D5 — Held-Out Mark Channel Burned

Defect: D5_HOLDOUT_MARK_CHANNEL_BURNED · Date: 2026-07-21 · Gate: H-AIF-G2 Status: VERIFIED_IRREVERSIBLE · Class: process defect, not a biological result Cause: the builder (me). Not the reviewer, not a subagent acting outside its brief.

1. What happened

While running UltraCode design tracks in parallel with the corrected B4 runs, I briefed a read-only track (Track C) to plan how the mark process — nextStateN, direction, jump — could be integrated into a motor-stack likelihood.

To answer that brief the agent read the mark fields on the 19 holdout motors, computed per-state mark marginals, and ran a held-out contrast (pooled vs state-conditional transition kernel, motor-equal NLPD, 2000-replicate motor-cluster bootstrap, four smoothing constants).

No file was written. No frozen artifact was touched. The agent obeyed its brief exactly. The channel was destroyed anyway, because reading held-out data is itself the irreversible act.

2. The exact instruction that caused it

From my Track C brief:

"1. What are the empirical marginals of direction/jump per state? How many events have a non-null mark? What happens at state boundaries (state 0, state 11)?"

That question has no train/holdout qualifier. It asks for marginals of the whole dataset. A competent agent answering it will read the holdout. The brief contained no split boundary, no forbidden-field list, and no data-access protocol. That omission is the root cause.

3. Fields burned

4. Fields already spent BEFORE D5 (not attributable to this incident)

I initially assumed the mark was entirely unused. That was wrong: direction had already been consumed by a committed gate. Only the mark magnitude (nextStateN, jump) was unspent, and that is precisely what D5 destroyed.

5. What is unaffected

• B3 duration scoring is unaffected. B3 reads durationS, stateN, rightCensored only. It never reads the mark (b3-model-competition-runner.py:109-110 filters on rightCensored and stateN). The leaderboard, the adverse M2-over-M3 headline, and every B3 interval stand.
• The corrected B4 runs are unaffected. B4C02, B4C10, B4C11 and B4C01 read no mark field. The runs currently executing are not contaminated.
• All B4 cells C03–C08 are unaffected.
• Package integrity receipts are unaffected.
• D1–D4 findings are unaffected.

6. What is now retrospective only

Any mark-process model on Wadhwa-2022 — transition kernels, direction/jump-bearing likelihoods, mark-conditioned hazards, or any motor-stack Lmotor-0 term including p(N', direction, jump | ...) — must be labelled:

RETROSPECTIVE_EXPLORATORY_ON_THIS_DATASET

Such work is still permitted and still useful. It simply cannot be presented as prospective held-out evidence, and cannot move P3 on that basis.

7. Claim impact

• No claim moves up. No P-level rises.
• Prospective mark-process mechanism evidence now requires an independent dataset with a new prospective split.
• D5 is a process defect, not a biological result. It says nothing whatsoever about flagellar motors. It is a statement about how this laboratory handled its own evidence.

Mitigating fact, recorded but not offered as an excuse: the held-out contrast the track ran came back NOT_ESTABLISHED — the CI crossed zero at every smoothing constant α ∈ {0.1, 0.5, 1.0, 2.0}, and the point estimate changed sign between α = 0.1 and α = 0.5. The channel had very little resolving power at 19 holdout motors. That bounds the practical cost of the loss. It does not undo it, and it would not have excused the omission had the result been decisive.

8. Existing P-level impact

No level rises. P8 unchanged: FULL_PARITY = false.

9. Process rule added (now binding)

Any brief that could cause an agent to read held-out data must declare the split explicitly and restrict analysis to the training partition, unless a prospective record is committed first.

Supporting machinery, now in force:

1. hierarchical-aif/docs/DATA-CHANNEL-SPEND-LEDGER.md — every field × split, who spent it, what it can still support.
2. A mandatory split-boundary label on every analysis, from a fixed vocabulary.
3. hierarchical-aif/protocols/<TASK>-DATA-ACCESS-PROTOCOL.md required before any task that might touch held-out fields — declaring fields to read, split, forbidden fields, claim boundary, prospective status, falsifier, stop condition.
4. The D5 firewall text is now pasted verbatim into every subagent brief that touches this repository, instructing agents to answer NOT_CHECKED — would require holdout access rather than read. That answer is explicitly framed as correct and valuable, never a failure.

10. Future prevention

• The firewall clause was applied immediately to the Track D/E/F verification workflow launched after D5 was discovered.
• Verification agents are now required to emit a split-boundary label on every claim.
• Recommended follow-on (not yet built): a check that greps analysis scripts under hierarchical-aif/ for reads of partition == "holdout" and fails unless a matching -DATA-ACCESS-PROTOCOL.md exists.

11. Honest summary

I built a firewall after driving through the wall. The rule in §9 is the rule that would have prevented this, and it did not exist because I did not think to write it until an agent had already spent the channel. The loss is permanent within this dataset, it is bounded, and it is recorded here rather than absorbed.