Mark-Process Transfer Rescue Protocol

Type: BUILDER-SUPPORT DESIGN PROTOCOL. This document moves no P-level, changes no frozen verdict, and creates no claim. It answers one question: what receipt would make the mark-process lane true, and what would kill it? Lane: LANE C (mark process). Gate touched: H-AIF-G8 (design only). Split boundary of this document: NO_DATA_ACCESS_NEEDED — no dataset field was read to write it. Every quantity below is quoted from an existing recorded artifact. D5 firewall compliance: held-out nextStateN, direction and jump were not read while writing this protocol. Everything cited comes from reports/D5-HOLDOUT-MARK-CHANNEL-BURNED.md, reports/D6-INGEST-NEXTSTATE-RANGE-CHECK-DEFECT.md, docs/DATA-CHANNEL-SPEND-LEDGER.md and ledgers/HIERARCHICAL-AIF-DEFECT-CLOSURE-LEDGER.md.

0. Status, stated up front

D5 and D6 do not kill the mark process. They reclassify it. The three standing statements:

1. Wadhwa-2022 nextStateN / jump mark claims are RETROSPECTIVE_EXPLORATORY_ON_THIS_DATASET — retrospective-only on this dataset. The held-out mark channel was spent by a read-only audit track (D5) with no prospective record in existence at the time. Reading is the irreversible act; nothing was written and the channel was destroyed anyway.
2. A prospective mark-mechanism receipt now requires transfer or a new split (INDEPENDENT_TRANSFER_REQUIRED). This dataset is one study with one holdout partition and there is no second holdout to spend. No modelling in this repository can supply it.
3. The closed mark chain is BLOCKED until the nextStateN range defect is resolved. D6 records 2 holdout events carrying nextStateN = -1, 5 holdout events with zero training support under an unsmoothed (N,N') kernel, and 15–17% of marks leaving {1..8}. The raw MAT archive that would discriminate step-fitting artifact from ingest defect is absent.

Nothing in this protocol asserts that the mark process has been prospectively validated on Wadhwa-2022 — that specific wording is forbidden and would be false.

1. Recorded facts this protocol is built on (no re-derivation)

The mitigating fact from D5 §7 is recorded, not offered as an excuse: at 19 holdout motors the mark channel had very little resolving power. That bounds the cost of the loss. It does not undo it, and an inconclusive interval is NOT_ESTABLISHED — it is never a statement that the pooled and state-conditional kernels are equivalent. Underpowered is not equivalence.

2. RAW ARCHIVE RE-DERIVATION

Required. Obtain data/remodeling_data.mat (the Wadhwa-2022 raw step-fitting archive), then re-run scripts/ingest-wadhwa-data.py against it under a fresh output path — the committed experiments/data/wadhwa-2022-events.json is the historical record and is not edited. Compare the re-derived extraction to the committed one event-by-event. The single discriminating check is the two below-physical-minimum target-state mark events identified in D6 §1 (event identifiers D12-redacted; the raw-archive re-derivation script resolves them internally against the committed dataset without reproducing them in this document):

• if the raw archive also carries a transition into a negative stator count → the step fitting produced an impossible transition, and the recorded value is faithful to the source;
• if the raw archive carries a legal target and the -1 appears only after ingest → the ingest mis-propagated the outOfRangeDwells exclusion into the predecessor's mark.

Licenses: a P0 provenance statement about which of the two mechanisms produced the impossible value; the removal of a blocking ambiguity from the D6 closure row; a documented, source-pinned handling decision for those two events.

Does NOT license: any prospective mark claim. Re-deriving the extraction does not un-spend the held-out mark channel — D5 is about reading, not about file contents. It also does not license editing the committed dataset, nor re-labelling any re-derived field OBSERVED beyond what the raw archive itself pins.

Falsifier: re-derivation shows the committed extraction is faithful and the raw archive itself contains the negative target. Then the reflecting-boundary assumption P(jump < 0 | N = 0) = 0 is contradicted by the source measurement, not by a pipeline artifact, and every mark model that enforces that boundary is refuted at that event rather than rescued.

Status today: BLOCKED_EXTERNAL — archive absent, verified.

3. RANGE-CHECK REPAIR

Required. A validation layer that refuses to let an out-of-physical-range mark pass silently into any mark model. Physical bounds are PHYSICAL_MIN_STATORS = 0, PHYSICAL_MAX_STATORS = 11 (src/motor_stack_aif/marks.py:20-21). The repair is detection plus a forced decision, never a correction of the recorded value.

Already built and executable in this repository today. flag_impossible_marks() returns one record per offending event carrying eventId, motorId, stateN, nextStateN, jump, partition and a reason from {NEXT_STATE_BELOW_PHYSICAL_MINIMUM, NEXT_STATE_ABOVE_PHYSICAL_MAXIMUM}. The regression receipt tests/motor_stack_aif/test_nextstate_range_check.py includes test_the_known_defect_is_still_present_and_unedited, which fails if anyone repairs the defect by editing the frozen dataset.

Licenses: a mark pipeline that cannot ingest a physically impossible stator count without an explicit, recorded operator decision; a P1 implementation-integrity statement about the mark path.

Does NOT license: any statement about biology. A range check is a software invariant. It says nothing about whether stator remodelling is Markovian, state-dependent, or mechanistically interesting. It also does not license patching scripts/ingest-wadhwa-data.py and re-emitting the frozen dataset in place — the defect is preserved on purpose.

Falsifier: the guard fails to fire on a synthetic event carrying nextStateN = -1 or nextStateN = 12; or the frozen dataset stops containing the two known defect events, which would mean the record was edited rather than annotated.

Status today: EXECUTABLE_NOW — BUILT AND TESTED (7 tests, 1.26s).

4. QUARANTINE POLICY

Required. Every mark analysis must declare exactly one impossible-mark policy in its own artifact, from the closed set in marks.py:23:

There is deliberately no policy that drops the events silently — prepare_mark_dataset raises ValueError on any policy name outside the three. The quarantined count is part of the result, not a footnote: an artifact that reports a score without reporting how many events were quarantined is incomplete.

Licenses: a reportable mark analysis whose event exclusions are auditable; comparison between two mark models only if both declare the same policy.

Does NOT license: comparing a quarantine run against a retain_labelled run as if the difference were a modelling result. Different policies score different event sets, so the contrast is confounded by the exclusion, not by the model. It also does not license treating quarantine as a repair — the underlying ambiguity stays open until §2.

Falsifier: a mark analysis is found that passes marks.py yet silently changes its scored event count; or two published mark contrasts are found to differ in policy while being compared head-to-head.

Status today: EXECUTABLE_NOW — BUILT AND TESTED (covered by test_preparation_refuses_silent_drop, test_quarantine_policy_preserves_and_labels, test_documented_defect_policy_retains_with_label, test_unknown_policy_is_rejected).

5. PREDECLARED SMOOTHING

This is the section that matters most, and the ordering is the whole point.

An unsmoothed (N,N') kernel HALTs: 5 of 233 holdout events have zero training support, each producing log p = -inf under the no-floor policy. So smoothing is not a stylistic choice — it is mandatory, and it is outcome-determining. The recorded evidence: in the Track C held-out contrast the point estimate changed sign between α = 0.1 and α = 0.5, while the interval crossed zero at all four constants α ∈ {0.1, 0.5, 1.0, 2.0}.

A hyperparameter that decides the direction of a result is not a nuisance parameter. If α is chosen after the outcomes are visible, the analyst — not the data — picks the sign of the headline. Predeclaring α does not make the analysis more accurate; it makes the analysis falsifiable. That is the entire function of the ordering: with α fixed in advance, an adverse result cannot be tuned away, and a favourable result cannot have been manufactured by selection. With α fixed afterwards, the contrast has no evidential content at all, however narrow its interval looks.

Required, in this order, before any new mark analysis:

1. Write and commit a prediction record naming: the smoothing family (e.g. additive-α Dirichlet / Laplace on the (N,N') count table), the exact constant α, the alphabet the counts are laid over, the impossible-mark policy from §4, the scoring rule, the resampling unit (MOTOR), the seed, and the falsifier — with zero observations of the new outcome in existence. Prospectivity is decided by the commit graph, not by prose: the prediction commit must be a proven strict ancestor of the result's introduction.
2. If α cannot be justified a priori, predeclare a sensitivity grid and predeclare the reading rule for it. The honest reading rule is: if the sign of the contrast is not stable across the whole predeclared grid, the result is reported as NOT_ESTABLISHED — SMOOTHING_DEPENDENT, and no direction is claimed. Declaring the grid after seeing the grid is the failure mode this clause exists to prevent.
3. Only then run.

Licenses: a mark contrast whose sign can be trusted as coming from the data, on whatever split its data-access protocol declares.

Does NOT license: promoting a predeclared-α result on this dataset to prospective status. Predeclaring α repairs the hyperparameter selection problem; it does not repair D5. On Wadhwa-2022 the held-out mark channel is already spent, so even a perfectly predeclared analysis there remains RETROSPECTIVE_EXPLORATORY_ON_THIS_DATASET. The two defects are independent and both must be cleared.

Falsifier: a mark result is published whose α was fixed after any outcome was visible; or the predeclared grid is found to have been widened, narrowed, or re-centred between the prediction commit and the result commit.

Status today: EXECUTABLE_NOW as a procedure (the prediction record can be written and committed today) but evidentially inert on this dataset — see §6.

6. INDEPENDENT DATASET REQUIREMENT

Required. A stator-remodelling dataset that is not Wadhwa-2022, carrying dwell state, dwell duration, and the transition target, with:

• a new split defined and committed before any held-out field is read;
• enough motors — the experimental unit is the motor, and 19 holdout motors gave the mark channel very little resolving power; the required count must be justified from the effect size a mark model is expected to produce, not chosen after seeing an interval width;
• a declared <TASK>-DATA-ACCESS-PROTOCOL.md naming fields to read, split, forbidden fields, claim boundary, prospective status, falsifier and stop condition, per the D5 process rule;
• provenance sufficient to pin the transition targets to a recorded measurement, so OBSERVED is earned rather than asserted.

Licenses: the only route to a prospective mark-mechanism receipt. This folds the mark question into the P4 transfer and P7 independent-replication requirements the project already carries — it does not create a new requirement, it strengthens two existing ones.

Does NOT license: anything retroactive. A clean second dataset does not restore prospective status to Wadhwa-2022; the Wadhwa mark result would become, at best, a retrospective consistency check against the new prospective one. It also does not license averaging the two studies into one verdict — cross-study pooling is its own gate with its own leakage risks.

Falsifier of the transfer itself: the mark structure that fits Wadhwa-2022 fails on the independent dataset under the predeclared α and the predeclared scoring rule. That is a real, informative outcome and must be reported as such, not absorbed.

Status today: BLOCKED_EXTERNAL — no such dataset is in this repository.

7. ONE-STEP CONDITIONAL MARK LIKELIHOOD

Required. Because 15.1% of training and 16.7% of holdout marks point outside {1..8}, the observed process is not a closed chain on the cohort. The only structurally honest object available on this extraction is a one-step-ahead conditional: p(N' | N, ...) scored on events whose originating state is in the cohort, with the target alphabet extended to cover the observed escape targets ({-1, 0, 9, 10, 11} alongside {1..8}) or with the escape mass assigned to an explicit OUT_OF_COHORT absorbing symbol. Either choice must be declared; both are legitimate; a third option — discarding the escape mass — is not, because it inflates the remaining likelihood.

The artifact must say, in its own words, that it is a one-step-ahead conditional prediction and not a trajectory likelihood.

Licenses: a scoreable mark-bearing term for the motor stack; a like-for-like contrast between a pooled kernel and a state-conditional kernel under the same alphabet, same policy, same α, same motor-cluster resampling; a P1/P3-shaped statement scoped to whatever split its data-access protocol permits.

Does NOT license: any statement about long-run occupancy, stationary distribution, mean first passage, or trajectory probability. Those are properties of a closed chain, and this object is not one. It also does not license reading a one-step conditional advantage as mechanism — predictive superiority is never promoted to mechanism anywhere in this program.

Falsifier: the escape fraction is found to be materially different from the recorded 15.1% / 16.7% under the declared cohort and policy, which would mean the alphabet was mis-specified; or the scored event count does not equal the declared eligible count minus the declared quarantine count.

Status today: EXECUTABLE_NOW on the training partition only (TRAIN_ONLY). On the holdout it is HOLDOUT_MARK_CHANNEL_BURNED_RETROSPECTIVE_ONLY.

8. CLOSED-CHAIN LIKELIHOOD ONLY IF VALID

Required. A closed-chain (trajectory) likelihood over {1..8} may be written only when all of the following hold simultaneously:

1. §2 raw-archive re-derivation has resolved the two impossible marks, or the cohort is redefined so that the escape mass genuinely vanishes;
2. assert_closed_alphabet(events, states=(1..8), partition=...) returns without raising — it currently raises, by design, and test_closed_chain_assumption_is_refused locks that behaviour in;
3. the impossible-mark policy from §4 is declared and the quarantined count is reported;
4. the smoothing constant from §5 is predeclared and committed.

Until all four hold, a closed-chain likelihood is invalid on this extraction, not merely imprecise. assert_closed_alphabet is the executable gate: it is not a warning, it raises OpenMarkAlphabetError and stops the analysis.

Licenses (only once valid): trajectory-level statements — occupancy, dwell-sequence probability, first-passage structure over the closed alphabet.

Does NOT license: trajectory statements obtained by narrowing the cohort until the escapes disappear, unless that narrowing was predeclared and its biological meaning is defended. Shrinking the state set to force a closed alphabet is alphabet-shopping, and it produces a chain that is closed by construction rather than by observation.

Falsifier: assert_closed_alphabet raises on the cohort a published closed-chain result claims to cover; or the escape targets recorded in D6 are present in the scored data of any artifact that asserts closure.

Status today: BLOCKED — conditions 1 and 2 are not met, and condition 1 is BLOCKED_EXTERNAL.

9. Executable-today ledger

10. What this protocol explicitly does not do

• It does not raise P2, P3, P4, P6, P7 or P8. No P-level is moved by this document.
• It does not restore prospective status to any Wadhwa-2022 mark field. Reading is irreversible.
• It does not re-threshold, re-open, or re-interpret any frozen B3/B4 verdict. B3 duration scoring never reads the mark; no B4 cell reads the mark; the F-side duration model does not read the mark.
• It introduces no new evidential threshold. The physical bounds 0..11 are the physics of stator occupancy, not a tuned criterion. Any α grid a future analysis predeclares is DESIGN_ONLY until it is committed inside a prediction record.
• It contains no claim that the mark process has been validated on any dataset.

11. Scoped standing statements (never unscoped)

• P6 for the Wadhwa mark-process mechanism is retrospective-only and transfer-required (D5, D6). P6 for duration-only B3/B4 is unchanged.
• P2 for mark fields carries the D6 limitation. P2 for duration fields is unchanged.
• P3 duration-only held-out evidence is untouched by D5 and D6.
• P4 and P7 requirements are strengthened, because transfer is now also the only route to a prospective mark receipt.

NEXT_ACT = Write and commit hierarchical-aif/protocols/MARK-KERNEL-TRAIN-ONLY-DATA-ACCESS-PROTOCOL.md declaring split boundary TRAIN_ONLY, forbidden fields (held-out nextStateN/direction/jump), impossible-mark policy quarantine, a predeclared additive-alpha grid with its sign-stability reading rule, motor-level resampling, and the falsifier — committed with zero new outcomes in existence — so that the §7 one-step conditional becomes runnable on the training partition without spending anything.