Choose a case
The contract is loading.
Expected policy decision—
Expected receipt—
Expected tool calls—
Receipt reason—
Source of truth: machine contract. The browser holds a copy checked against that file in repository tests.
PUBLISHED CONTRACT · 13 CASES · NO LIVE TOOLS
Can your boundary hold?
An AI agent proposes a tool call. Your host freezes the proposal, checks its exact policy, binds approval to the digest, and reports what happened at dispatch.
These are project-authored expected results. This board does not run an adapter or claim an independent pass.
AGENT REQUEST → HOST BOUNDARYEXAMPLE / 03
01send_messageproposed
02outsider@example.testchanged destination
03No exact host rulepolicy decision
04BLOCKED0 tool calls
Illustration of the published changed-destination fixture; no message is sent.
13published adversarial cases
0live services or real approvals
PASS / FAIL / MIXEDoutside reports invited
THE FIVE-MINUTE IMPLEMENTER LOOP
One case. A precise difference. Then all 13.
Keep your adapter in your own repository. From a separate aml-core checkout, point the harness to its executable. Begin with the changed-destination denial; click another case below to change the command.
01 / BUILDRead one caseInput and expected receipt are published. Your process reads the case file and prints one JSON result.
02 / FIXRun that caseThe report names mismatched fields, with expected and observed values. A focused PASS covers one case only.
03 / VERIFYRun the full setThirteen cases plus manifest and vector hashes. Share PASS, FAIL, or MIXED from your own code.
node scripts/run-action-boundary-challenge.mjs --case changed-destination -- ../my-adapter/adapterRun from the aml-core checkout. Replace ../my-adapter/adapter with your executable. A focused report sets case_scope: "single" and full_challenge_passed: false, even when that case passes.
Every edge has an expected result.
Select a case to inspect its host fixture, expected receipt, and tool-call count. Filters show the expected outcome, not observed performance.
Loading the published cases…
Select a case to inspect its JSON.
YOUR CODE · YOUR REPOSITORY · REPRODUCIBLE REPORT
Make the claim falsifiable.
After the focused loop, run all 13 published cases, then add six seeded probes. The JSON report includes stable failure codes, field diagnostics, the seed, and hashes of the exact challenge and vector bytes. Keep your implementation and evidence outside the project reference code.
node scripts/run-action-boundary-challenge.mjs -- ../my-adapter/adapterWant a readable artifact from your actual run? Add --html-out. It creates a new offline HTML file with observed verdicts and diagnostics while keeping the machine JSON on stdout.
node scripts/run-action-boundary-challenge.mjs --html-out ./my-boundary-run.html -- ../my-adapter/adapterFor the stronger run, add --probes before --html-out. The Action does this by default and records the seed needed to repeat a run.
Run from a checkout of aml-core; replace the sibling path with your executable. The harness uses simulated callbacks only. A full PASS is compatibility evidence for these fixtures, subject to outside review of the implementation.