add evidence-grounded ai assistance

2026-09-03 12:16:12 +02:00
parent c93dbd1ab4
commit c0909132f2
12 changed files with 352 additions and 2 deletions
@@ -111,3 +111,13 @@ There is no send button, message composer, SMTP probe, validation email, campaig
 ## Remaining limitations

 The static client has no client-side crawler, scanner, contact extractor, enrichment scheduler, outreach integration, availability provider, or SSE delivery. It can display server-provided Phase 9 observations, but production still requires server-side official-site scoping, SSRF/DNS-rebinding/redirect controls, hard extraction/page/byte/time/candidate budgets, durable history/cache isolation and retention/deletion, abuse/rate controls, suppression regression tests, and authenticated provenance/audit coverage. For domain intelligence, display registrable-domain/PSL version and unresolved reasons, DNS status and freshness/TTL (not “available”), independent MX/NS/TXT uncertainty, and association confidence with explainable evidence. Never auto-attach candidate domains or treat `nxdomain`/`no_data` as availability. CSV preview is capped for display and is not an import workflow.
+
+## Phase 13 optional AI assistance UI contract
+
+The UI may offer AI drafting only when the authenticated API reports an approved, enabled capability for the current tenant and task. Provider choice, fallback, prompt construction, redaction, budgets, tenant authorization, and suppression checks are server-side; the browser must never receive provider secrets or call an AI vendor directly. Show provider/model/version and `unknown`/`unavailable`, timeout, partial, stale, or policy-blocked states distinctly from an empty or successful result.
+
+Every suggestion must display its evidence citations, tenant-scoped evidence IDs, exact evidence hash/snapshot identifier, observed time, uncertainty/conflict reasons, and policy/provider/model versions. A citation points to the evidence used; it is not proof that the source is correct, and an AI explanation is not an independently verified fact. Do not render unsupported or invented facts about names, roles, contact details, dates, outcomes, consent, deliverability, ownership, or other claims as facts. Preserve missing and conflicting evidence instead of filling gaps. Suppressed/do-not-contact records remain visible with the safety state and never become actionable because an AI suggestion is confident.
+
+AI output must be visibly labeled **AI suggestion — human review required** and remain read-only until an authorized human explicitly approves it. Approval must show the proposed change, citations/hash, freshness, tenant scope, and safe reason; rejection and expiry must be available. The UI must require re-review when the evidence hash or policy version changes and must display partial/failed approval rather than implying persistence. Approval does not authorize contact or verification.
+
+No Phase 13 control may send email/SMS, probe SMTP, create a campaign, schedule follow-up, alter pipeline/interactions/outcomes as if communication occurred, merge records, acquire a domain, or perform autonomous CRM/outreach actions. The browser must not hide or export suppressed data as eligible, and exports/reports must retain safe AI provenance and redaction labels where applicable. Production remains limited until browser/API tests cover citations and hash mismatch, redaction, fallback boundaries, approval/rejection, stale/conflicting evidence, suppression precedence, tenant non-disclosure, and no-autonomy controls; the current Compose stack has no configured AI provider.