MATERIALITY MAP — ALL 57 HELD-OUT CONTRASTS

Status: BUILDER-SUPPORT PROBE. Interpretation layer only. Artifact: hierarchical-aif/results/motor_stack_aif/materiality_map_all_contrasts.json Builder: hierarchical-aif/scripts/build_materiality_map.py Channel: DURATION_ONLY — nextStateN / direction / jump were never requested (D5). No new science. No refit, no bootstrap, no data load. Every number is copied from a frozen artifact or derived from it by subtraction. No frozen verdict is altered. No P-level moves.

1. Headline

Of the 57 contrasts this repository holds, 25 are scored in NLPD nats and can be compared against the resolution floor at all. The other 32 are CRPS and are excluded on units grounds (section 3).

Among the 25 nats-scale contrasts:

22 of 25 nats-scale contrasts sit at or below the resolution floor.

And reportableAsAWin — frozen verdict resolved and effect above the floor — is true for exactly one contrast in the entire repository:

FSIDE | derived_eligible_1_to_8 | NLPD_motor_equal | M0_EXPONENTIAL effect +0.11529094279542929 nats, percentile interval [+0.013039958895400987, +0.22749707235626093], frozen verdict RESOLVED_ABOVE, effect/floor 2.745.

That is the F-side motor stack out-predicting a single-rate exponential on held-out durations. It is the weakest adversary in the set. Nothing else clears both bars.

Frozen verdicts across all 57 rows: INCONCLUSIVE 48 · NOT_ESTABLISHED 6 · RESOLVED_ABOVE 3. All 48 frozen B3 contrasts are INCONCLUSIVE — the frozen record already said so; this map only adds why.

2. The resolution floor — what it is and what it is not

resolutionFloor = 0.042 nats/event. It is a HALF-width.

Derivation: the BCa half-width of the narrowest frozen B3 contrast — cohort derived_eligible_1_to_8, rule NLPD_motor_equal, model M4_MIXTURE_K3. BCa width 0.08414086126525253, half 0.04207043063262626, used as 0.042 (the same constant already hard-coded in scripts/recompute_m4_m6_m7_per_motor.py and recorded as resolution.halfWidthFloorNats in the F-side result).

It is a HEURISTIC. It was not pre-specified, it is not an equivalence margin, and no verdict anywhere in this repository was decided by it. It describes what this assay can resolve; it does not describe what the bootstrap will call. The superseded figure 0.064 came from the mislabelled width field (D7).

Cohort provenance: the floor was derived on derived_eligible_1_to_8. Rows from primary_states_0_to_8 carry floorTransferCaveat = true in the JSON — same units, different cohort, so the comparison is indicative rather than exact.

MATERIAL means large enough to matter. It does not mean resolved. Two of the three MATERIAL rows carry the frozen verdict INCONCLUSIVE (section 5).

3. Units fence — the CRPS rules do not get a floor

B3 scores three rules: NLPD_motor_equal, CRPS_normalized, CRPS_seconds. The 0.042 figure is a nats/event quantity. CRPS_seconds is in seconds; CRPS_normalized is dimensionless CRPS on normalised y. Neither is nats.

Therefore for all 32 CRPS rows the JSON emits:

resolutionFloor      = NOT_APPLICABLE_DIFFERENT_UNITS
effectToFloorRatio   = NOT_APPLICABLE_DIFFERENT_UNITS
scientificReading    = NOT_APPLICABLE_DIFFERENT_UNITS
reportableAsAWin     = NOT_APPLICABLE_DIFFERENT_UNITS

Carrying a nats floor onto a CRPS scale would be a units error of exactly the kind the truth contract forbids. Their frozen verdicts, point estimates, BCa widths and percentile widths are still recorded in full — only the floor comparison is withheld. A CRPS-side materiality floor is NOT_COMPUTED; deriving one would need the same narrowest-contrast construction repeated in CRPS units, which this probe did not do.

4. BCa availability — NOT_COMPUTED for 9 of 57

The F-side harness and the M4/M6/M7 extension emit percentile intervals only — they never ran a BCa. BCaWidth for those 9 rows is the literal string NOT_COMPUTED. This probe does not compute, estimate, impute or approximate a BCa for them. Doing so would manufacture a number that no run ever produced, and it would not be comparable with the frozen 48 in any case, because it would come from a different harness.

D7 re-checked independently here, by subtracting the frozen bca / percentile arrays rather than trusting the ledger text:

• publishedWidthField equals the percentile width in 48 / 48 contrasts, the BCa width in 0 / 48.
• intervalUsed equals the bca array in 48 / 48, the percentile array in 0 / 48.

So the published width describes an interval that decided nothing, and the interval that decided every verdict had its width unpublished. Confirmed, from the artifact.

5. The MATERIAL three

Every material effect in this repository involves M0_EXPONENTIAL, and nothing else. The single largest signal the assay contains is "a one-parameter exponential is worse".

Note the first two rows: on the B3 sign convention contrast = S(M3) - S(M0), a negative point estimate means the reference M3_TWO_TIMESCALE scored better than M0_EXPONENTIAL by ~0.11 nats — 2.7× the floor — and the frozen BCa interval still crossed zero. The largest effect in the frozen B3 NLPD set is not resolved by the frozen B3 bootstrap. That is the power limit speaking, not the models.

6. The D10 class — frozen verdict resolved, effect below the floor

Two contrasts, both already carrying the interpretation in their own artifacts:

1. FSIDE | derived_eligible_1_to_8 | NLPD_motor_equal | M5_GAMMA effect +0.031036 nats, percentile [+0.004459…, +0.054473…], frozen verdict RESOLVED_ABOVE, effect/floor 0.739 → SUB_FLOOR_EFFECT. The interval reaches outside the floor on one side, so this is not SCIENTIFICALLY_NULL, but the point estimate sits below it. The resolution rests on consistency of sign across the 19 motors, not on effect magnitude.

2. M4M6M7 | derived_eligible_1_to_8 | NLPD_motor_equal | M7_HIERARCHICAL_MOTOR effect +2.506984e-07 nats, percentile [+1.604451e-07, +3.374688e-07], frozen verdict RESOLVED_ABOVE, effect/floor 5.969e-06 → SCIENTIFICALLY_NULL. This is D10's exhibit: ~168 000× below the floor, resolved because the F-side hierarchy is M7 to numerical precision (exp(mu)=0.6596326697… vs M7 k=0.6596322379…; tau=0.18372083 vs 0.18372186). Its own artifact already carries the mandatory WARNING.

Both frozen verdicts are reported verbatim and unaltered. The reading sits beside the verdict, never on top of it. Neither is reportableAsAWin.

One further row deserves naming even though its verdict did not resolve: FSIDE | … | M1_WEIBULL, effect +6.145101e-04, interval [−0.010881…, +0.011304…] — the entire interval lies inside ±0.042, so it is SCIENTIFICALLY_NULL rather than merely inconclusive. Together with M4M6M7 | … | M6_SEMI_MARKOV_STATE_DEPENDENT (−2.803367e-03, interval [−0.018990…, +0.013913…]) these are the only two contrasts where the assay can say the difference is small and bounded — which is a stronger statement than NOT_ESTABLISHED, and still not a statement of sameness.

7. underpowered flags already present in the frozen record

The frozen B3 artifact carries its own underpowered boolean. It is true in 25 of the 48 frozen contrasts. Verbatim list:

derived_eligible_1_to_8

• CRPS_normalized: M0_EXPONENTIAL, M2_LOGNORMAL, M4_MIXTURE_K3, M5_GAMMA
• CRPS_seconds: M0_EXPONENTIAL, M4_MIXTURE_K3, M5_GAMMA
• NLPD_motor_equal: M0_EXPONENTIAL, M5_GAMMA, M6_SEMI_MARKOV_STATE_DEPENDENT, M7_HIERARCHICAL_MOTOR

primary_states_0_to_8

• CRPS_normalized: M0_EXPONENTIAL, M2_LOGNORMAL, M4_MIXTURE_K3, M5_GAMMA, M8_EMPIRICAL_KDE
• CRPS_seconds: M0_EXPONENTIAL, M2_LOGNORMAL, M4_MIXTURE_K3, M5_GAMMA, M7_HIERARCHICAL_MOTOR, M8_EMPIRICAL_KDE
• NLPD_motor_equal: M0_EXPONENTIAL, M5_GAMMA, M7_HIERARCHICAL_MOTOR

The F-side and M4/M6/M7 artifacts carry no underpowered field, so those 9 rows record frozenUnderpowered = NOT_COMPUTED. Their harness reported power differently (via resolution.canResolve / atOrBelowResolutionFloor), and this probe does not back-fill a flag that run never emitted.

8. What a wall of sub-floor contrasts actually means

It is a statement about the assay, not about the models.

The binding constraint is 19 holdout motors. The experimental unit is the MOTOR; the 233 holdout events inside them are not independent replicates. A paired motor-cluster bootstrap over 19 clusters cannot separate models that differ by a few hundredths of a nat unless the sign of the per-motor difference happens to be near-unanimous — which is precisely how M7 "resolved" at 2.5e-07 nats and how M5_GAMMA resolved at 0.031.

Consequences that must travel with any use of this map:

• An interval containing 0 is NOT_ESTABLISHED. It is never "no difference" and never "equivalent". Underpowered is not equivalence.
• 22 of 25 nats-scale contrasts being at or below the floor does not rank the models. It says the instrument cannot separate them here.
• Replicates were not increased after seeing a width, and must not be.
• Nothing in this map is prospective. The F-side scoring rows are NOT_SATISFIED on prospectivity (D9); the M4/M6/M7 rows are NOT_SATISFIED and additionally POST_HOC_EXPLORATORY. The frozen B3 rows retain whatever standing their own record gives them; this probe does not re-adjudicate it.
• No mechanism is established by any row here. Predictive ordering on held-out durations is not mechanism, and this is duration-only, one dataset, one species (E. coli, Wadhwa 2022), retrospective-only.

9. Reproduction

python hierarchical-aif/scripts/build_materiality_map.py

Input hashes (recorded in the JSON under inputs):

audits/** was read only; git status --porcelain audits/ is empty after this probe.

Output sha256 of materiality_map_all_contrasts.json at build time: 1aecdec94e37d402f73e821ff237e3a7af3b6c620545ba4f60eee6bdeb2e287b.

Self-check built into the script: for the three M4/M6/M7 rows the re-derived scientificReading is compared against the classification already stored in M4_M6_M7_PER_MOTOR_CONTRASTS_RESULT.json, and the script HALTS on any disagreement. It did not halt — all three agree, which is evidence that the rule transcribed here is the same rule that artifact used.

Field additions beyond the nine requested. Each row carries the nine required fields (frozenVerdict, effectSize, intervalUsed, BCaWidth, percentileWidth, resolutionFloor, effectToFloorRatio, scientificReading, reportableAsAWin) plus identification keys (contrastId, source, sourceArtifact, cohort, rule, units, reference, challenger, signConvention, nHoldoutMotors) and provenance flags (intervalUsedValues, publishedWidthField, scientificReadingWhy, frozenBeatsM3, frozenUnderpowered, floorTransferCaveat, prospectivity). The additions are identification and caveats; none of them is a computed result.

10. Limitations

• The floor is a heuristic derived from one contrast on one cohort. It is not a registered margin.
• MATERIAL is a magnitude statement only. It carries no verdict.
• CRPS materiality is NOT_COMPUTED — 32 of 57 rows are therefore uncharacterised on materiality.
• BCa widths for 9 of 57 rows are NOT_COMPUTED and were deliberately not imputed, so BCa-vs- percentile width divergence cannot be assessed for the F-side or M4/M6/M7 families.
• Cross-cohort floor transfer to primary_states_0_to_8 is flagged, not validated.
• This probe read no held-out mark channel and therefore says nothing about nextStateN, direction or jump. Any question about those is NOT_CHECKED — would require holdout access.

What would make this lane true, and what would kill it

An earlier draft of this map diagnosed the assay without naming a discriminator. That is the gap this section closes.

The lane: the motor-equal held-out scoring assay can distinguish the candidate models it is being asked to distinguish. On the present evidence it largely cannot: of 25 NLPD-nats rows, 3 are MATERIAL, 19 are SUB_FLOOR_EFFECT and 3 are SCIENTIFICALLY_NULL, and the frozen record already carries 25 underpowered flags. Exactly one row is reportableAsAWin (FSIDE | derived_eligible_1_to_8 | NLPD_motor_equal | M0_EXPONENTIAL) — beating a unit exponential. That is a statement about the assay, not about the models.

Receipt that would make the lane true. An independent held-out cohort with enough MOTORS that the motor-equal half-width falls below the stated floor for the contrasts that matter — so that a real difference between M2_LOGNORMAL, M4_MIXTURE_K3, M6_SEMI_MARKOV_STATE_DEPENDENT, M7_HIERARCHICAL_MOTOR and the F-side candidate could resolve at all. The required motor count is NOT_COMPUTED here; the companion probe reports/POWER-ATLAS-MOTOR-EQUAL-SCORING.md estimates it on synthetic data, and any target resolve-rate used there is DESIGN_ONLY. Because such a cohort is external data, this receipt sits on P4 transfer — which no modelling in this repository can close.

What would kill it.

1. The single reportable win fails to reproduce. If, on an independent cohort, the F-side-over-M0_EXPONENTIAL contrast no longer excludes 0, the one reportableAsAWin row in this map is withdrawn and the assay retains no material win at all.
2. A larger cohort still resolves nothing material. If motor count rises substantially and the serious contrasts remain SUB_FLOOR_EFFECT, then these models are genuinely indistinguishable on duration alone, and discriminating evidence must come from a different channel — intervention, transfer, or the quarantined mark process — not from more motors.
3. A sub-floor contrast is reported as a win anywhere downstream. That would be D10 recurring; reportableAsAWin exists precisely to make it mechanically detectable.

What this map cannot do: it changes no frozen verdict, moves no P-level, and adds no claim. It supplies the materiality reading the frozen CI rule does not.

NEXT_ACT = Derive a CRPS-units resolution floor by the same narrowest-frozen-contrast construction (narrowest BCa half-width among the 16 CRPS_seconds and separately the 16 CRPS_normalized frozen B3 contrasts), so the 32 CRPS rows currently NOT_APPLICABLE_DIFFERENT_UNITS can be given a units-correct materiality reading; label it DESIGN_ONLY and do not let it touch any frozen verdict.