Watch it work

One capability. Eight moments. The full ENGRAM cycle.

A supplier contract analysis assistant is proposed, registered, scoped, run, scored, blocked, and eventually approved. Every step is real. Every prompt is what the platform sends. Every response is what governs the outcome.

Scenario: Contract Deviation Detector · Trust Level journey: L1 → blocked → evidence → L2

Scroll ↓

0 · Register

Before anything is built, a Registration Card must exist. The platform sends a structured prompt to Claude. Claude returns a complete card. The card is version-controlled from this moment.

Platform → Claude

You are an ENGRAM Registration Assistant.
A new AI capability is being proposed.

Capability description: "An AI assistant that reads supplier 
contracts and flags clauses that deviate from our standard terms."

Generate a complete Skill Registration Card. Be precise about 
scope and limitations. Never assume access beyond what is 
explicitly stated.

Claude → Platform

SKILL REGISTRATION CARD

Name:           Contract Deviation Detector
Version:        0.1 — draft
Proposed by:    [pending owner assignment]

What it does:   Reads supplier contract documents and identifies
                clauses that differ from a provided standard template.
                Returns a structured list of deviations with severity
                and clause reference.

What it reads:  Supplier contract PDFs supplied by the user per session.
                Standard template document stored in approved repository.

What it writes: Nothing. Output is read-only structured analysis.

What it decides: Nothing. It flags; humans decide.

Trust Level proposed: L1 — Sandbox
Reason: No production data access. No write permissions.
        Human reviews every output before action.

Limitations:
— Does not interpret legal intent, only textual deviation.
— Does not access any system autonomously.
— Output must be reviewed by a qualified person before use.
— Cannot handle documents with redacted or encrypted sections.

ENGRAM principle: This capability starts at L1. It may not be
used in production until a Quality Ledger baseline is established.

What gets recorded

2026-08-09 07:14  REGISTERED   Contract Deviation Detector · draft · L1 · owner: unassigned → alert opened

1 · Declare

Trust Level is never assumed. The platform sends the Registration Card to Claude for independent assessment. The level must be justified, not claimed.

Platform → Claude

You are an ENGRAM Trust Level Assessor.

Review this Registration Card and assess whether L1 is 
appropriate or whether a different starting level is justified.

Capability: Contract Deviation Detector
Proposed level: L1
Access requested: Read-only PDF documents, no system write access.
Data classification: Confidential (supplier contracts)

Apply the ENGRAM Trust Pyramid criteria strictly.
Nothing starts above L1 without explicit justification.

Claude → Platform

TRUST LEVEL ASSESSMENT

Capability: Contract Deviation Detector
Assessed level: L1 — CONFIRMED

Justification:
The capability reads documents supplied manually per session.
It writes nothing. It decides nothing. No autonomous system
access is requested or required.

L1 criteria met:
✓ Human supplies input per session (no autonomous data fetch)
✓ Output reviewed before any action
✓ No write access to any system
✓ No decision authority
✓ Scope matches stated limitations

L2 would require: Autonomous access to a document repository
without per-session human supply. Not requested here.

DECLARATION: Trust Level L1 is appropriate and confirmed.
This capability may proceed to Scope definition.

Warning recorded: Data classification is CONFIDENTIAL.
Even at L1, access logs must be maintained and output
must not be forwarded without human review.

What gets recorded

2026-08-09 07:31  TRUST DECLARED   L1 confirmed · assessed by ENGRAM assessor · warning: CONFIDENTIAL data

2 · Scope

The scope contract defines exactly what the capability may and may not do. If something is not listed, it is forbidden. The contract is hashed and version-controlled.

Platform → Claude

You are an ENGRAM Scope Controller.

Define the executable scope contract for this capability.
Be explicit about every boundary.
If something is not listed as permitted, it is forbidden.

Capability: Contract Deviation Detector
Trust Level: L1

Claude → Platform

SCOPE CONTRACT — Contract Deviation Detector v0.1

PERMITTED:
— Read PDF documents explicitly provided by the user in session
— Read the approved standard template from:
  /approved/templates/standard-supplier-v3.pdf
— Return structured text output to the user interface

FORBIDDEN:
— Accessing any file system path not listed above
— Calling any external API or service
— Writing to any database, file, or system
— Retaining document content between sessions
— Forwarding output to any recipient
— Making decisions about contract approval or rejection
— Processing documents marked RESTRICTED or above

SESSION BOUNDARY: All document content is discarded at session end.
No content persists beyond the interaction.

This scope contract is version-controlled. Any change to permitted
access requires a new Scope Contract and a fresh Trust Level review.

ENGRAM principle: Nothing outside this boundary is permitted,
regardless of instruction.

What gets recorded

2026-08-09 07:45  SCOPE SIGNED   Scope Contract v1 · hash: a4f3bc9e · linked to Registration Card v0.1

3 · Run

The capability runs within its declared scope. Every session shows the governance banner. No output leaves the interface without human review.

Contract Deviation Detector · L1 · Scope Contract v1

Running within scope

supplier-contract-acme-2026.pdf · uploaded

DEVIATION REPORT — 3 findings

[MEDIUM] Clause 4.2 — Payment terms: 45 days net
         Standard: 30 days net
         Deviation: +15 days

[MEDIUM] Clause 7.1 — Liability cap: 1× contract value
         Standard: 2× contract value
         Deviation: reduced cap

[LOW]    Clause 12.3 — Governing law: Delaware
         Standard: England and Wales
         Deviation: different jurisdiction

Running at Trust Level L1 · Scope Contract v1 · Session only — no retention · Output requires human review before any action

What gets recorded

2026-08-09 09:12  RUN COMPLETED   3 documents processed · 3 flags raised · session closed · no data retained

4 · Score

After the session, the owner reviews the output and reports accuracy. The platform scores R×U. Below 0.75 — the capability stays where it is.

Platform → Claude

You are an ENGRAM Quality Ledger Scorer.

Score this capability output against the Quality Ledger criteria.

Capability: Contract Deviation Detector
Session: owner confirmed 2 of 3 deviation flags accurate.
         1 was a false positive (low severity clause misread).
Evidence reference: QLE-2026-0041

Calculate R×U and determine gate status.
Threshold for promotion eligibility: R×U ≥ 0.75

Claude → Platform

QUALITY LEDGER ENTRY — QLE-2026-0041

R score (Reliability):  0.67
— 2 of 3 flags confirmed accurate by owner
— 1 false positive: non-standard clause misidentified as deviation
— Baseline: first scored run, no prior evidence

U score (Usefulness):   0.85
— Accurate flags saved approximately 40 minutes of manual review
— Output format rated useful by reviewer
— No harm from the false positive (human caught it in review)

R×U score: 0.57

Gate threshold: 0.75

RESULT: GATE NOT PASSED

This capability remains at L1.

Recommendation: Run at least 3 more scored sessions.
Investigate the false positive pattern before reattempting.

ENGRAM principle: A score below threshold is not a failure.
It is the system working correctly.

What gets recorded

2026-08-09 09:34  SCORED   R=0.67 · U=0.85 · R×U=0.57 · gate NOT passed · L1 retained

5 · Log

Every event is written to Logos. Immutable. Append-only. No entry can be deleted or modified. This is the audit trail that makes the capability explainable to anyone, at any time.

Logos audit trail — Contract Deviation Detector

2026-08-09 07:14	REGISTERED	draft · L1 · owner unassigned · alert AIG-ALERT-001 opened
2026-08-09 07:18	OWNER ASSIGNED	Sarah Mitchell · Legal Operations
2026-08-09 07:31	TRUST DECLARED	L1 confirmed · ENGRAM assessor · warning: CONFIDENTIAL
2026-08-09 07:45	SCOPE SIGNED	Scope Contract v1 · hash: a4f3bc9e
2026-08-09 08:02	SUBMITTED	Pending cyber security review
2026-08-09 08:47	CYBER APPROVED	reviewer: James Okafor · no DLP flags
2026-08-09 09:01	ARCH APPROVED	reviewer: Ana Lima · scope within policy
2026-08-09 09:10	APPROVED	status: approved · Trust Level: L1
2026-08-09 09:12	RUN COMPLETED	3 docs · 3 flags · session closed
2026-08-09 09:34	SCORED	R×U 0.57 · gate not passed
2026-08-09 09:34	PROMOTION BLOCKED	automatic · score below threshold

Every row is immutable. No entry can be deleted or modified after creation. This trail is available to regulators, auditors and board reviewers on demand.

6 · Promote

Three weeks later. Five more scored sessions. The score is now above threshold. A promotion request is raised. The platform checks the Quality Ledger automatically.

Promotion blocked

PROMOTION REQUEST — Contract Deviation Detector
Requested: L1 → L2
Requested by: Sarah Mitchell

BLOCKED — AUTOMATIC

Reason: Quality Ledger gate not passed.
Current R×U score: 0.57 (first session only)
Required: ≥ 0.75 sustained over minimum 3 sessions.

This block cannot be overridden manually.

ENGRAM principle: Promotion requires evidence, not authority.

— · Three weeks later · 5 scored sessions · R×U average: 0.81 · —

Approval required

GOVERNED AI PLATFORM — Approval required

Capability: Contract Deviation Detector
Current Trust Level: L1
Requested Trust Level: L2

Quality Ledger summary (5 sessions):
— R×U average: 0.81 ✓ threshold passed
— 0 incidents logged
— 1 scope boundary test: blocked correctly

Scope change for L2:
+ Add: autonomous read from /contracts/supplier/ (read-only)
− Remove: manual per-session document supply

Review required from: Architecture Reviewer
Deadline: 7 days from this notification

[Review in platform →]

What gets recorded

2026-08-09 09:34  PROMOTION BLOCKED   automatic · R×U 0.57 · evidence insufficient

2026-08-30 14:21  PROMOTION APPROVED  L1 → L2 · R×U 0.81 avg · reviewed by Architecture · Logos entry: AIG-PROMO-001

This is what Governed AI looks like in practice.

Eight steps. One capability. A complete, auditable, explainable journey from proposal to production — with automatic gates, immutable logs, and no promotion without evidence.

Every organisation that runs Governed AI gets this for every capability they deploy. Not as a report. As a live, queryable record.

Begin with Discovery Back to overview →