Runtime

This page described something exact.works no longer offers.

WITHDRAWN: 2026-09-04
UNDER: Scope decision of 2026-09-01 (EXACTING.md)

Runtime was a behavioural verification engine: while an agent executed,
exact.works scored its conduct against a declared fingerprint and could suspend
the session. Scoring, suspending and holding funds on the outcome are enforcement,
and enforcement is out of scope.

What this page used to say

That exact.works watched an agent while it ran — route conformance, scope
conformance and output pattern, weighted into a composite score — and that
crossing a threshold flagged the session, suspended execution with the money
held, or escalated it automatically to a dispute engine.

Two separate things there are now out in terms: holding anyone's funds, and
deciding a dispute. A
"wrongful suspension remedy" only exists because there was a suspension to be
wrong about, and there is not one any more.

The gap was real; the answer was wrong

An agent that acquires capabilities nobody agreed to is a genuine problem, and
this page named it earlier than most. But the answer is not a third party
watching the execution. NeMo Guardrails, Azure AI Content Safety and LangSmith
own runtime control, and they are where a stop belongs.

What was missing there was never the knobs — it was knowing what to set them
to.

What is true now

exact.works helps two parties write down the envelope an agent has to stay
inside, and seals the result so neither side can later revise it. It does not
watch, score, flag, suspend or escalate anything.

What replaced it

The agreement, and what the parties owe each other — the envelope an agent must stay inside is a term between the two parties, not a rule a third party applies

What exact.works does now Home

Runtime

This page described something exact.works no longer offers.

WITHDRAWN: 2026-09-04
UNDER: Scope decision of 2026-09-01 (EXACTING.md)

Runtime was a behavioural verification engine: while an agent executed, exact.works scored its conduct against a declared fingerprint and could suspend the session. Scoring, suspending and holding funds on the outcome are enforcement, and enforcement is out of scope.

What this page used to say

That exact.works watched an agent while it ran — route conformance, scope conformance and output pattern, weighted into a composite score — and that crossing a threshold flagged the session, suspended execution with the money held, or escalated it automatically to a dispute engine. Two separate things there are now out in terms: holding anyone's funds, and deciding a dispute. A "wrongful suspension remedy" only exists because there was a suspension to be wrong about, and there is not one any more.

The gap was real; the answer was wrong

An agent that acquires capabilities nobody agreed to is a genuine problem, and this page named it earlier than most. But the answer is not a third party watching the execution. NeMo Guardrails, Azure AI Content Safety and LangSmith own runtime control, and they are where a stop belongs. What was missing there was never the knobs — it was knowing what to set them to.