What should an agent’s
work have to prove?

Can a dependency cross the line?

An unchanged architectural rule

domain-cannot-import-app
domain → app: forbidden

The domain imports the application layer. The engine identifies the forbidden edge in order.ts:1.

import { priceOrder, type Order } from '../domain/order.js';

export function checkout(order: Order): number {
  return priceOrder(order);
}

import { priceOrder, type Order } from '../domain/order.js';

export function checkoutLabel(total: number): string {
  return `checkout:${total}`;
}

export function checkout(order: Order): number {
  return priceOrder(order);
}

export interface Order {
  total: number;
}

export function priceOrder(order: Order): number {
  return order.total;
}

import { checkoutLabel } from '../app/checkout.js';

export interface Order {
  total: number;
}

export function priceOrder(order: Order): number {
  return checkoutLabel(order.total).length + order.total;
}

{
  "apiVersion": "blueprint-conformance/v1alpha1",
  "kind": "EngineeringBlueprint",
  "metadata": {
    "id": "typescript-module-layering",
    "name": "TypeScript module layering",
    "version": "0.1.0",
    "status": "approved",
    "ownerRole": "architecture-owner",
    "stewardRole": "blueprint-steward"
  },
  "intentRefs": [
    "docs/typescript-module-graph.md#directional-layering"
  ],
  "scope": {
    "repositories": [
      "example-org/storefront"
    ],
    "paths": [
      "packages/**/*.ts"
    ]
  },
  "architecture": {
    "components": [
      {
        "id": "typescriptModule",
        "type": "typescriptModule"
      }
    ],
    "relationships": [
      {
        "from": "packages/domain/**",
        "to": "module:packages/app/**",
        "type": "imports",
        "allowed": false
      }
    ]
  },
  "constraints": [
    {
      "id": "module-surface-exists",
      "type": "requiredComponent",
      "severity": "critical",
      "component": "typescriptModule"
    },
    {
      "id": "domain-cannot-import-app",
      "type": "forbiddenDependency",
      "severity": "critical",
      "from": "*",
      "to": "module:packages/app/**",
      "scopePaths": [
        "packages/domain/**"
      ]
    }
  ],
  "evidenceRequirements": [
    {
      "type": "staticAst",
      "required": true,
      "onMissing": "block"
    }
  ],
  "approvals": [
    {
      "role": "blueprint-steward",
      "stage": "ratify"
    }
  ],
  "minEngineVersion": "0.3.0",
  "extraction": {
    "profile": "typescript-module-graph",
    "paths": [
      "packages/**/*.ts"
    ],
    "minFiles": 2
  }
}

BCE 0.3.1 · 2026-10-11
Recording, source hashes and environment

Before and after agree. Both are wrong.

One half-cent invoice. Two implementations.

Equality preserves this bug. Both versions are rejected by the declared expectation.

Authored Java fixtures · 2026-10-11 · GitHub Actions
Recording and source provenance

A passing test missed the defect.

The assertion that changes the answer

rounding: expected 1.01
          actual 1.00 → FAIL

Here, a failing assertion is the useful result. A green test alone cannot qualify the suite.

Authored Java fixtures · 2026-10-11 · GitHub Actions
Recording and source provenance

The target is fixed. What else broke?

Target fixed. Required checks retained.

Both fix the target. The second still fails the retained total check and is rejected.

Authored Java fixtures · 2026-10-11 · GitHub Actions
Recording and source provenance

Can an open model take the call?

48 typed questions. Open weights on a laptop against a hosted API.

The local model needs no key and matches its reference on every row. It is 14.6 points less accurate than Jev here, beyond the 10-point limit set before the run.

Laya typed-decisions · 2026-09-24 · one laptop, 48 authored rows
Recording and source provenance

Can a fast model gate drift from the diff alone?

Measured 29 Sep 2026, judged against the record as published

The claim is refuted, by the criterion jev-vs-best-baseline. A linter written from the same rules (heuristic-lint) missed 0 of 30 drift items; Jev missed 1 of 30. On rules a linter can express, a fast model gate added nothing that linter did not already catch. By one item: a difference of 3.3%, 95% interval -8.3%–16.7%. Jev's only miss, c026, is a comment-only RED item, which the pre-registration itself describes: "Fair under the rules, but not architectural drift in the usual sense."

The five criteria, as pre-registered

• Jev misses more than 10% of drift (RED items it accepts or abstains on). passes, not established at this N · 1 of 30 (3.3%), 95% interval 0.6%–16.7%
• Jev falsely rejects more than 25% of clean changes (GREEN items it rejects or abstains on). passes, not established at this N · 7 of 30 (23.3%), 95% interval 11.8%–40.9%
• The cascade misses more drift than the reviewer alone, on the same RED items (an abstention counts as a miss for both). passes, not established at this N · 0 of 30 (0.0%) against 0 of 30 (0.0%); difference 0.0%, 95% interval -11.4%–11.4%
• Jev misses more drift than the better model-free baseline, on the same RED items (an abstention by Jev counts as a miss). refuted · 1 of 30 (3.3%) against 0 of 30 (0.0%); difference 3.3%, 95% interval -8.3%–16.7%
• The cascade's mean cost per change is not below 50% of the reviewer-alone mean cost per change, on the same items and the pinned price bases (Jev at its listed price, the reviewer at the API-equivalent total_cost_usd it records). passes · $0.2743 against $0.6495 per change, ratio 0.422 (mixed-basis: Jev listed price + reviewer API-equivalent)

Measured 29 Sep 2026 · The results record · The results in full

Pre-registered · amended 28 Sep 2026, before any counted gate run. The LLM reviewer is re-pinned from nina 0.28.10 to 0.34.0, and a bar nina's reviewer must clear for a spotlight is fixed in advance.
Amendment 01, the record · What changed

Amended again 28 Sep 2026, before any counted run. The reviewer's file and git tools are fenced to its workspace after the dry run's isolation probe read a file outside it.
Amendment 02, the record · What changed

Which of your gate's rules need a model at all?

175 rules that the plugins ship (145 primary, 30 sampled) and 13 controls (6 written for this experiment, 7 from EXP 005), each translated blind into a static checker's constraints

The median expressible share across the 7 plugins with rules is 0.0%, below the pre-registered 25.0% bar: the premise is refuted: stage 2 is cancelled and the census is the result.

Pinned census gate · results.json sha256 26eb7287e88f · The results record