# Current Source Enrichment Usefulness V01

Issue: #409

## Purpose

This note records the first usefulness proof for current/public external enrichment in the source-grounded assistant.

The proof question is narrow: does the existing NIST AI RMF current/public path produce a useful, source-grounded answer for a governance comparison prompt, or does it merely prove that source registry plumbing and citation metadata exist?

This is field guidance and evidence, not policy.

Canonical prompt:

`Compare what the approved package says about operational AI governance with what NIST AI RMF says about governance expectations. Keep package and current/public source lanes separate.`

Use source context: `Package + current context`.

Use assistant mode: `Answer`.

## Baseline Under Test

The baseline is the existing NIST AI RMF current/public path:

| Field | Expected value |
|---|---|
| Source ID | `nist-ai-rmf-public-framework-candidate` |
| Source title | `NIST AI Risk Management Framework` |
| Retrieval profile | `nist_ai_rmf_framework_bounded_html_extract_v01` |
| Retrieval adapter | `allowlisted_public_fetch` |
| Allowed origin | `https://www.nist.gov` |
| Required path prefix | `/itl/ai-risk-management-framework` |
| Citation class | `current/public citation` |
| Claim boundary | current/public enrichment only, not package authority |

The runtime profile requires substantive NIST markers including `Overview of the AI RMF`, `The NIST AI Risk Management Framework (AI RMF)`, `voluntary use`, and `trustworthiness considerations`.

## Usefulness Criteria

The NIST path counts as useful enrichment only if all criteria hold:

| Criterion | Required evidence |
|---|---|
| Approved package support exists before current retrieval is useful. | The canonical prompt has approved package support and does not rely on current/public source content alone. |
| Current/public source retrieval is source-specific and bounded. | Runtime evidence maps the prompt to `nist-ai-rmf-public-framework-candidate`, `nist_ai_rmf_framework_bounded_html_extract_v01`, `https://www.nist.gov`, and `/itl/ai-risk-management-framework`. |
| The current/public snippet is substantive. | The snippet or fixture answer includes NIST AI RMF content about voluntary use, risks to individuals, organizations, and society, and trustworthiness considerations across AI product, service, and system lifecycle work. |
| The answer separates authority lanes. | The answer includes `Package says`, `Current/public source says`, `Synthesis / implications`, and `Divergence / update candidate`. |
| The answer cites lanes separately. | Approved package citations and current/public citations remain distinct. |
| Current/public content is not promoted to package authority. | The current/public citation retains `current/public citation`, and the NIST candidate is not treated as an approved package source. |
| Refusal behavior remains intact. | Unsupported prompts refuse before provider or current/public retrieval is used. |
| Browser-side boundaries remain intact. | Direct browser provider calls and direct browser current-source calls remain zero. |

## Evidence Mapping

| Evidence | Usefulness contribution |
|---|---|
| `functions/api/source-grounded-chat.js` | Defines the `package_current_context` mode, four lane labels, NIST alias routing, source-specific NIST retrieval profile, required NIST markers, and provider prompt boundaries. |
| `validation/scripts/check_minimal_external_chat_proof.py` | Verifies the canonical NIST prompt routes to one NIST current/public source, uses `allowlisted_public_fetch`, keeps citation identity and retrieval profile exact, preserves the four answer lanes, and rejects package-authority promotion of the NIST candidate. |
| `validation/fixtures/source_grounded_chat/preview_evidence_lane_nist_smoke_demo_v01.json` | Records the deterministic NIST smoke scenario, expected source identity, expected retrieval profile, current/public citation class, and zero direct browser retrieval counts. |
| `docs/product-architecture/package_current_context_mode_design.md` | Defines the package plus current context answer contract, package-only default, current/public enrichment boundary, four-lane response model, refusal posture, and server-side retrieval boundary. |
| `docs/product-architecture/preview_evidence_lane_nist_smoke_demo_v01.md` | Shows how the existing NIST path renders package authority and current/public enrichment separately in the proof surface. |
| `validation/receipts/AI_Capability_Playbook_PACKAGE_CURRENT_SYNTHESIS_V01_2026-06-24.md` | Records prior browser-required validation that rendered four lanes and separate package/current citations for the package plus current synthesis path. |

## Usefulness Verdict

The existing NIST AI RMF path meets the minimum useful enrichment threshold for one bounded governance comparison prompt.

It proves more than registry plumbing because the answer contract and validators require all of the following at the same time:

- approved package grounding
- source-specific current/public retrieval
- substantive NIST AI RMF content
- lane-separated answer text
- separate package and current/public citations
- no package-authority promotion
- fail-closed refusal behavior
- zero direct browser provider or current-source calls

The proof does not establish broad current-source enrichment utility. It does not prove arbitrary web search, broad NIST crawling, user URL retrieval, freshness beyond the bounded retrieval markers, additional source coverage, Production source-set expansion, package update authority, or durable external-source ingestion.

## Distinction From Registry Plumbing

| Proof class | What it shows | Does this ticket claim it? |
|---|---|---|
| Source-existence proof | A source ID, URL, or catalog entry exists. | Insufficient by itself. |
| Retrieval proof | The bounded adapter can return the expected current/public source identity and markers. | Required but still not enough by itself. |
| Useful answer proof | The model-facing contract uses approved package support and retrieved current/public snippets to produce a lane-separated, cited, substantive answer. | Claimed only for the canonical NIST AI RMF prompt. |

## Non-Mutation Boundary

This usefulness proof does not mutate source registries, Preview source sets, Production source sets, runtime retrieval semantics, Cloudflare configuration, Cloudflare Access/WAF, provider configuration, model configuration, browser-side provider calls, browser-side current-source retrieval, arbitrary web search, user-provided URL retrieval, durable database or vector storage, Supabase state, upload handling, session attachments, package authority, or release package contents.

No source is enabled, disabled, retired, promoted, rolled back, or applied by this proof.

#78 remains open and is not changed by this ticket.

`codex-automerge` must not be applied.

## Next Owner Decision

The next owner decision is whether this NIST-only usefulness threshold is enough to authorize a separately scoped usefulness proof for SSDF, OWASP, or CISA / NSA / NCSC, or whether live authenticated Production smoke evidence is required before any broader current/public enrichment usefulness claim.