how lacuna detects omitted facts

lacuna audits what sources must disclose, illuminates silent omissions, and flags discrepancies.

the t3 wording rule (mandatory standard)

“Not found in the sources reviewed is never evidence that something does not exist in the world.”

lacuna surfaces dark spots among the specific documents reviewed. silence on a website does not prove a rule does not exist elsewhere offline or on unindexed portals. all findings are framed as questions to verify with official issuing authorities.

the six-stage staged pipeline

1. planai + template

classifies the problem domain and generates 6 to 12 critical facets from standard policy expectation templates.

target: ≤ 6 s

2. acquiredeterministic code

ssrf-guarded web fetcher and robots-compliant search discovery. extracts cleaned plain text with readability.

target: ≤ 15 s

3. profileheuristic code

heuristic domain classification (.gov, .edu, official) and date extraction from meta tags, json-ld, and visible cycle text.

target: ≤ 3 s

4. extractai + exact verifier

extracts atomic claims mapped to facets. discards any claim whose quote does not verbatim match the stored source text.

target: ≤ 25 s

5. analyzecode comparators + ai

code comparators test dates, numbers, and currencies. distinguishes direct conflicts from version mismatches across cycles.

target: ≤ 15 s

6. actionscritic + checklist

critic generates targeted open questions; synthesizes prioritized verification actions with official source references.

target: ≤ 8 s

the three provenance labels (ux-3)

sourced

sourced (solid border, document icon)

directly asserted by a reviewed source with a verbatim verified quote span.

ai signal

ai signal (dashed border, radar icon)

computed observation or discrepancy identified by comparing two or more verified claims.

ai suggestion

ai suggestion (dotted border, sparkle icon)

synthesized hypothesis, suggested question to ask, or recommendation for follow-up.

coverage score & evidence strength

the coverage score (0 to 100) measures how thoroughly the reviewed sources address the critical facets necessary for an informed decision.

Score = (Weighted Sum of Facet Coverages / Total Facet Weights) × 100

Critical facets: weight 3.0. full credit requires at least one official reviewed source.
Important facets: weight 2.0.
Helpful facets: weight 1.0.

strong: score ≥ 80partial: 50 ≤ score < 80weak: score < 50

golden set benchmark & dark spot recall

To measure detection accuracy objectively, Lacuna evaluates against 3 curated golden fixtures containing 13 seeded dark spots across distinct real-world problem domains (PRD §13).

Scholarship

100%

5 of 5 seeded spots detected (0 false positives)

Government Service

100%

4 of 4 seeded spots detected (0 false positives)

Product Policy

100%

4 of 4 seeded spots detected (0 false positives)

Aggregate recall: 100% (target ≥ 80%) with 0 false positives across 13 seeded findings.

system guardrails & limits

• source pool cap: maximum 8 sources reviewed per analysis (user urls plus auto-discovery).

• document size cap: maximum 2 mb per source. oversized html is truncated to readable text.

• paste limit: maximum 30,000 characters per pasted text snippet.

• retention policy: live analyses automatically expire and are purged after 7 days (nfr-7).

• plain text sandbox: source html is never rendered in the client; all stored text is sanitized plain text (nfr-4).