example.osint-proof

OSINT proof

Deterministic conformance and scale qualification for the reusable company, job, and people system. The proof materializes 100 companies, 1,000 jobs, and 1,000 people with unique evidence locators; the scale pass processes 100,000 entity identities, 1,000,000 assertions, and 1,000,000 embedding identities. It also measures local detail and filtered-search work against the reference p95 budgets.

bun run proof prints a synthetic JSON receipt. It is a deterministic development baseline only and never satisfies the production database gate.

bun run proof:france-corpus -- <receipt.json> validates a live corpus receipt starting from france-corpus.pending.json. Qualification requires the exact release SHA, France/EU scope, live provider and source receipts, at least 100 companies, 1,000 active-or-recent jobs, and 1,000 policy-approved people. Collector, identity resolver, and reviewer must be independent. Representative review must find zero false or name-only silent merges, zero unsupported fields, zero jobs closed without explicit evidence, and authorized evidence for every reviewed displayed fact. Pending, stale, malformed, or policy-pending receipts exit non-zero and never satisfy the production corpus gate.

bun run proof:postgres is the opt-in managed qualification harness. It requires:

• OSINT_QUALIFICATION_DATABASE_URL: migration-owner URL for replay-safe load and concurrent HNSW index creation;
• OSINT_QUALIFICATION_RUNTIME_DATABASE_URL: distinct non-owner runtime URL for FORCE-RLS benchmark queries;
• OSINT_QUALIFICATION_CONFIRM_LOAD=YES;
• Optional OSINT_QUALIFICATION_DIMENSIONS (default 1536) and OSINT_QUALIFICATION_BATCH_SIZE (default 10000).

The loader writes a dedicated qualification scope, 100,000 projections, 1,000,000 evidence-linked assertions, and 1,000,000 evidence-linked vectors. It is idempotent, cancellable between batches, creates a generation-specific HNSW index, verifies exact cardinalities, then captures detail/hybrid p95 latency and EXPLAIN (ANALYZE, BUFFERS, FORMAT JSON) plans through the runtime role. A missing URL, explicit load confirmation, pgvector/migration, target count, or latency budget fails closed. Credentialed providers, object storage, and the France corpus remain separate live gates.

bun run proof:postgres:docker runs the same 100k/1m/1m loader and benchmark against a pinned disposable pgvector container, using a unique host temporary data directory and distinct runtime login. Its default 64-dimensional vector profile is a real PostgreSQL scale/concurrency proof designed for bounded local resources; it does not replace the managed 1536-dimensional production gate. OSINT_SCALE_DOCKER_DIMENSIONS and OSINT_SCALE_DOCKER_BATCH_SIZE may increase the local profile deliberately. The runner never prunes or reuses Docker state.

OSINT_RECOVERY_RELEASE_SHA=$(git rev-parse HEAD) bun run proof:recovery:docker performs a separate bounded local restore proof. It applies every migration, seeds canonical evidence/projection/vector/audit/outbox state whose snapshot references a real content-addressed object in a separately pinned disposable MinIO service, dumps the source, and restores into a fresh sibling database. It compares exact row and row-content checksums, downloads and rehashes the unchanged evidence object after restore, verifies that both source and restored snapshot references match its manifest, verifies pgvector and HNSW restoration, then proves correct-scope visibility and wrong-scope denial through osint_runtime. It requires a clean worktree and emits a local_disposable_sibling receipt; it does not satisfy managed fresh-environment RPO/RTO, evidence-bucket restore, or application smoke gates.