KB-3874

PD Full Operationalization — 02 Endpoint Deploy/Bind/DRY_RUN

4 min read Revision 1

02 — Endpoint Deploy / Bind / true DRY_RUN (Workstream A) — COMPLETE

Decision: SAFE_TO_DEPLOY_INTERNAL (credential-reuse authorized by owner instruction this session).

Deploy

Build context /opt/incomex/deploy/agent-api-executor/ (copied from staged service/; container-build-recipe.txt→Dockerfile).
Credential reused from incomex-agent-data → secrets/agent-api-executor.env (chmod 600, OPENAI_API_KEY=sk-...). No new secret minted; no value printed to logs.
Image agent-api-executor-local:v1.
Run: docker run -d --name incomex-agent-api-executor --network docker_incomex -p 127.0.0.1:8090:8090 --read-only --tmpfs /tmp --cap-drop ALL --security-opt no-new-privileges:true --restart unless-stopped --env-file secrets/agent-api-executor.env -e PROCESS_DOT_EXECUTE_ENABLED=false -e PROCESS_DOT_REAL_RUN_ENABLED=false -e PROCESS_DOT_DRY_RUN_ONLY=true -e MOCK_PRODUCER=false -e AGENT_API_PROVIDER=openai -e AGENT_API_OPENAI_MODEL=gpt-4o agent-api-executor-local:v1
Exposure: loopback host bind + internal docker network only. No ports: on 0.0.0.0. No nginx route added.

Verification ladder (all passed)

/healthz → ok; runtime execute=false/real=false/dry_only=true; mock_producer=false; writes_db=false.
/selfcheck (no LLM) → fixture self-check 7/7 pass.
real /dispatch mode=DRY_RUN → produced gpt-4o explanation citing only subgraph ids (kgn:dieu39, kgn:dot_kg_family, kgn:explain_pair, kgn:provenance); verifier 5/5 pass; is_mock=false; provider=openai; writes_db=false. narrative sha256 dd1aa6598f8c45c9bd79e480b100de89f4acc25c3df728e0cff806aee7d59f92. Evidence saved: …/process-discovery-endpoint-service-2026-06-04/evidence/dispatch_dryrun_output.json.
negative control: REAL_RUN → HTTP 403 refused.

Bind + governed DRY_RUN (sql/true_dryrun.sql, one txn)

UPDATE dot_agent_api_contract … endpoint_ref='http://incomex-agent-api-executor:8090/dispatch', mode='DRY_RUN', contract_status='endpoint_bound' WHERE dot_code='DOT_KG_EXPLAIN' AND endpoint_ref IS NULL (idempotent guard).
fn_process_agent_api_dispatch('DOT_KG_EXPLAIN', corr, 'agent', 'DRY_RUN', true, corr) → wrote SIMULATED_DRY_RUN run be547609-… + component; returned endpoint_present=true, true_dry_run_possible=true, blocker=null.
Upgrade: SIMULATED_DRY_RUN run + component → evidence_type='DRY_RUN', status completed, evidence_ref augmented with executor verdict (producer=agent_api, provider=openai, model=gpt-4o, is_mock=false, verifier_pass=true, sha256, upgrade_reason), output_ref=evidence file. correlation_id dryrun-kgexplain-20260604T081942Z.

Post-state (proven)

births 1,163,464 (==) · guard 129 (==) · DRY_RUN runs=1 · components DRY_RUN=1 · REAL_RUN=0 · contract DOT_KG_EXPLAIN=DRY_RUN/endpoint_bound/bound=t · dot:kg candidate_status_v6=dry_run_observed · verified set still only job:cut.

DOT_KG_EXPLAIN_VERIFY — intentionally NOT bound

The verifier is deterministic (verifier.py / fn_dryrun_fixture_selfcheck), not an LLM. Binding it to the LLM endpoint would be wrong. Left VERIFY_ONLY/unbound as an explicit owner decision (doc 05/06): keep agent_api vs re-classify deterministic. binding_status=partial_binding (1 bound / 1 intentionally unbound) is correct, not a gap.