SallyIP

Home / No evidence, no assertion

No evidence, no assertion

SallyIP retrieves evidence first and generates second. Five gates stand between a question and an answer — and zero results is a valid result.

The five gates

  1. Hybrid retrieval — lexical plus vector plus jurisdiction-pack lookup with reciprocal-rank fusion and relevance thresholds.
  2. Exact-quote verification — every generated quotation checked word-for-word against retrieved passages; exact, fuzzy or missing recorded per quote.
  3. Citation-integrity guard — every source label must map to retrieved evidence for the current answer; dangling labels are removed and the proposition marked unverified.
  4. Entailment grading — each proposition-citation pair graded entails, partial, context, contradicts or does not support.
  5. Answer modes — VERIFIED, QUALIFIED, or RESEARCH REQUIRED, with workflow runs, verification events and citation ledgers audit-logged.
What this does not prove. Passing the gates means the answer is grounded in real sources — not that its legal conclusion is correct. The frozen P0 run showed this precisely: 100% citation integrity of scored answers alongside 66.7% entailment (FAIL) and 81.0% unsupported-proposition (FAIL). Both halves are published.

Audit trail

Every answer carries its provenance: workflow runs record what executed, verification events record what each gate decided, and the citation ledger ties each displayed source label to the exact passage retrieved for the current answer. The Verify panel exposes this trail for inspection rather than asking for trust.

All benchmark results · Citation verification · Legal AI hallucination

Run references: adversarial-v1 live run 3368daf1 (12 items, 2026-09-10); stanford-24 run fe26da44 (2026-09-09); grounding-100 eval_runs 2026-09-09; frozen P0 full report 2026-09-09; ablation simulation 2026-09-10. Reports: benchmarks/cross-bench-report.md, benchmarks/regression_report_p0_full.md, benchmarks/ablation-report.md.