ĀRU MEANING LANGUAGE™ · RESEARCH PROTOTYPE

AI proposes.
The host decides.

An agent asks to use a tool. An app asks to show an interface. ĀML makes the host's policy decision and its evidence inspectable at both boundaries. Change one input below and follow the exact path.

proposal → exact rule → host decision → receipt

Connecting to the project preview. No tool runs, no message is sent, and no person is asked to approve.

ĀML / TWO BOUNDARIES

AGENT PROPOSAL · SEND_MESSAGE

Requested destinationoutsider@example.test
HOST RULE send_message → team@example.test only
Project planCALCULATING
01 / PROPOSALsend_message
02 / HOST RULEExact route
03 / PLANCalculating
04 / DISPATCHNot connected

No tool callback is connected to this page.

CANONICAL PROPOSAL / SHA-256Connecting to planner…Project vector self-check pending.
Project-authored preview · no live messageInspect this path ↗
01 / TRYStart with one decision ↗Change a tool destination or a render value, then follow the evidence. 02 / BUILDMake your first gate ↗Change declared inputs, inspect the browser decision, and download runnable HTML. 03 / AGENT TOOLGuard one function tool ↗Run a safe callback rehearsal, then wire policy and approval in your host. 04 / INTEGRATEStart from a clean package ↗Use the consumer rehearsal and inspect the release boundary. 05 / WRITE ĀMLOne line, new decision ↗Edit a program, watch its live browser gate, and share your source. 06 / CHALLENGEBreak a saved run ↗Open a checked example, change one saved outcome, and watch the claim fail.
FOR DEVELOPERS / EXECUTABLE CHALLENGE

Can your adapter survive the next case?

Start with one agent tool call. Run your own host adapter against 13 published cases, then six seeded changes that test whether it actually binds the proposal, rule, and outcome.

WHY A PERFECT PUBLIC SCORE ISN'T ENOUGH
Canned answer adapter13 / 19
Project reference19 / 19
Project-authored examples, executed locally. Scores do not certify an outside implementation or a real tool deployment.
THE RECORD CAN DISAGREE

Can you find the fracture?

Choose one controlled change to a captured run. Predict which check catches it, then reveal the answer and reproduce it in your browser.

Reproduce this change ↗A project-authored, unsigned consistency check. It cannot prove that a historical event occurred.
CAPTURED TWO-GATE EXAMPLECHALLENGE / GATE CLAIM
ORIGINAL RECORDCONSISTENTSHA-256 ✓   compiler ✓   gate claims ✓
FLIP ONE SAVED OUTCOME
WHICH CHECK WILL FAIL?
Prediction only · the checker runs the actual local recheck.
RESEARCH ATLAS / 13 LABSFind the experiment behind the boundary ↗Search by failure mode or choose a path through the 13 labs.
13 labs shownTopic only · search text stays here
NEW / CONNECTED MCP PILOTOne call crosses the wire. Zero pass the rule.

An official SDK client talks to a child MCP server over stdio. A misleading hint lets an outside call reach the local handler once; an AML host rule stops the same proposal before the server. Inspect six measured cases and the local audit.

Trace the connected call ↗
MODEL CONTEXT PROTOCOL / IN-MEMORY CASESA tool hint is not permission.

Nine additional in-memory callback cases distinguish policy denial, stale approval, tool errors, unknown outcomes, and unreadable results.

Inspect the earlier cases ↗
NEW / SHARED HOST CLAIMOne ID can cross workers.

Two actual Node workers each reach a simulated callback when their claim stores are separate. A private shared claim lets one through; the other receives a replay error. Inspect four executed cases.

Watch the worker race ↗
NEW / FALSE PERFECT-SCORE COUNTEREXAMPLE13/13 can still be wrong.

A canned adapter copies fixed answers and passes every public case. Six seeded transformations expose it at 13/19 while the project reference executes 19/19. Inspect the case-by-case evidence.

Watch the score break ↗
OUTSIDE IMPLEMENTER TOOLShow your own boundary result.

Run an adapter against 13 published cases plus six seeded probes and export a standalone visual report of actual verdicts and diagnostics. The published 13-case example is a project reference self-test.

Inspect the report format ↗
PREPARED REQUEST BOUNDARYCheck what will be sent.

Eight executed cases reveal a BCC added by host routing after an earlier check. Moving request construction before policy changes simulated provider callbacks from one to zero.

Follow the final request ↗
COMPLETE MAIL TOOLOne tool. Every recipient.

Seven executed cases compare a To-only function-tool mapper with a complete-recipient host adapter. An added BCC goes from one simulated callback to zero.

Follow the callback ↗
INTERACTIVE RECIPIENT WORKBENCHPut your own recipients through the target check.

Edit intended and proposed To, CC, and BCC fields. Compare exact resource keys and the local target outcome.

Try the workbench ↗
COMPLETE RECIPIENT MANIFESTOne visible recipient. Two actual destinations.

A to-only target check misses an added BCC. Seven local cases compare it with a complete, host-owned recipient manifest.

Inspect the second recipient ↗
EXECUTABLE TARGET BINDINGDeclared route. Different actual recipient.

Seven local cases show why a host must map the destination a tool will use before approving or dispatching it.

Inspect the mismatch ↗
REAL-SCREEN FIELD STUDY / TWO CAPTURED VIEWPORTSThe last call path is the test.

At the top, a repeated number can look expendable. After scrolling, the fixed dock may be the only surviving call route. Compare two project-authored candidates against the captured desktop states.

CAPTURED ORIGINAL · 3 call paths top / 2 after scroll

TOP OF PAGE2 call paths
AFTER SCROLL1 call path

Continuity guard passes · a call and email route remain in both states.

Project-authored candidates evaluated against two hashed desktop captures. The production site was not changed; no visitor outcome was measured.

ACTION BOUNDARY STUDIODraw the line before the call.

Edit the proposal and the host policy together. Inspect the exact decision, then export and recheck a portable calculation bundle.

Open the Studio ↗
POLICY X-RAYSame address. Different message.

An unreviewed exact route admits a changed body. See what a digest-bound host decision changes in eight executable cases.

Inspect the counterexample ↗
PRESERVE / RECHECK / EVOLVE

Built for the next interface. And the next.

An ĀML evidence capsule packages a decision receipt with a readable summary and two integrity digests. Keep the evidence when the screen, framework, or organization changes.

01 / CAPTUREKeep the original receiptDeclared intent, policy decisions, output, and audit evidence travel together.
02 / VERIFYRecheck every bindingRecalculate capsule digests, receipt integrity, summary, and any present signature.
03 / RENEWLink the next recordPreserve the original bytes. A SHA3-512 successor requires verifier-trusted witness quorum.
04 / RECOVERVerify across runtimesOpen a cold archive offline. Python checks capsules, signatures, and witnesses with external trust.
CAPSULE v1 · SHA-256 + SHA-512 · STRICT JSONExplore the preservation contract ↗

Read the append-only renewal contract ↗ · Pack an offline archive ↗ · Review versioned handoff ↗ · Repair from two shares ↗ · Model common storage failures ↗

Integrity alone does not establish truth, independent validation, trusted identity, or a guarantee of long-term cryptographic security.

Existence is not justification.

Conventional markup presents every declared element. The ĀML prototype inserts an inspectable policy decision before presentation.

Legacy rendering

Valid markup → rendered element.

ĀML accountable rendering

Declared meaning → policy evaluation → decision.

Operate the gate.

Change claimed restoration value. The same element may be allowed, degraded, or suppressed—and the rationale updates immediately.

The difference is inspectability.

01 · Meaning

Purpose and semantic value become explicit inputs, not invisible author assumptions.

02 · Decision

A named policy evaluates the element before the browser receives its final state.

03 · Accountability

The build emits both interface output and a machine-readable decision record.