Skip to content

Golden Path

BRYAN DAVID WHITE edited this page Feb 23, 2026 · 5 revisions

Golden Path

The Golden Path is a 7-step end-to-end proof loop that exercises the full DeepSigma governance pipeline in a single command. It demonstrates: ingest → canonicalize → extract → score → detect drift → patch → recall.

Source: src/demos/golden_path/


Quick Start

# Fixture mode (no credentials needed)
deepsigma golden-path sharepoint \
  --fixture src/demos/golden_path/fixtures/sharepoint_small --clean

# Live mode (requires env vars)
deepsigma golden-path sharepoint --list-id Documents

The 7 Steps

Step Name What It Does
1 Ingest Fetch records from source connector (or fixture)
2 Canonicalize Transform raw records to canonical envelopes
3 Extract Extract truth/reasoning/memory claims
4 Score (Baseline) Compute initial coherence score (0–100, A–D)
5 Detect Drift Run drift detection against extracted claims
6 Patch Apply corrective patches for detected drift
7 Recall (IRIS) Query IRIS to verify knowledge retrieval

Supported Sources

Source Flag Requirements
SharePoint --source sharepoint SP_TENANT_ID, SP_CLIENT_ID, SP_CLIENT_SECRET, SP_SITE_ID
Snowflake --source snowflake SNOWFLAKE_ACCOUNT, auth credentials
Dataverse --source dataverse Power Platform credentials
AskSage --source asksage ASKSAGE_API_KEY, ASKSAGE_EMAIL

All sources support --fixture <path> for offline/test mode.


Output

============================================================
  GOLDEN PATH — Complete
============================================================

  [1] INGEST               OK
  [2] CANONICALIZE          OK
  [3] EXTRACT               OK
  [4] BASELINE_SCORE        OK
  [5] DETECT_DRIFT          OK
  [6] PATCH                 OK
  [7] IRIS_RECALL           OK

  Records ingested:    5
  Claims extracted:    12
  Baseline score:      72.0 (C)
  Patched score:       88.0 (A)
  Elapsed:             340ms
============================================================

Use --json for machine-readable output.


Related Pages

Clone this wiki locally