ĀML™ / ĀRU INTELLIGENCE
ENGINEERING EVIDENCE / LOCAL DISPATCH

The dispatch proof atlas.

One host boundary. Two resource scopes, four policy states, six approval behaviors, four claim behaviors, two callback outcomes. Explore every recorded combination, then inspect the receipt and the local effects that led to it.

CALLBACK RETURNED—

Dispatched in a local fixture

CALLBACK THREW—

May follow a local write

STOPPED BEFORE CALLBACK—

Claim may still be consumed

FINITE COMBINATIONS384

Exhaustive only over these axes

FOUR IMPORTANT EDGES

What a receipt hides.

These counts describe the finite matrix, not a production rate. Each card opens one exact example below.

THE COMPLETE CROSS-PRODUCT

Every local path.

Each square is one vector. Columns are policy states; rows are approval-hook behaviors. Each row has an original-scope lane and an other-resource-rule lane. Squares follow the four claim behaviors, with a returning callback then a throwing callback. Select one to inspect its path.

Loading vectors…
DispatchedUnknownBlocked↗ callback returns · ! callback throws

Colors are supported by text labels and each square has an accessible name. Horizontal scrolling reveals all four policies on narrow screens.

SOURCE MUTATION CHALLENGE

A new gap, then a smaller proof.

The first mutation challenge found a resource-rule gap and expanded this matrix from 192 to 384 cases. A new challenge found two more blind spots: weakened tool and effect checks. An eight-case rule-match lattice covers those comparisons.

Explore the three-field gate ↗ · Inspect the receipt-binding fix ↗ · Mutation report ↗
10/10SELECTED SOURCE CHANGES
DETECTED ACROSS THREE SUITES
192OTHER-RESOURCE CASES
BLOCKED BEFORE CALLBACK

The 384-case matrix detects seven selected source changes, including all 192 other-resource rows under a weakened resource check. The smaller three-field lattice detects tool and effect changes the matrix cannot see. A changing-getter fixture exposes a receipt that once hashed different bytes than the callback received. These are finite project-authored fixtures, not arbitrary defect coverage.

What was actually checked?

The Node runner calls the reference dispatchAction for all 384 local fixtures. A separate project-authored Python oracle checks each row, including receipt, digest, claim consumption, callback target, and marker write. CI rejects five deliberately falsified reports.

Reproduce with node pilots/action-boundary/dispatch-matrix.mjs --check, python3 pilots/action-boundary/verify_dispatch_matrix.py docs/action-dispatch-matrix.json, , python3 pilots/action-boundary/verify_policy_tuple.py docs/action-policy-tuple.json, and node pilots/action-boundary/dispatch-mutants.mjs --check. Read the method and limits ↗

Scope: Finite project-authored callbacks and selected mutations, no human approval, provider call, arbitrary interleavings, real delivery, outside witness, or formal proof of production safety. A browser view of the report does not independently verify its provenance. Never silently retry an unknown result.