Copy this file to audits/<YYYY-MM>_<target>/report.md and fill every section. A report with an empty or failing positive control is not published.

---

Incomplete-cross-section leakage screen — <TARGET NAME>

Target	<project / dataset name> + link
Version / vintage	<tag, release, or the exact data vintage screened>
Screened on	<YYYY-MM-DD>
Screened by	<name>
Screen version	pit-release-gate <version> (or: inline implementation, see §Reproduce)
Verdict	benign (by construction) / benign (measured) / susceptible (details)

One-paragraph finding. State the result before the method. What was screened, what came out, and the single most important caveat. A reader who stops after this paragraph should not end up with a wrong impression.

---

1. Data context

• What the cross-section is — the entities, the period, and what makes them one group.
• Where arrival times come from — the target rarely ships them, so say exactly how they were reconstructed and from which fields.
• How good that reconstruction is — is it filing staggering or only availability staggering? Which direction does the error go? State the bound explicitly.
• Scale — rows, entities, periods, and the period range actually screened.

2. Method

The screen is one statistic, estimated honestly:

ρ̂ = partial correlation between arrival latency and the complete-cross-section residual, conditional on an observable known before the period opens — fitted on prior completed periods, frozen, then applied to the period being graded.

• Conditioner used and why it is legitimate (must be knowable before the period opens).
• Threshold — the |ρ̂| above which a signal is flagged, and where it comes from.
• Trailing window K, evaluation periods, minimum entities per period.
• Method references — doi:10.6084/m9.figshare.33061955 (screen), doi:10.6084/m9.figshare.33158615 (release controller).

3. Positive control

Mandatory. Run the identical screen on planted-truth cross-sections where the selection strength is a knob you set, and show it classifies all of them correctly. Include at least one benign-but-strongly-selected-on-observables case — that is the one that distinguishes a real screen from a completeness threshold with extra steps.

planted case	ρ̂	φ_req	screen says	should say	correct
clean (c_a=0.0, c_x=0.3)				benign
composition (c_a=0.0, c_x=2.0)				benign
mild leak (c_a=0.3, c_x=0.7)				susceptible
strong leak (c_a=1.0, c_x=0.7)				susceptible

Control result: <n>/<n> correct. If this is not all-correct, stop — the screen is miscalibrated for this data shape and the findings below mean nothing.

4. Results

Per-signal summary

signal	mean ρ̂	max |ρ̂|	periods flagged / screened	mean φ_req	verdict

Per-period detail — one row per (signal, period). Aggregates hide dispersion; show it.

signal	period	n	ρ̂ (frozen)	flagged	φ_req	release point	completeness at release

5. Conclusion and limitations

What was found. The finding, restated with the numbers behind it.

Why. The mechanism. If the verdict is benign (by construction), name the construction — this is the most useful sentence in the report, because it tells a reader exactly which change to their own pipeline would remove the protection.

What this does not license. Be specific about what the reader must NOT conclude. Scope of the screen, of the arrival reconstruction, of the period covered.

Limitations, honestly.

• Coverage: how many signals, how many periods, one arrival proxy?
• The conditioner: is it the one you would have chosen with unlimited data access?
• Conservative choices made and which direction they bias the result.

Suggestion for the target. Optional, and only if there is a concrete, small change that would let users check this themselves — usually an arrival/availability column. Offer help; do not lecture.

6. Reproduce

Every number above must come out of this command.

<exact commands, from a clean checkout, including how to get the data without a licence>

• Runtime: <approx>
• Requires a licence or account: <no — or say exactly what>
• Artifacts in this directory: <notebook / script / cached inputs>

---

Report: CC BY 4.0. Screening code: MIT.