iter 8 — volatility regimes are labelled with lookahead QUANTIFIED

Attack item X-2 · 2026-07-29 · harness harness/future_perturbation_invariance.py · draft remedy PRE-REGISTRATION-DRAFT-regime-tercile-lookahead.md

The rule had to be restated before it could be a gate

The operator's standing rule is perturb test rows, train-side discovery must be unchanged — phrased for a predictive walk-forward. These probes are redundancy screens with no train/test split, so the rule does not transfer verbatim, and iteration 7 flagged that waving the analogy through would produce a gate that looks like compliance and tests nothing. The invariant that does transfer:

Perturbing FUTURE rows must not change any label or reading assigned to PAST rows.

Where the leak is

regime_labels() computes volatility with a trailing window — correct — then buckets it against globally computed terciles:

v = vol[idx]
q1, q2 = np.nanquantile(finite, [1/3, 2/3])          # over the WHOLE cell
vol_lab = where(v <= q1, "low", where(v <= q2, "mid", "high"))

A bar's volatility regime is assigned relative to volatility that had not happened yet.

Prediction, and how it fared

I predicted vol_lab would relabel 5–25% of past rows while trend and session stayed at 0%. The direction and the discrimination were right. The magnitude was wrong in both directions, because it turns out to depend entirely on the scenario — which is itself the finding.

Result

Scenario	% of PAST value rows relabelled
Operational — append 5% more bars (i.e. a re-run)	2.25%
Operational — append 10%	2.47%
Operational — append 25%	5.24%
Operational — append 50%	5.24%
Stress — future half replaced by a 2.5× volatility regime	65.74% (max 81.43%)

Controls, which are what make this a measurement rather than an alarm:

sess_lab   0.00%  invariant
trend_lab  0.00%  invariant
xi sub-slice reading on untouched rows: bit-identical

The test discriminates. It is not flagging everything — it isolates the one axis that consults a global statistic.

What is and is not being claimed

The 65.74% is a stress bound and should not be quoted as the operational number. The operational figure is 2–5%: every re-run relabels a few percent of historical value rows simply because time passed. That is smaller, but it is structural and unavoidable under the current rule.

This is not a prediction leak — nothing is being forecast. The harm is irreproducibility of the stratification, and it lands on the campaign's own bookkeeping:

The coverage ledger accumulates across runs. The rotation mechanism is a "coverage ledger driven, deficit-first loop… keeps attacking whichever condition is least-tested until every reachable one is covered N times." If a bar's stratum changes between runs, the ledger sums observations taken under different stratifications — "low-vol covered 20 times" may be a mixture.
Verdicts drift as data accrues, independently of any change in what is being judged.
Coverage claims are snapshot-relative — "tested in 12 of 18 regimes" is true of one fetch.

No verdict in the existing record is invalidated by this alone. Within a single run the stratification is internally consistent; the defect is about comparability between runs.

Not fixed, and not wired into check-full

Four remedy options are drafted — expanding-window terciles, fixed absolute thresholds, frozen per-cell terciles from a declared reference window, or accept-and-document. Each carries a decision: the first three change every historical stratum assignment and therefore invalidate all accumulated coverage. That is the operator's call, not an unattended loop's.

The harness is runnable as /xi:future-perturbation but is deliberately not added to /xi:check-full. Wiring a currently-failing gate into the suite would turn the whole suite red and force the decision by making CI unusable — which is coercion, not reporting.

Limits

Synthetic price path built by resampling returns from the committed BTCUSDT fixture. Real cells have autocorrelated volatility, which will change the numbers in an undetermined direction — clustered volatility could make terciles either more stable (long regimes) or less (regime shifts).
The append-sweep plateaus at 5.24% for both 25% and 50% appended, which is suspicious enough to name: it suggests saturation of the tercile shift rather than a smooth dose-response, and it was not investigated further.
Only vol_lab was traced to its source. Other global statistics may exist elsewhere in the pipeline; this iteration establishes one leak, not the absence of others.

Evidence: harness/future_perturbation_invariance_evidence.json. Reproduce: mise run xi:future-perturbation