Harness reference

---

title: Harness Reference description: Complete CLI reference for adjust360_harness.py: every flag, mode, IDP track, exit code, and worked invocation examples. kind: authored status: current source_date: 2026-10-02 tags: [adjust360_harness.py, CLI, modes, IDP, carrier_structured, extract, subro, full, receipt, stability]

---

What it is

adjust360_harness.py is the single command-line entry point for the Adjust360 FNOL pipeline. It accepts a claim document or a pre-built manifest, routes through the chosen IDP track, executes the selected pipeline mode, and writes structured artefacts to disk or MongoDB. The API that serves adjust360.enowclear360.com sits behind the harness: no tool touches a database directly, and no pipeline stage is accessible except through this entry point.

IDP Tracks

The --idp flag selects how the claim document is ingested before any GKR tool sees it.

Track	Value	How it works	Document input
Native	native	FNOLExtractorGKR performs LLM-based extraction from raw files	--doc or --bundle
Reducto	reducto	Pre-processed Reducto output loaded from manifest	--manifest
Databricks	databricks	Pre-processed Databricks output loaded from manifest	--manifest
Carrier Structured	carrier_structured	Validated AXA v1.0 payload loaded from manifest	--in (alias for --manifest)

The reducto, databricks, and carrier_structured tracks skip FNOLExtractorGKR entirely; idp_normalizer.py converts the manifest to the canonical envelope shape that all downstream GKR tools expect.

Pipeline Modes

The --mode flag controls how far the pipeline runs and what it produces.

Mode	Steps executed	Primary output
extract	FNOLExtractorGKR only	Classification + core_facts printed; stops before any GKR analysis tool
subro	Steps 1–6, through SubrogationScreenerGKR	Thesis record, trace, optional brief
full	All 7 tools including AuthorityGateGKR	Complete run log, thesis, brief
stability	Full pipeline × N runs with cached extraction	Per-field agreement rate report
receipt	Extraction + JurisdictionResolverGKR only	HTML data receipt; no determination

extract is the lowest-cost diagnostic pass. It confirms the extractor reached a clean classification and that core_facts are populated before committing to a full pipeline run. It accepts --idp native only; carrier-structured and manifest-based tracks bypass extraction entirely.

subro runs the carrier path through the subrogation determination and stops there. Liability and coverage tools are recorded as not run. Use this for AXA POC scenarios and for any run where only the subrogation thesis matters. The --in flag is the alias for --manifest on the carrier_structured track; --store DIR writes the eight run artefacts (payload.json, receipt.json, receipt.html, thesis.json, trace.json, brief.html, brief.pdf, manifest.json) to DIR/lob/claim_id/run_id/. See the harness-and-modes page for the receipt and thesis relationship; see this page's Flag Reference for the complete argument list.

full is the standard production invocation. It adds AuthorityGateGKR as the seventh step after the subro screener, producing the complete authority tier assignment alongside the subrogation determination.

stability catches non-determinism in rules tools. It runs the full pipeline N times using a single cached extraction result, then reports the agreement rate for every field in the thesis record. An agreement rate below 1.0 on a rules tool is a defect: rules tools carry the llm_construction_count == 0 assertion.

receipt tells the sender what its payload contains before any determination is made. It validates the manifest through the normalizer, builds the receipt record, and renders receipt.html with data-trace on every element. It runs no analysis tool and makes no subrogation call. Useful for carrier onboarding: the sender can confirm all decisive fields are present and that jurisdiction readiness is green before submitting a production claim.

Flag Reference

Flag	Type	Default	Notes
--idp	choice	native	native | reducto | databricks | carrier_structured
--doc PATH	appendable	:	Single file, directory, or repeated for ordered list. Native track only. Unsupported extensions skipped with warning.
--bundle ZIP	path	:	ZIP archive extracted to temp dir and processed as directory. Native track only.
--manifest FILE	path	:	Canonical manifest JSON. Required for reducto, databricks, carrier_structured.
--in FILE	path	:	Alias for --manifest when --idp carrier_structured. Validated against v1.0 schema + SHA-256.
--mode	choice	full	extract | subro | full | stability | receipt
--lob	choice	auto-detect	wc | auto | gl | property
--state	string	derived from doc	Two-letter US state code
--reference-date	YYYY-MM-DD	:	Reference date for SOL/deadline calculations. Required for --idp carrier_structured.
--case-id	string	TEST-{timestamp}	Explicit case identifier
--model	string	openai/gpt-oss-120b	LLM model for FNOLExtractorGKR (native track only)
--out PATH	path	:	Write full run log JSON to this path
--brief PATH	path	:	Write adjuster-facing Subrogation Intelligence Brief. Requires --mode subro or full. PDF via Playwright if installed, HTML fallback.
--store DIR	path	:	Write run artefacts to DIR/lob/claim_id/run_id/ (eight files: payload, receipt, thesis, trace, brief, manifest)
--n-runs	int	5	Number of pipeline runs for --mode stability only
--phase-a-seed FILE	path	:	JSON with pre-captured Phase A output (semantic_facts, extraction_metadata). Parity harness only: not for production runs.
--skip-date-gate	flag	off	Bypass null date_of_loss intake gate. Uses sentinel UNKNOWN-DOL. Local fixture testing only; never in Conductor.
--compare-to CASE_ID	string	:	Compare this run against stored case in fnol_claim_intel. Produces IMPROVED / REGRESSED / SAME report via LLM.
--write-mode	choice	ephemeral	Only valid value: ephemeral. Instructs FNOLExtractorGKR and CoverageAnalyzerGKR not to write intermediate results to MongoDB. Pass explicitly for test clarity.
--query-workflow	string	:	ClickHouse: show all tasks for a Conductor workflow ID. Mutually exclusive with pipeline run.
--query-trace	string	:	ClickHouse: drill into a specific task trace ID. Mutually exclusive with pipeline run.
--service	choice	:	Filter --query-trace by tool-sandbox or sandbox-service

Worked Examples

Native extraction: AUTO claim, California, subro mode with brief:

python adjust360_harness.py \
  --idp native \
  --doc /data/claims/morales_police_report.pdf \
  --lob auto \
  --state CA \
  --mode subro \
  --case-id AXA-POC-001 \
  --brief /tmp/AXA-POC-001_brief.pdf

Native extraction: verify extraction only before committing to full run:

python adjust360_harness.py \
  --idp native \
  --doc /data/claims/martinez_fnol_packet.pdf \
  --lob auto \
  --state TX \
  --mode extract \
  --case-id EXTRACT-CHECK-001

Full pipeline: ordered multi-document bundle:

python adjust360_harness.py \
  --idp native \
  --doc /data/claims/fleet_001_fnol.pdf \
  --doc /data/claims/fleet_001_police.pdf \
  --doc /data/claims/fleet_001_photos/ \
  --lob auto \
  --state FL \
  --mode full \
  --case-id FLEET-2025-001 \
  --store /data/runs/

ZIP bundle: property claim:

python adjust360_harness.py \
  --idp native \
  --bundle /data/bundles/property_claim_TX_2025.zip \
  --lob property \
  --state TX \
  --mode subro \
  --out /tmp/property_run_log.json

Carrier-structured v1.0: AXA AUTO payload, subro mode:

python adjust360_harness.py \
  --idp carrier_structured \
  --in /data/axa/AXA-SAMPLE-AUTO-001.json \
  --lob auto \
  --state TX \
  --reference-date 2025-07-22 \
  --mode subro \
  --brief /tmp/AXA-SAMPLE-AUTO-001_brief.pdf \
  --store /data/axa/runs/

Receipt mode: preflight validation before production submission:

python adjust360_harness.py \
  --idp carrier_structured \
  --in /data/axa/new_claim_payload.json \
  --lob auto \
  --state NY \
  --reference-date 2025-09-15 \
  --mode receipt \
  --out /tmp/preflight_receipt.json

Stability mode: detect non-determinism across five runs:

python adjust360_harness.py \
  --idp native \
  --doc /data/claims/morales_police_report.pdf \
  --lob auto \
  --state CA \
  --mode stability \
  --n-runs 5 \
  --case-id STABILITY-CHECK-001

Ephemeral run: local fixture testing without MongoDB writes:

python adjust360_harness.py \
  --idp carrier_structured \
  --in /data/fixtures/blind_auto.json \
  --lob auto \
  --state CA \
  --reference-date 2025-06-01 \
  --mode subro \
  --write-mode ephemeral \
  --skip-date-gate

Regression comparison: compare new run against stored baseline:

python adjust360_harness.py \
  --idp native \
  --doc /data/claims/morales_police_report.pdf \
  --lob auto \
  --state CA \
  --mode subro \
  --case-id AXA-POC-001-V2 \
  --compare-to AXA-POC-001

ClickHouse Observability

The --query-workflow and --query-trace flags are mutually exclusive with a pipeline run: they connect to ClickHouse and return trace data without executing any claim logic.

# Show all tasks that ran in a Conductor workflow
python adjust360_harness.py --query-workflow <CONDUCTOR_WORKFLOW_ID>

# Drill into a specific task trace, optionally filtered by service
python adjust360_harness.py --query-trace <TRACE_ID> --service tool-sandbox

The default ClickHouse host is http://184.73.82.115:8123. The CH_PASSWORD environment variable is required for both query flags. See observability-query-guide.md at the vin-tools repo root for the full SQL query patterns and table schema.

Environment Variables

Variable	Required for	Notes
ARTIFI_MONGO_CONNECTION_STRING	All tool runs that read GKR or write claims	Required for every mode except ephemeral dry-run fixture tests
LITELLM_API_KEY	--idp native	FNOLExtractorGKR LLM calls route through the LiteLLM proxy
LITELLM_BASE_URL	--idp native	Dev: https://api.groq.com/openai/v1. Prod: http://44.220.92.58:4000/v1
CH_PASSWORD	--query-workflow, --query-trace	ClickHouse authentication. Default host: http://184.73.82.115:8123

Exit Codes

Code	Meaning
0	Success
2	V1IntegrityError: carrier_structured v1.0 schema validation or SHA-256 failure. See the carrier-structured-v1-adapter page for remediation.
non-zero (other)	Pipeline error: check stderr and the run log written to --out