Evidence Eligibility Gate
Retrieved material is not automatically evidence. Put a gate between intake and reasoning that checks whether a source is admitted, current, approved, sensitivity-appropriate, and permitted for this actor, claim, and use. Only eligible evidence may ground a conclusion.
Problem
AI-assisted and automation-heavy systems often treat retrieval as permission. If the search layer found it, the model may use it. If a feed delivered it, the workflow may cite it. If a document is in the corpus, the answer may lean on it.
That collapses transport, availability, and authority into one false state. The material may be real but stale, unauthorized, pending review, sensitive, out of scope, poisoned, or licensed for a different purpose.
Context
Use this pattern when conclusions, recommendations, approvals, security decisions, publication decisions, or delegated actions depend on retrieved or ingested evidence. It is especially important when the same corpus contains drafts, pending intake, restricted material, stale snapshots, contradictory sources, or sources approved for one use but not another.
Forces
- Retrieval latency rewards pulling first and asking questions later.
- Reasoning needs enough context to be useful.
- Sensitive or pending material may be present in the system for review but forbidden for grounding.
- Freshness and approval can change after a source was indexed.
- Users often interpret citations as proof even when the source was ineligible.
- Blocking a conclusion is frustrating, but using ineligible evidence is worse.
- The system needs a repair path when evidence fails eligibility.
Solution
Install an evidence eligibility gate before reasoning. The gate evaluates source admission, snapshot identity, source health, freshness, approval state, sensitivity, license or permitted use, actor rights, and claim scope. If the material passes, it can enter the evidence bundle and ground retrieval-grounded reasoning. If it fails, the system blocks, downgrades, routes review, or creates a data request.
The gate is explicit state, not a ranking hint. "Not eligible" must be representable and preserved so reviewers can see whether the system refused because the evidence was missing, stale, pending, restricted, rejected, or outside the permitted use.
Structure
See diagram-evidence-eligibility-state-flow for the canonical state flow: intake, admission, source snapshot, freshness, approval, sensitivity, eligibility, grounded reasoning, and refusal or review.
Consequences
The system stops laundering retrieved material into authority. Evidence can be present without being usable. Review queues become visible. Stale or restricted sources stop producing confident answers. Audit trails can explain why a conclusion proceeded, waited, or refused.
Tradeoffs
The gate adds schema, policy, and operational work. Corpus operators must preserve source metadata, review states, freshness rules, sensitivity labels, and permitted-use constraints. Some answers will be slower or blocked until evidence is approved or refreshed.
Failure Modes
- The retrieval service returns snippets without approval or freshness state.
- Eligibility is used only for ranking and not enforced before reasoning.
- Pending intake appears in prompts because it was indexed early.
- Sensitive evidence grounds an answer for an actor who cannot see the source.
- The system refuses without naming the missing eligibility fact or repair path.
Anti-Patterns
context.
gate.
refused.
- Corpus equals evidence - anything indexed can ground a conclusion.
- Citation theater - links are attached after generation even though the model used ineligible
- Policy in the prompt - eligibility rules are written as instructions instead of enforced by a
- Silent downgrade - bad evidence is ignored without preserving why the answer changed or
Known Uses
and evidence eligibility before security reasoning can cite or rely on retrieved material.
the delegated-action lifecycle.
- AISDR - tracks corpus approval, freshness, source health,
- EDDA - treats evidence as a governed domain in
Implementation Notes
Store eligibility inputs with the source snapshot, not only in the search index. Separate admission from retrieval: transport success says the material arrived; admission says it may enter this context. Evaluate freshness, approval, sensitivity, license, actor rights, and claim scope before the model sees the evidence. Preserve the gate result in the evidence bundle or decision record.
Review Checklist
- Which source states are eligible, pending, rejected, stale, restricted, or expired?
- Can retrieved material enter the prompt before the gate passes?
- Does the system preserve the exact source snapshot and gate result?
- What repair path exists when evidence is not eligible?
- Can a reviewer explain why a cited source was eligible for this actor and claim?
Full source pattern: source.md.