SKIP TO CONTENT

CASE STUDY

INTERNAL TOOLING

sage-kernel + Course Auditor

Internal SDLC platform + content-audit harness behind my own products — a proof-first, 140-tool MCP server with claim/source integrity checks.

Topology: a 140-tool MCP surface backed by AST engines writes into an HMAC-anchored append-only proof ledger and proof graph, which feed the 68-of-68 release:check gates. In parallel a Python course auditor extracts 17k-plus claims into claim/source ledgers behind a scored gate (95 approved, 85 pilot, below 70 blocked) that reports into the same release check.

PROBLEM

"Nothing stated, everything proven" fails the moment the tooling cannot back its own claims with re-runnable evidence.

BUILT

A 140-tool MCP SDLC server with an HMAC-anchored proof ledger and AST engines, plus a Python course auditor extracting claim/source ledgers behind scored gates.

MCP tools served over stdio
140
release gates green
68/68
claims extracted at baseline
17k+
lessons audited across 23 courses
460+
FAILURE MODES
  • Quietly edited proof ledgers
  • Corrupt ledger locks
  • Fail-open DAGs
  • Claim over-extraction from prose

INTERACTIVE — TRY THE GROUNDING CHECK · PICK A QUESTION

SIMULATION OF MY REAL SCRIPT
RETRIEVED + CITED SOURCES
policies/refunds.mdCITEDlegal/terms.mdRETRIEVED
GROUNDED — SHIPS AUTOMATICALLY

Every citation is inside the retrieved set. The grounding check passes and the answer goes out — no human needed.

EVIDENCE — REAL ARTIFACTS, CAPTURED FROM DISK

Node.jsMCP (stdio)PythonpytestHMAC-anchored ledgerAST analysisINSPECT THE REPO ↗

IMPROVE NEXT: Stand up a public, sanitized demo of the proof ledger and terminal cockpit so the 68/68 gate output is verifiable outside my machine.