D6 — Ingest Writes Out-of-Range `nextStateN`

Defect: D6_INGEST_NEXTSTATE_NOT_RANGE_CHECKED · Date: 2026-07-21 · Gate: H-AIF-G2 Classification: INGEST_MARK_FIELD_DEFECT + RAW_ARCHIVE_REQUIRED_TO_RESOLVE_OR_CONFIRM Status: VERIFIED (independently reproduced by the builder) Split boundary of this analysis: the two affected events are in the holdout partition; the mark channel there was already burned by D5, so inspecting them adds no new spend.

1. Finding

Two events record a physically impossible stator count (a below-physical-minimum target-state mark, nextStateN below the physical floor of 0 stators) in the mark field. Both events are in the holdout partition, both from the same holdout motor.

D12 redaction notice. The exact event identifiers, the source motor identifier, and the per-event stateN/nextStateN/jump/direction values were removed from this distributable report per the D12 incident-remediation ruling (event-level held-out mark data must not survive in a shipped artifact). The scientific finding below is unchanged: two holdout events carry nextStateN values below the physical floor. The event-level record used to be reproducible internally via tests/motor_stack_aif/test_nextstate_range_check.py, which is itself excluded from external distribution (see D12 report) because its regression-testing purpose requires keeping the literal identifiers.

2. Root cause — exact source lines

scripts/ingest-wadhwa-data.py:

141            if dwell["state"] < 0 or dwell["state"] > 11:
142                exclusions["outOfRangeDwells"] += 1
143                continue                      # <-- the DWELL's own state IS range-checked
...
147            next_state = dwell["nextState"]   # <-- read with NO range check
...
158                    "nextStateN": next_state,                     # written through
159                    "direction": None if next_state is None else ("on" if next_state > dwell["state"] else "off"),
160                    "jump": None if next_state is None else next_state - dwell["state"],

grep -n next_state returns only lines 147, 158, 159, 160 — there is no range check on next_state anywhere.

Mechanism: the out-of-range dwell itself is correctly excluded (outOfRangeDwells: 3), but its predecessor is retained and keeps a mark pointing at the excluded, impossible state. So the exclusion is applied to the dwell but not propagated to the mark of the event before it.

3. Reproduction

python -c "import json;print([(e['eventId'],e['stateN'],e['nextStateN'],e['jump'],e['partition']) \
 for e in json.load(open('experiments/data/wadhwa-2022-events.json'))['events'] \
 if (e.get('nextStateN') if e.get('nextStateN') is not None else 0)<0])"

4. Consequence — a falsified boundary assumption

A mark model that enforces the physical reflecting boundary P(jump < 0 | N = 0) = 0 assigns log p = −inf to the first below-physical-minimum event named above. Under the runner's declared no-floor policy (b3-model-competition-runner.py: non-finite log density HALTS), such a model cannot be fitted or scored on this extraction without an explicit decision about these two events.

So: the reflecting-boundary assumption at N = 0 is falsified by the recorded data at exactly one event. Either the step-fitting produced an impossible transition, or the ingest mis-propagated an exclusion. This is undecidable here — the raw MAT archive (data/remodeling_data.mat) is absent (P2 observational is already BLOCKED_EXTERNAL for that reason).

5. Related structural constraints on any mark model

Both independently verified by the builder:

Implications:

1. Smoothing is mandatory and outcome-determining. An unsmoothed kernel HALTs. And the smoothing constant is not innocuous: in the Track C contrast it flipped the sign of the headline effect between α = 0.1 and α = 0.5. A hyperparameter that decides the direction of a result must be frozen in advance, not chosen after seeing outcomes.
2. The process is not a closed Markov chain on the cohort. With 15–17% of marks leaving {1..8}, either the alphabet must include out-of-cohort states or that mass is silently discarded and the remaining likelihood is inflated. A mark-bearing likelihood here is a one-step-ahead conditional prediction, not a trajectory likelihood, and must be described as such in its own artifact.

6. Impact on existing work

7. Existing P-level mapping

No level rises.

8. Corrective action

1. Do NOT edit the committed dataset. It is the historical record; the defect is preserved.
2. Any future mark model must declare its handling of these two events explicitly — reject, quarantine with an explicit status, or retain under a documented raw-data-defect label — and must never silently drop them.
3. Regression test added: hierarchical-aif/tests/motor_stack_aif/test_nextstate_range_check.py.
4. If the raw archive is ever obtained, re-derive these two events first; that single check discriminates "step-fitting artifact" from "ingest defect".