The AI decided something.Who authorized that?
Every other tool in this catalog ends at a verdict a human reads. This lane is about what happens after — whether the verdict binds a machine, whether anybody can stop it, and whether anyone can answer for it afterwards.
- Runs offline. No accounts, no telemetry, nothing to connect.
- Returns a genuine verdict, including the one you did not want.
- Grades a system or a record — never a person.
- Reports at least one number it refuses to score you on.
- Ships a worked example that fails, because most real ones do.
Between the model being right and anyone being able to answer for it.
An enormous amount of attention goes into whether an AI system decides well. Comparatively little goes into the machinery around the decision: whether the verdict actually stops the tool, whether the conditions attached to an approval do anything when they are missed, whether the thing can be halted once it is running, and whether — six months later — anybody can reconstruct what happened and why.
That machinery is the execution layer. It is where automated decisions quietly stop being decisions and start being events nobody owns.
Three stages, in order.
Every instrument stands alone and answers its own question honestly. The order matters because a later instrument cannot repair an earlier failure — perfect audit evidence about a verdict that never bound a machine is evidence about nothing.
Where most estates should start, and the stage most often skipped — its findings are the least presentable and they change what everything downstream means.
What must a decision payload carry before a machine can act on it?
Does the gate hold the tool, or merely sit in front of it?
When an approval comes with conditions, does anything happen if they are missed?
Intervention and accountability for intervening are one stage, deliberately. A stop nobody can be named for is not meaningfully different from no stop at all.
When it escalates, does it reach somebody who can actually decide?
What has your kill switch actually been shown to do?
When somebody bypasses a control, is anybody's name on it?
The most-visited stage and the most frequently moot. Four instruments can read at the very top here while describing a decision that never bound anything.
The agent acted — did your books ever hear about it?
Is a verdict you are still acting on still good?
Could somebody who was not there assemble the account?
Could you produce that answer again?
None of these instruments has a model in its decision path.
A language model in a decision path is an unpinnable source of variance. Temperature zero does not survive a model version change, a hardware change or a provider-side update, and none of those are events you control or are reliably told about. So a decision made with a model in the path cannot be reproduced — at best it can be reconstructed from a record.
That is a fine property for a great many systems and a poor one for an instrument you are using to check your systems. Every engine in this catalog is deterministic code deciding from a config: same inputs, same answer, on any machine, on any day. Run one twice and you get the same verdict — which is the least you should expect from something telling you your controls are decorative.
Each drop ships a Python engine with no dependencies, a workbook that reproduces its arithmetic, and a browser demo that agrees with both — checked byte-for-byte before release.
Every instrument here can return the answer that costs us the sale. One in five pieces of our writing shows a tool declining, because a tool that only ever agrees is not measuring anything.
Each instrument reports at least one figure and compares it against nothing — how many gates you have, how often a path runs, how long you keep records. Volume is not safety.
Three instruments run in order beat ten run at random.
Verdict Envelope Standard — it sits at the top of the binding stage, and it is the only instrument the other nine can emit their findings into.
The bundle adds a roll-up engine that reads all ten in order and names the stages whose readings are moot. That mechanic cannot exist inside any single instrument.
Instruments you have not run are never guessed at. Whatever you skip is reported as not run — counted neither for you nor against you.
Before you pick one.
Somebody will ask who authorized it.
Better to find out now.
Ten instruments and the roll-up. $940 bought separately, $719 together.
Scope: these are engineering instruments. They grade systems and records, never people, assert no regulatory position, and are not legal advice.
Sold by RedHub AI LLC · Secured by Stripe · redhub.ai