disclosure-check skill

/disclosure-check — Statistical-Disclosure-Limitation pre-screen

Scan analysis outputs built on restricted or confidential data (Census FSRDC, IRS SOI, administrative registers, linked health records, proprietary firm panels) for the disclosure-avoidance problems that get an export request rejected — before it reaches the data provider's official disclosure review. The skill is a pre-screen, not a substitute for that review.

Core principle: A single un-suppressed n=3 cell, an exact count that pins down one firm, or a p-percent dominance failure can re-identify a person or establishment. Catch it on your machine, not in the rejection email from the RDC analyst.

When to use

• Before requesting an export from a Census FSRDC / secure data enclave / RDC.
• Before depositing restricted-data results to openICPSR, a journal, or a co-author outside the enclave.
• Before sharing any figure, table, or log derived from confidential microdata.
• As a release gate. Pair with a pre-commit / pre-deposit invocation so no restricted-data output ships un-screened. This is the foundation of the data-management plan for any restricted-data project.

Inputs

• $0 — outputs directory to scan. Defaults to output/, where every language's pipeline writes; if it does not exist but a pre-v2.6 scripts//outputs/ does, scan that and say so. Also accepts any export-staging directory (e.g., a toreview/ folder the analyst stages for the RDC).
• --provider — selects which disclosure-rule profile to load (Phase 0). One of census / irs / irb / generic. Providers differ — thresholds and rules are not interchangeable; default generic is deliberately conservative.
• --threshold N — override the minimum cell count (default n<10). Census FSRDC commonly uses 10 for establishments; IRS and many IRBs differ. Always reconcile with your provider's written rules.

Workflow

Phase 0: Load the provider's disclosure rules

1. Read .claude/rules/confidential-data.md for the project's restricted-data handling contract and the rule-profile placeholder.
2. Load the --provider profile (a placeholder config the forker fills in from their signed agreement — Census, IRS, and IRB rules differ and supersede any default here):

• min cell count (default n<10),
• dominance rules: p-percent (a cell is unsafe if the largest respondents contribute > p% of the total) and (n,k) (top n units > k% of total),
• rounding required for sensitive statistics (counts, totals, ratios),
• top-coding / bottom-coding thresholds for extreme values,
• geographic minimum population for any geocoded statistic.

1. If no signed-rule values are recorded, fall back to the conservative generic profile and flag prominently in the report that real provider thresholds must be substituted.

Phase 1: Scan the outputs directory

Glob the outputs dir for .tex, .csv, .txt, .log, .smcl, .out, .md tables and figure-data files. For each:

• Cell counts — parse table cells / frequency columns; flag any count 0 < n < threshold that is not already suppressed.
• Complementary-suppression gaps — if one cell in a row/column is suppressed but the margin total and the other cells let a reader back it out by subtraction, the suppression is incomplete.
• Dominance — for any total/mean cell where unit-level contributions are available (or inferable), apply the p-percent and (n,k) rules.
• Exact re-identifying counts — small exact integers (e.g., "4 hospitals", "1 firm", a max/min that is a single observation) that single out a unit.
• PII leakage — regex for names, SSNs (\d{3}-\d{2}-\d{4}), exact dates of birth, addresses, exact lat/long or fine geocodes, record IDs that survived into an output.
• Unrounded sensitive statistics — exact unrounded counts/totals where the provider requires rounding.

Phase 2: Classify each finding — CRITICAL / WARNING / OK

When two findings interact (a suppressed cell + a recoverable margin), report them together — the gate cares about the joint disclosure risk, not each cell in isolation. Be economics-aware: DiD / event-study cell counts per (cohort × period), IV first-stage subsamples, RCT arm × stratum balance tables, and panel firm-counts are the usual offenders.

Phase 3: Suggest remediation

For each CRITICAL / WARNING, propose the standard SDL fix, in order of preference:

• Suppress the offending cell (and its complement, if a margin allows back-out).
• Round counts/totals to the provider's base (e.g., nearest 10 or 15).
• Top-code / bottom-code extreme values.
• Aggregate — collapse thin categories, coarsen geography, widen bins until every cell clears the threshold.
• Drop the statistic if no remediation preserves both safety and meaning.

Each suggestion names the file, the cell/location, the rule it violates, and the concrete edit — never auto-applies it (the analyst owns the disclosure decision).

Phase 4: Gate

Exit non-zero on any CRITICAL. WARNINGs surface but do not block. See Exit behavior.

Output / Report format

Write qualityreports/disclosurecheck_[outputs-dir-slug].md:

# Disclosure Check: [outputs dir]

**Date:** [YYYY-MM-DD]
**Provider profile:** census | irs | irb | generic   (rules source: confidential-data.md)
**Min cell count:** [N]   **Dominance:** p=[p]%, (n,k)=([n],[k]%)   **Rounding base:** [b]

## Summary
| Disposition | Count |
|---|---|
| CRITICAL | M |
| WARNING | W |
| OK | P |
| **Verdict** | **PASS / FAIL** (FAIL iff M > 0) |

## CRITICAL (blocks release)
| File | Location | Rule violated | Observed | Suggested remediation |
|---|---|---|---|---|
| tab3_by_cohort.tex | row "2008", col "n" | min cell (n<10) | n=4 | suppress cell + suppress complement in margin |

## WARNING (human judgment)
| File | Location | Concern | Suggested action |
|---|---|---|---|

## OK
[counts only, or a short list]

## Next steps
1. Resolve every CRITICAL — suppress / round / top-code / aggregate, then re-run.
2. Review WARNINGs with the agreement's written rules in hand.
3. Re-run until zero CRITICAL, THEN submit to the provider's OFFICIAL disclosure review.

Exit behavior

• Zero CRITICAL: exit 0; report printed. (WARNINGs allowed — they are surfaced, not blocking.)
• Any CRITICAL: exit 1; summary to stderr. This makes the skill usable as a release / pre-deposit gate. Mirrors /audit-reproducibility's gate semantics: WARNING ≠ FAIL, only CRITICAL blocks.
• No rules loaded (generic fallback): exit 0 with a prominent warning that real provider thresholds were not supplied — the pre-screen ran but at conservative defaults, not the actual agreement.

Flags

• --provider — Load that data provider's disclosure rules (e.g. census, irs, irb). Default: the generic small-cell ruleset.
• --threshold — Override the minimum cell-count threshold (default n<10); match your data-use agreement's actual rule.

Cross-references

• .claude/rules/confidential-data.md — restricted-data handling contract + the provider-rule profiles this skill loads.
• .claude/rules/replication-protocol.md — for restricted-data papers the replication package ships code + access path, not the microdata; screen every released output first.
• .claude/skills/audit-reproducibility/SKILL.md — numeric paper↔code verification: run it on the retained values, this skill on the released ones.
• .claude/skills/data-analysis/SKILL.md, .claude/skills/stata-replication/SKILL.md — produce the R / Stata / Python outputs this skill screens.
• AEA Data Editor checklist and the DCAS standard — disclosure + access expectations for restricted-data deposits (openICPSR restricted-access stub).

What this skill does NOT do

• It does not replace the data provider's official disclosure review. Census/RDC, IRS, and IRB analysts run the authoritative review; this skill pre-screens so the official review is more likely to pass on the first pass. A PASS here is not clearance to release.
• It does not certify your rules are correct. It applies the thresholds you load from your signed agreement; if the loaded --provider profile is wrong, the scan is wrong. Reconcile with the written agreement, not a default.
• It does not move, encrypt, or transmit data, never exfiltrates microdata from the enclave — it reads only the staged outputs you point it at.
• It does not catch every disclosure risk. Differencing across released tables, longitudinal re-identification, and model-based inferential disclosure can evade a per-file scan. A clean run is necessary, not sufficient.