CENSORING STRESS DISCRIMINATOR — PROBE 6

Type: BUILDER-SUPPORT PROBE. Not a verdict. Moves no P-level, changes no frozen verdict, creates no claim. Date: 2026-07-22 · Repo HEAD at probe time: 28ce738 Lane: D (motor-stack AIF) with a bounded read into lane A (duration-only B3/B4). Split boundary: duration_only. No held-out nextStateN, direction or jump was read, loaded, printed or reasoned from. The D5 firewall was not approached. Experimental unit: the MOTOR. All held-out scores are motor-equal. Compute: three short single-process runs, 2.4 s + 14.5 s + 11.4 s wall. No parallel compute. No write to results/. No write under audits/.

0. The central fact this probe is built around

The frozen B3 cohort excludes right-censored events entirely. The rule is one line:

audits/phase-b/b3-model-competition-runner.py:109-110
        elig = [e for e in events
                if (not e["rightCensored"]) and e["stateN"] in self.states]

Consequences that govern everything below:

1. The F-side motor stack's censoring branch (hazard_survival.log_event_density: uncensored log h + log S, censored log S) is correct-by-test but UNEXERCISED on the frozen cohort. Every event it has ever scored had rightCensored = False. Its censoring correctness has therefore contributed zero to the F-side held-out result.
2. Therefore no censoring comparison in this probe can be run on the real cohort. Sections 2 and 3 are SYNTHETIC with a KNOWN generating shape. Section 1 is the only OBSERVED section, and it reports the consequence of the exclusion, not a censoring comparison.

The two are labelled sharply throughout: [OBSERVED] vs [SYNTHETIC].

1. [OBSERVED] What the exclusion removes from the real ingest

Source: Wadhwa 2022 E. coli ingest via _bridge.b3().load_events() — the duration channel only. These are recorded measurements. Reproduction commands in §7.

Censored-event state distribution (raw): {4:1, 5:3, 6:5, 7:17, 8:28, 9:29, 10:19, 11:7} — zero censored events in states 0, 1, 2, 3.

The censored events are the long ones.

Within-motor position of the censored dwell (104 motors with ≥1 uncensored event):

• 48 of 104 motors (46.1538%) have their censored dwell longer than every uncensored dwell they contribute.
• Mean within-motor quantile rank of the censored dwell among that motor's uncensored dwells: 0.802900165.

Removed dwell-time mass:

• All states: censored 13438.2000 s vs uncensored 19453.7200 s → 40.8556% of recorded dwell-time sits in censored events.
• States 1–8 only: censored 6240.6800 s vs uncensored 12775.3200 s → 32.8180% of in-band recorded dwell-time is removed by the exclusion.

Reading (fact, not inference). The frozen filter drops 5.0% of in-band events but 32.8% of in-band dwell-time, and it does so non-randomly: it removes each motor's terminal dwell, which is the longest observation that motor supplies in 46% of cases. The frozen cohort is a tail-depleted subpopulation, and every model on the B3 leaderboard was fitted and scored on that subpopulation.

2. [SYNTHETIC] Design of the censoring stress

A world where the truth is known, built to resemble the cohort without being it.

Matched oracle. Every replicate also fits the same draw with no censoring at the true scale. All treatment effects are reported as differences against that matched oracle (dk, dsigma, dNLPD), so the shared sampling fluctuation is differenced out. Oracle performance:

• k_hat = 0.626741 (sd 0.014650, se 0.002316); bias vs k_true = +0.001652.
• sigma_hat (mean-one lognormal, B3 M2 form) = 1.793632 (sd 0.041098).
• Held-out motor-equal NLPD in seconds = Weibull 3.122483, lognormal 3.247492; gap (lognormal − Weibull) = +0.125009 (se 0.001874). Positive = the true family wins.

Two normalisation modes, because normalisation is a second censoring handle.

• Mode A SCALE_TRUE — normalise by the true scale_N. Isolates the likelihood treatment. Counterfactual; no pipeline does this.
• Mode B SCALE_FROZEN_RULE — per-state scale = mean duration of uncensored events only, which is exactly Cohort.scale_N at lines 115–120. This is what the real pipeline does, regardless of which likelihood is used afterwards.

All held-out NLPD is reported in seconds (B3 convention: normalised-y NLPD + log scale), so scores remain comparable across treatments that disagree about the scale.

Five stress treatments. T1 correct survival likelihood · T2 drop censored · T3 treat censored as completed events · T4 upper-tail trim · T5 state-specific censor sensitivity (§3.4).

Realised censoring rates (mean over 40 replicates): W=60 → 0.043375 (sd 0.006865); W=20 → 0.173250 (sd 0.014765).

Window calibration against the real cohort — this decides which row to read.

W = 60 s is the row that resembles the real cohort, on both event rate and removed time mass. W = 20 s is a deliberate over-stress, roughly four times the real exposure.

3. [SYNTHETIC] Results

dk = shape minus matched oracle shape. dk negative = apparently HEAVIER tail. dsigma = lognormal sigma minus oracle sigma. dsigma positive = apparently HEAVIER tail. dNLPD_s = held-out motor-equal seconds-NLPD penalty vs the oracle, on the full population. All values are means over R = 40; se is the standard error over replicates.

3.1 Mode A — likelihood treatment only, true scale held fixed

Findings.

1. T1 recovers the truth. At both windows the correct survival likelihood is statistically indistinguishable from the matched no-censoring oracle: dk = −0.000068 (se 0.000707) and +0.000244 (se 0.000852), both intervals containing 0. The censoring branch does what it is documented to do.
2. Direction of induced error is NOT fixed. T2 (drop) moves the shape up at 4.3% censoring (+0.011479, lighter tail) and down at 17.3% censoring (−0.032968, heavier tail), with the lognormal sigma inflated by +0.093467. T3 (censored-as-events) moves the shape up at both rates. So "mistreating censoring manufactures a heavy tail" is regime-dependent, not a rule.
3. At the real-like exposure the likelihood handle alone is immaterial. Every W=60 Mode A dNLPD_s is ≤ +0.001313 nats — roughly 32× below the corrected motor-equal resolution floor of ≈0.042 nats. On the frozen cohort's censoring exposure, getting the censoring likelihood wrong would not be detectable by the frozen CI machinery.

3.2 Mode B — the frozen normalisation rule. This is where the cost lives.

Findings.

4. The normalisation handle dominates the likelihood handle by one to two orders of magnitude. At W=60 the frozen-rule scale costs +0.035502 seconds-NLPD even under the correct likelihood, against +0.000088 for the same likelihood at the true scale — a factor of ~403.
5. Two of four W=60 Mode B cells sit at or above the 0.042-nat resolution floor (+0.042561 drop-censored, +0.043165 trim). The other two sit below it: +0.039388 censored-as-events and +0.035502 correct-survival — the latter was omitted entirely from an earlier draft of this sentence, which also miscounted two as three. At the real cohort's censoring exposure, the cost of estimating scale_N from the tail-depleted set is at the edge of scientific materiality in this synthetic world.
6. Under the frozen rule the induced error points toward a LIGHTER apparent tail, not heavier: dk positive and dsigma negative in all eight Mode B cells. This is the opposite direction to the Mode A drop-censored cell at high censoring, and it matters for §4.

3.3 The estimand split — why the frozen score cannot see any of this

Same Mode B fitted models, scored two ways. gap = lognormal − Weibull seconds-NLPD; positive = the true (Weibull) family wins. Oracle gap = +0.125009 (se 0.001874).

Findings.

7. At the over-stress window the true family's advantage collapses. W=20 drop-censored under the frozen normalisation takes the gap from +0.125009 to −0.012847. The paired shift is −0.137856 (se 0.006084), 95% normal interval [−0.149781, −0.125932], excluding 0 → the collapse is resolved. The absolute gap interval is [−0.026279, +0.000585], which contains 0 → an actual preference inversion is NOT_ESTABLISHED. Underpowered here is not equivalence, and a point estimate below zero is not a verdict.
8. The B3-style score is blind to the collapse. The identical fitted models, scored on the eligible (uncensored-only) subpopulation, report +0.197343 (se 0.002410) — the true family winning by more than the oracle does. A score computed only on the surviving subpopulation cannot detect a mis-specification that lives in the removed tail. This is the single most actionable receipt in this probe.
9. At the real-like window there is no collapse. W=60 drop-censored keeps the gap at +0.133403 (se 0.002203), statistically indistinguishable from the oracle's +0.125009. In this synthetic world, at the frozen cohort's actual censoring exposure, the exclusion does not reorder the families.

3.4 T5 — state-specific censor sensitivity

Because scale_N ranges from 3.4826 s (state 2) to 24.5206 s (state 8), a fixed observation window censors high-N states far more. Realised per-state censoring rates, R = 40:

The rate tracks scale_N, not the state index — a 157× spread at W=60 (state 2 → state 8). Censoring is therefore a state-confounded nuisance: any state-dependent structure inferred from a windowed dataset is partly a censoring gradient.

Single-draw stratified fit (separate configuration: 200 motors × 8 events, true scale, R = 1, no uncertainty, indicative only — not evidential):

The drop-censored bias flips sign between strata at W=20: upward (lighter tail) in low-N, downward (heavier tail) in high-N. A single pooled shape estimate averages two opposite biases.

Methodological note preserved. A first pass used a quantile-based 5% trim. At W=20 it removed zero observations and T4 collapsed byte-for-byte onto T3, because 17–18% of observed durations pile up exactly at the window boundary, putting the 95th percentile at the window. Upper-tail trimming is inoperative once a hard window has already truncated the tail. The reported T4 uses a rank-based trim instead. This is a real property of trimming windowed data, not a fixed bug.

4. The builder question, answered

Does mistreating censoring manufacture an apparent heavy tail?

Conditionally yes, and the condition is not met by the frozen cohort.

• Yes in one specific regime: dropping censored dwells while the normalisation is held at the untruncated scale, at ~17% censoring, moves the shape to dk = −0.032968 and inflates the lognormal sigma by +0.093467 — an apparently heavier tail (Mode A, W=20).
• No in the regime the frozen pipeline actually occupies: under the frozen scale_N rule the bias runs the other way in all eight Mode B cells (dk positive, dsigma negative) — an apparently lighter tail. The two handles partially cancel, and the normalisation handle wins.
• At the real-like exposure (W=60, matched to the real cohort on both 5.0% event rate and 32.8% removed time mass) the family ordering is untouched: gap +0.133403 (se 0.002203) vs oracle +0.125009 (se 0.001874).

4.1 Relevance to the retained M2-over-M3 adverse result — stated as a hypothesis, and

the probe's own evidence is against it

The target hypothesis would be: the M2_LOGNORMAL-over-M3_TWO_TIMESCALE held-out advantage is partly an artifact of the frozen cohort's exclusion of right-censored dwells.

This probe does not support that hypothesis, and supplies one datum against it. In a world whose truth is a Weibull, at the censoring exposure that matches the real cohort, the frozen exclusion did not reorder the families (finding 9), and the induced bias pointed toward a lighter apparent tail (finding 6) — the wrong direction to explain a lognormal advantage.

Three limits on that negative reading, all of which must travel with it:

1. The generator is a Weibull, not a two-timescale mixture. M3's structure is not represented. Whether the exclusion differentially damages a mixture is NOT_COMPUTED here.
2. The real censoring is structurally different from a fixed window. The real data has exactly one censored dwell per motor — a terminal segment, median 89.26 s, longest in that motor 46% of the time. A fixed window censors independently per event. The two mechanisms match on rate and on removed time mass but not on within-motor structure, and the motor is the experimental unit.
3. The frozen adverse result already has a stronger explanation on the table. B4C02 established that the M2-over-M3 result is GENERATOR-SPECIFIC (gensWithM2overM3 = 1 of 3), refuting the frozen GENERATOR-ROBUST_ADVERSE reading. That is a landed result on real machinery; this is a synthetic caricature. B4C02 governs.

Nothing in §4.1 changes any frozen verdict. The M2-over-M3 adverse result is retained exactly as it stands.

5. Claim impact

Bounded, and the bound is tight.

• The F-side model's censoring correctness is currently carrying no observable weight. The frozen cohort contains no censored event, so the censoring branch has never fired on real data. Its held-out advantage — M0 +0.115291 RESOLVED_ABOVE, M5 +0.031036 RESOLVED_ABOVE, everything else NOT_ESTABLISHED — is entirely attributable to the hierarchy and the Weibull hazard shape, not to censoring handling. Any narrative that credits the censoring-correct likelihood for the F-side result is unsupported.
• "Its advantage vanishes under every treatment" is not what was measured. At the real-like exposure the likelihood treatment is immaterial (≤ +0.001313 nats, ~32× below the 0.042 floor), so the F-side advantage is neither created nor destroyed by censoring handling. It is untested on this axis, which is a different status from tested-and-null.
• The result that does bind: at the real-like exposure, the frozen scale_N rule costs +0.035502 to +0.043165 seconds-NLPD against the full untruncated population — three of four cells at or above the 0.042 resolution floor — and the B3-style eligible-subpopulation score cannot see any of it (finding 8). This is a statement about the estimand the frozen leaderboard targets, not about any model's correctness on that estimand.
• The frozen leaderboard is valid for what it measures. It scores the tail-depleted eligible subpopulation and it does so correctly. It is not a leaderboard on the untruncated dwell-time distribution of the motor, and no result in this repository should be read as one.

Scoping, per contract. P6 for duration-only B3/B4 is unchanged by this probe. P6 for C11 U4 remains withdrawn pending the corrected run. No P-level moved. The first unsatisfied level remains P4 transfer.

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

Receipt that would exercise the censoring branch on real data (the only one available): build a new, separately named cohort — not the frozen one — that retains the 54 in-band right-censored events; recompute scale_N on it (it will differ, so it is a different cohort by construction); fit the F-side motor stack with hazard_survival.log_event_density and the real censor flags; and pre-register the held-out comparison in a commit that is a strict ancestor of the result commit, or the receipt is retrospective from birth (D9).

Falsifier that would kill the lane. If the censored-inclusive cohort's F-side held-out contrasts, motor-equal and CI-bound, are NOT_ESTABLISHED against M1_WEIBULL with a half-width below ~0.042 nats, then the censoring-correct likelihood is not carrying the mechanism on this dataset and should be reported as machinery that is correct but not load-bearing.

What would kill the §4.1 hypothesis outright. Re-run §3 with a two-timescale generator matching the frozen M3 parameters (w = 0.393356, lambdaFast = 0.444859, lambdaSlow = 5.239879) and a per-motor terminal censoring mechanism instead of a fixed window. If the M2-vs-M3 ordering is unchanged at 5.0% event / 32.8% time exposure, the censoring-artifact explanation is dead. NOT_RUN — COMPUTE_BUDGET in this probe (two long runs are in flight; this needs a replicated mixture-fitting sweep, which is not a sub-two-minute single-process job).

Threshold hygiene. The two observation windows (60 s, 20 s) and the 5% trim fraction are DESIGN_ONLY and carry no evidential weight. The 0.042-nat resolution floor is the frozen corrected motor-equal BCa half-width (M4_MIXTURE_K3, BCa width 0.08414086126525253) and is used here only as a materiality yardstick, not as a pass/fail criterion. No threshold was introduced.

7. Reproduction

cd [redacted: operator-path]/Documents/UNI-Flagellum/UNI-FLAGELLUM

# Section 1 - OBSERVED real-data counts (duration channel only)
python -c "import sys; sys.path.insert(0,'hierarchical-aif/src'); \
from motor_stack_aif import _bridge; raw=_bridge.b3().load_events(); \
print(len(raw), sum(1 for e in raw if e['rightCensored']))"

# Sections 2-3 - SYNTHETIC (scripts held in the session scratchpad, not committed)
python probe6_censoring.py   # v1, single draw, 2.4 s   - T5 stratified + trim degeneracy
python probe6_v2.py          # R=40, both modes, 14.5 s - tables 3.1 / 3.2
python probe6_v3.py          # R=40, estimand split, 11.4 s - table 3.3

Scratchpad location: [redacted: operator-path]\AppData\Local\Temp\claude\C--Users-mpolz-Documents-UNI-Flagellum\42f0792e-ab52-4a4d-8991-5e9fdc602268\scratchpad\ (probe6_censoring.py, probe6_v2.py, probe6_v3.py, and the three JSON outputs). These are probe scripts, not repository machinery; nothing under src/, results/ or audits/ was modified.

8. Limitations and what was NOT computed

• NOT_MEASURED — any real-data censoring comparison. Structurally impossible: the frozen cohort has zero censored events. Only the consequence of the exclusion is observable.
• NOT_COMPUTED — a two-timescale (M3-structured) generator under censoring stress. Reason: COMPUTE_BUDGET, see §6.
• NOT_COMPUTED — a per-motor terminal censoring mechanism matching the real one-censored- dwell-per-motor structure. The fixed-window caricature matches on rate and time mass, not on within-motor structure.
• NOT_COMPUTED — the hierarchical (mu, tau) F-side model under censoring stress. §3 uses a single shared shape so that parameter recovery is unambiguous. Whether the hierarchy absorbs or amplifies censoring bias is untested.
• NOT_COMPUTED — a joint ML fit of per-state scale and shape under censoring. Mode B uses the frozen plug-in rule because that is what the pipeline does; a censoring-corrected scale estimator was not built.
• NOT_CHECKED — any held-out mark quantity. Would require holdout nextStateN/direction/ jump access, which the D5 firewall forbids. This is the correct answer, not a gap.
• The synthetic intervals in §3 are normal-approximation intervals over 40 replicate draws. They are not the frozen motor-cluster BCa bootstrap and must never be compared to a frozen verdict interval.
• All §2–§4 numbers are simulation output. They are model output and are never OBSERVED. Only §1 is OBSERVED.

Probe status: COMPLETE. No P-level moved. No frozen verdict changed. No claim created.

NEXT_ACT = Draft hierarchical-aif/protocols/CENSORED-INCLUSIVE-COHORT-PREDICTION.md pre-registering, BEFORE any fit is run, the F-side held-out contrast on a new separately-named censored-inclusive cohort retaining the 54 in-band right-censored events, with the falsifier from §6 written into the protocol and the prediction commit made a strict ancestor of the result commit.