Integrator verification

The canonical integrator guide — first-run CLI/SDK flow, previews, proofs, governance pointers — lives in docs/integrate.md.

This document holds the generated capability matrix (below) plus the authoritative LangGraph checkpoint trust subsection. For certificate field meanings and runKind semantics site-wide, see docs/outcome-certificate-normative.md and docs/outcome-certificate-integrator.md.

LangGraph checkpoint trust

Checkpoint trust verifies persisted orchestration checkpoints alongside registry-backed SQL so a LangGraph-shaped run cannot claim alignment from trace narrative alone.

Objective

Produce the same deterministic Outcome Certificate v3 contract as agentskeptic check, but with runKind: contract_sql_langgraph_checkpoint_trust, checkpointVerdicts on eligible runs, and shared kernel wiring (verifyRunStateFromBufferedRunEvents, createLangGraphCheckpointTrustGate in src/langGraphCheckpointTrustGate.ts).

Eligibility

For the selected workflowId, every buffered tool_observed event considered for checkpoint trust must:

• Validate as a tool_observed NDJSON row.
• Carry schemaVersion: 3 (mixed or schemaVersion-missing tool_observed rows make the workflow ineligible with reason such as LANGGRAPH_CHECKPOINT_TRUST_NON_V3_TOOL_OBSERVED).
• Include structured langgraphCheckpoint payloads consistent with emitter contracts (ordering and strict emitter rules for Node reference tooling: scripts/lib/langgraphReferenceVerifyCore.mjs, test/langgraph-reference-emitter-before-cli-spawn.test.mjs).

Ineligible runs terminate in a certificate-only outcome (no database verification passes); see docs/outcome-certificate-integrator.md (checkpoint trust bullet).

Generic agentskeptic check without --langgraph-checkpoint-trust against v3 tool_observed input for that workflow exits 3 with empty stdout (product guardrail).

Minimum CLI path

Authoritative shell snippets (SQLite and Postgres parity) live in docs/partner-quickstart-commands.md — use the --langgraph-checkpoint-trust variants of agentskeptic check with the examples/partner-quickstart/partner.events.ndjson + partner.tools.json pairing.

Regression driver (CI only, not consumer quickstart): scripts/langgraph-reference-verify.mjs (see docs/langgraph-reference-boundaries.md).

Minimum Python-first path

Run examples/python-verification/README.md (run_partner_kernel_demo.py, Postgres optional). Direct kernel helpers (e.g. verify_langgraph_checkpoint_trust) reuse the same policy as CLI.

User-visible outcome

Stream	Contents
stdout	One Outcome Certificate v3 JSON object with schemaVersion: 3; eligible runs include checkpointVerdicts after SQL reconciliation. runKind is contract_sql_langgraph_checkpoint_trust.
stderr	Leading **`truth_check_verdict: trusted
Exit codes	0 = matches_expectations; 1 = does_not_match; 2 = not_established; 3 = operational / usage errors. Details: docs/outcome-certificate-integrator.md.

Canonical highStakesReliance derivation for this runKind: docs/outcome-certificate-normative.md (derivation table).

Production irreversibility gate

assertLangGraphCheckpointProductionGate(certificate) (SDK / hosting) allows progression only when the certificate already reflects successful SQL alignment, checkpoint verdict rollup fully verified, trustDecision safe, highStakesReliance: permitted — the lgRowBPasses invariant in src/langGraphCheckpointTrustGate.ts. Intended for gating irreversible commits (deploy, ledger movement) after checkpoint trust—not a substitute for covering every branching path that needs a human risk review.

Preview vs enforced contract

Mode	Notes
agentskeptic quick	Infer SQL-only previews; runKind: quick_preview; highStakesReliance: prohibited; no checkpoint trust path (see docs/quick-verify-normative.md).
agentskeptic check + --langgraph-checkpoint-trust	Authoritative bounded multi-step truth with checkpoint rollup + SQL (this section).
agentskeptic enforce	Stateful baselines and drift (commercial lifecycle). See docs/ci-enforcement.md and docs/commercial-enforce-gate-normative.md.

Related tests & boundaries

Terminal-outcome regressions: test/langgraph-checkpoint-trust-terminal.contract.test.mjs. Oracle vs Python primacy narrative: docs/langgraph-reference-boundaries.md.

Generated Capability Matrix

Behavior	Capability	TS	Python
bigquery.sql_row.strong	unsupported	supported	supported
http_witness.strong	supported	supported	supported
mongo_document.strong	supported	supported	supported
mysql.sql_row.strong	unsupported	supported	supported
object_storage_object.strong	supported	supported	supported
postgres.sql_effects.eventual	supported	supported	supported
postgres.sql_effects.strong	supported	supported	supported
postgres.sql_relational.strong	supported	supported	supported
postgres.sql_row_absent.strong	supported	supported	supported
postgres.sql_row.bounded	supported	supported	supported
postgres.sql_row.eventual	supported	supported	supported
postgres.sql_row.strong	supported	supported	supported
sqlite.sql_row_absent.strong	supported	supported	supported
sqlite.sql_row.bounded	unsupported	supported	supported
sqlite.sql_row.eventual	supported	supported	supported
sqlite.sql_row.strong	supported	supported	supported
sqlserver.sql_row.strong	unsupported	supported	supported
vector_document.strong	supported	supported	supported