Bundle · Build & Ship

Three ways to say no,at the three momentsan agent can do damage.

An agent can be stopped before the call is made, before the write commits, or after a check says no. Most teams have built one of those three and assume they are covered. They fail independently, and the one nobody tests is the third: the approval step that quietly approves when nobody answers it.

Deterministic and offline. Nothing connects, calls, or changes anything.

Grades contracts, actions and paths — never people.

Instant download. Three engines, three workbooks, six playbooks, one chain rollup.

01

Door, ledger, catch

Three refusals, in the order a request meets them. Each catches something the other two cannot see.

01

The door

Before the call is made

A parameter that would accept anything. A refund amount with no ceiling, a delete filter with no bound, an identifier with no pattern.

Tool-Input Schema Turnstile

TYPED · LOOSE · OPEN DOOR · $79

02

The ledger

Before the write commits

An action that is perfectly well-formed and still contradicts what your records already say. Two managers for one employee. A party that cannot be both things at once.

Agent Action Admissibility Engine

ADMISSIBLE · REVIEW · INADMISSIBLE · $99

03

The catch

After the check says no

A review step nobody answers, on a timer nobody chose. An approval that auto-approves is a delay, not a control.

Validation-Failure Backstop

FAILS CLOSED · FAILS SOFT · FAILS OPEN · $79

Why fixing one does nothing for the other two

A perfectly bounded parameter still lets through an action that contradicts your ledger. A flawless admissibility check still hands the decision to a review step that approves on timeout. And the best escalation path in the world never sees the call that a typed-but- unbounded parameter waved through. They are independent, which is why doing one and stopping feels like progress and is not.

02

What the chain rollup prints

Run all three, feed their JSON to one script, and it tells you which layer to work on first.

THE REFUSAL TRIO - CHAIN ROLLUP
==========================================================================

The door     UNGUARDED               <-- weakest
    Can the tool contract refuse a bad call before it is made?
    Tool-Input Schema Turnstile
    fix first here: issue_refund.amount

The ledger   CONTRADICTIONS FOUND
    Does this action contradict what is already known?
    Agent Action Admissibility Engine
    fix first here: A-1002

The catch    NO BACKSTOP
    What happens when the check says no and nobody answers?
    Validation-Failure Backstop
    fix first here: P-06 (timeout_behavior)

--------------------------------------------------------------------------
CHAIN VERDICT: CHAIN BROKEN
  0 of 3 layers holding; weakest is door (Tool-Input Schema Turnstile)
  start at: issue_refund.amount

  Ties go to the earliest layer. A loose door makes the ledger's job
  harder and sends more traffic to the catch, so repairing upstream
  beats compensating downstream.

This rollup computes no score. It reports the worst of three verdicts
the three engines already produced. Three layers holding proves those
three checks pass - not that your agent is safe.

This is a real run against the three Systems’ own shipped samples, which is why every layer is at its worst verdict. Two things are worth noticing. It reports the worst of the three rather than blending them, because a blended number would hide which link is weak. And when all three tie, it points at the door — repairing upstream beats compensating downstream, since a loose door is what sends the extra traffic to the other two in the first place.

03

What you get

Three complete Systems, each usable on its own, plus the piece that only makes sense once you have more than one.

Three engines

Zero-dependency Python, one per layer. Each runs offline, reads only the files you point it at, and prints a verdict you can paste into a ticket.

Three workbooks

Every engine reproduced as live formulas — 164, 119 and 88 of them — with no macros. Excel, Google Sheets, or Numbers.

Six playbooks

Two per System: one on how to fill it in honestly, one on how to read the verdict and who owns each fix.

Three worked samples

Twelve parameters, eight proposed actions, six validation paths — each built around a teaching pair that isolates exactly what changes a verdict.

The chain rollup

One script, three JSON files, one answer: which layer is weakest and what that engine says to fix first. No combined score.

Editable rule packs

Every threshold, weight, guard requirement and constraint lives in a JSON file you own. Change what money requires; the engine checks your weights still sum to 100.

04

What it will not do

Three honest tools do not add up to a guarantee, and this page will not pretend otherwise.

Built for

  • Teams shipping agents that write — refunds, payouts, deletions, grants, status changes.
  • Anyone who has done one of these three and wants to know honestly whether the other two hold.
  • Reviewers who need a defensible reason to send something back rather than a feeling that it looks thin.

It will not

  • Read your code, parse your schemas, query your database, or call any tool. You transcribe; they grade.
  • Enforce anything. They tell you which parameter is a blank cheque and which path fails open; building the fix is yours.
  • Test whether your timeout behaves the way you marked it. That is a drill you run, and the playbook says how.
  • Turn three passing checks into a safety claim. CHAIN HOLDS means those three checks passed on the data you entered.

These Systems grade contracts, actions and paths you transcribe. They are working aids, not legal, security, or compliance advice, and not audits or certifications. None of them scores, ranks, or assesses people. Three layers holding proves those three checks pass on the data you entered — not that your agent is safe.

05

Questions people actually ask

Short answers first, with the caveat after it rather than in front of it.

Three complete Systems, each sold separately, plus a chain rollup that reads all three. Tool-Input Schema Turnstile ($79) grades whether your tool contracts could refuse a bad call. Agent Action Admissibility Engine ($99) grades whether a proposed action contradicts what your records already say. Validation-Failure Backstop ($79) grades what happens when a check says no and nobody answers. $257 separately, $199 together.

The layer nobody tests is the one that decides.

Three complete Systems and the rollup that names your weakest link. $199 together, $257 apart.

These Systems grade contracts, actions and paths you transcribe. Working aids, not legal, security, or compliance advice, and they never score or rank people.

Get all three — $199

Sold by RedHub AI LLC · Secured by Stripe · redhub.ai