terms.
  2. Phase 6 real-data case study, under the §4 condition.
  3. Consider making the null-dispersion materiality rule the default. It is opt-in here so the committed Phase 5 report/ keeps regenerating byte-identically. Switching the default is a reporting change that should come with a recorded re-issue of any affected localization output — and note the interpretive consequence: because population checkpoints carry analytically zero nulls, the rule answers "is this construction exactly pure?" rather than "is the impurity large enough to matter". Severity is read from the response magnitude, not the label.

7. What this review does not do

It adopts no construction change, re-runs no study, and edits no predeclared artefact — the config, gate roles, acceptance targets, and report contract stay frozen, and the committed Phase 5 report/ is untouched. The conditional pass is a statement about what Phase 5 established, not a waiver of the Phase 4 gate's verdict, which stands as recorded in results/phase5-2026-09-10/report/phase4_gate_decision.json.