A correct decision can expire.

Reasoning is not authorization.

An AI agent can make the right decision at T0 and still arrive at execution after the conditions that justified it have changed.

Current-state preflight asksIs the action acceptable now?

Valid Until asksIs this still the same justified action?

Open Live Proof Lab →

AI agent interprets + proposes → Binance Agent OS supplies fresh observations + capabilities → Valid Until seals + revalidates policy, T0 state and exact action

Portable skill · skills/valid-until/SKILL.md · MCP contract · valid_until_begin → valid_until_revalidate

Execution validity replay · BTCUSDT

6-case red team →

01 · Decision contract / T0

BTCUSDT

Exact actionBUY · loading

Policy max$100

Mid at decision—

Spread—

ReceiptV2 · ACTION-BOUND

StatusELIGIBLE

SEALED BEFORE EXECUTION
policy loading…
snapshot loading…
action loading…

02 · Cross-time validity window

Fresh does not mean same premise.

00:05validity remaining

0 bps2035.47

0.00

bps
from T0

The exact action remains valid only while fresh evidence still belongs to the sealed premise that justified it.

03 · Execution boundary / T1

STILL VALID

Current-state checksPENDING

Exact action hashMATCH

Cross-time premiseWITHIN BOUND

Execution path openNo sealed contract invariant has failed.

01Policy + action frozen

02Binance state read

03Receipt v2 bound

04Fresh state + action rechecked

05Allow / replan

Current state checks · T1PENDINGspread · 1m movement · bid depth · ask depth

≠

T0 → T1 premise driftWITHIN BOUNDT0 → T1 drift against the sealed contract

A correct decision can expire.Policy + T0 state + exact action are one contract.

LIVE
PROOF

Official Binance CLI · real BTCUSDT · fail-closed

Public market data only. No API key, no account, no trade. Live authenticity remains separate from the deterministic judge replay.

Core CI8/8 + MCP 5/5Red team6/6

Three distinct evidence classes.LIVE observations, CONTROLLED replay and VERIFIED EXECUTION are separate. Inspect historical Testnet order 13634770 →. No new order.

Agent-native by design

The model can propose. It cannot authorize itself.

01Human policy

02AI proposal

03Agent OS state

04Validity contract

Portable skillnpx skills add Faadil1/valid-until-agent-os --skill valid-until -y

Deterministic MCP companionnpm run mcp · valid_until_begin · valid_until_revalidate

Four different questions

Reasoning verifier

Was the model's reasoning supported?

Reasoning

Current preflight

Is the action acceptable under current conditions?

Now

Permissions

May the agent use this capability?

Scope

Valid Until

Is this exact old action still the same valid decision contract?

Cross-time