Bundle · Build & Ship
Ten ways to find outwhat your agentcannot refuse.
A model that is confident and wrong sounds exactly like a model that is confident and right. The only thing that separates them is something deterministic standing in the way — a declared rule, a bounded parameter, a contradiction check, a person who is actually reached. This is the whole line: ten checks for whether those things exist, in the order they depend on each other.
Deterministic and offline. Nothing connects, queries, or changes anything.
Grades models, contracts, actions and decisions — never people.
Instant download. Ten engines, ten workbooks, eighteen playbooks, one rollup.
01
Ten checks, in the order they depend on each other
Not the order they were built. The order that makes each one worth running.
01
01 · The model
Is the domain written down at all
Six dimensions per domain, and a gate that counts your entity types — because at twenty-eight types the distinctions nobody declared are where the contradictions come from.
MODELED · PARTIAL · UNMODELED · $89
02
02 · The map
Which decisions have which layers
One matrix read three ways. A layer covering three decisions out of forty is a pilot, not a layer — and the row with every cell full still is not four-layer.
FOUR-LAYER · THIN · UNVALIDATED · $99
03
03 · The values
Do the declared sets hold
The variant carrying 6,488 rows that is a second status rather than a typo. Volume tells them apart; spelling does not.
SETS HOLD · DRIFT FOUND · SETS BROKEN · $49
04
04 · The identity
Can the key tell two records apart
The pair your evidence says is one company and your data says cannot be. Merging destroys a fact; not merging leaves two records claiming to be one.
DISTINCT · SAME ENTITY · IDENTITY CONFLICT · $69
05
05 · The door
Before the call is made
A refund amount with a $10,000 ceiling that still cannot stop a $9,999 refund on a $12 order.
TYPED · LOOSE · OPEN DOOR · $79
06
06 · The ledger
Before the write commits
An action that is perfectly well-formed and still contradicts what your records already say.
ADMISSIBLE · REVIEW · INADMISSIBLE · $99
07
07 · The blast radius
What one wrong call reaches
A rollback you cannot finish inside its window is not a rollback. And a perfectly reversible action with no cap is still not unattended-safe.
RUN UNATTENDED · RUN WITH APPROVAL · DO NOT AUTOMATE · $89
08
08 · The catch
After the check says no
A path scoring 84 out of 100 that still fails open, because an approval step that auto-approves is a delay rather than a control.
FAILS CLOSED · FAILS SOFT · FAILS OPEN · $79
09
09 · The deliverable
Which facts were ever sourced
A fact that names its inputs, states its rule, and is still an assumption — because something three rows down was never sourced.
ASSERTED · DERIVABLE · ASSUMED · $89
10
10 · The next change
What the model change breaks
The change where everything still validates, every consumer still runs, and the answers are quietly different from now on.
ADDITIVE · BREAKING · SILENT BREAK · $79
You are not meant to run them all at once
Ten assessments in a week produces ten backlogs and no decisions. Run one, act on it, run the next — the rollup is built for exactly that, reporting whatever you have not reached as NOT RUN rather than quietly leaving it out, and naming which one is next. The order matters: four of these check against what the first one grades, so hardening tool contracts before the domain is written down is work you will do twice.
02
What the rollup prints
Point it at whichever runs you have. It answers what is worst, and what to run next.
THE ONTOLOGY LAYER - LANE ROLLUP
==============================================================================
1. Domain Ontology Readiness Survey NO MODEL
is the domain written down in a form a machine could check
that check says start at: D-04
2. Validation-Layer Coverage Map BLIND SPOTS
which decisions have which validation layers at all
that check says start at: D-06
3. Enum & Status-Value Drift Sweep SETS BROKEN
do the declared value sets hold in the data
that check says start at: order.status
4. Entity Identity & Functional-Key Sweep IDENTITY CONFLICTED
can the declared key tell two records apart
that check says start at: P-04
5. Tool-Input Schema Turnstile UNGUARDED
can a tool contract refuse a bad call
that check says start at: issue_refund.amount
6. Agent Action Admissibility Engine CONTRADICTIONS FOUND
does a proposed action contradict what is known
that check says start at: A-1002
7. Agent Side-Effect & Blast-Radius Checkpoint UNSAFE TO AUTOMATE
which actions can run without a human watching
that check says start at: A-04
8. Validation-Failure Backstop NO BACKSTOP
what happens when a check says no and nobody answers
that check says start at: P-06
9. Inferred-Fact & Unstated-Assumption Trace RESTS ON ASSUMPTION
which facts in a finished deliverable were ever sourced
that check says start at: Vendor risk memo
10. Ontology Change-Impact Tripwire SILENT BREAKS PRESENT
what will the next change to the model break, quietly
that check says start at: C-05
------------------------------------------------------------------------------
LANE VERDICT: LAYER BROKEN
10 of 10 run (0 clean / 0 with gaps / 10 broken)
worst finding: model (Domain Ontology Readiness Survey) -- D-04
Ties go to the earliest check, because the later ones assume it.
This rollup computes no score. It reports the worst of the verdicts those
engines already produced. LAYER HOLDS means all ten ran and all ten came
back clean on the data you entered - not that your agent is safe.This is a real run against all ten Systems’ own shipped samples, which is why every check is at its worst verdict. Two things are worth noticing. It reports the worst of the ten rather than blending them, because a blended number would hide which check is weak. And it breaks ties toward the earliest check — the later ones assume the earlier ones, so there is no point hardening a tool contract against a domain nobody has written down. One rule holds it together: a check you have not run is not a check that passed, so the lane cannot read LAYER HOLDS while anything is NOT RUN.
03
What you get
Ten complete Systems, each usable on its own, plus the rollup that only makes sense once you have run more than one.
Ten engines
Zero-dependency Python, one per check. Each runs offline, reads only the files you point it at, and prints a verdict you can paste into a ticket.
Ten workbooks
Every engine reproduced as live formulas, no macros. Excel, Google Sheets, or Numbers.
Eighteen playbooks
How to fill each check in honestly, how to read its verdict, and who owns each fix.
Ten worked samples
A teaching pair inside each one that isolates exactly what moves a verdict — the $9,999 refund on a $12 order, the variant carrying 6,488 rows, the fact that was never sourced.
The lane rollup
One script, whichever runs you have. It reports each verdict, takes the worst, forwards that engine's own fix-first, breaks ties toward the earliest check, and reports anything unrun as NOT RUN. No combined score.
Editable rule packs
Every dimension, threshold, weight, guard requirement and constraint lives in a JSON file you own.
04
What it will not do
Ten honest checks do not add up to a guarantee, and this page will not pretend otherwise.
Built for
- Teams putting an agent on a domain nobody has written down, who want an order to work in.
- Anyone asked to show that guardrails exist, and finding that “we have guardrails” is not an answer.
- Architects with a finite budget deciding between deepening one decision and widening one layer.
It will not
- Read your code, parse your schemas, query your database, or call any tool. You describe; they grade.
- Build, merge, migrate or enforce anything. They name the field, the parameter, the path and the decision; the work is yours.
- Tell you a check you skipped would have passed. Anything unrun is reported as unrun.
- Turn ten passing checks into a safety claim. LAYER HOLDS means ten checks came back clean on the data you entered.
These Systems grade models, contracts, actions, paths and decisions you describe. They are working aids, not legal, security, or compliance advice, and not audits or certifications. None of them scores, ranks, or assesses people. LAYER HOLDS proves all ten ran clean 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.
All ten Systems and Quick Kits in the line, each sold separately, plus a lane rollup. Together they cover the whole sequence: whether the domain is written down, which decisions have which validation layers, whether the declared value sets and identity rules hold in the data, whether tool contracts can refuse a bad call, whether an action contradicts what is known, what one wrong call can reach, what happens when a check says no, which facts in a deliverable were ever sourced, and what the next model change will break quietly. $820 separately, $629 together.
It answers the two questions a ten-part assessment actually raises: what is worst, and what should I run next. It reads whichever engine outputs you have, reports each verdict, takes the worst, and forwards that engine's own fix-first. Ties go to the earliest check in dependency order, because the later ones assume it — there is no point hardening tool contracts against a domain model that does not exist yet.
Because a check you have not run is not a check that passed. Three clean results out of ten is three clean results, not a clean layer — and a rollup that reported LAYER HOLDS on a third of the evidence would be the most dangerous thing in the box. Until all ten have run, the best available verdict is GAPS FOUND, and the rollup names the next one to run.
No, and the rollup is built to be useful from the first one. Point it at whatever you have and it reports the rest as NOT RUN rather than guessing or quietly omitting them. That is the intended way to work through the line: run one, act on it, run the next.
Yes, and it is the order the rollup lists them in. The Domain Ontology Readiness Survey comes first because four of the others check against what it grades. The Validation-Layer Coverage Map comes second because it tells you which decisions are worth pointing the rest at. Then the two data sweeps that test the model's claims, then the door, the ledger, the side effects and the catch, then deliverables and change impact.
That all ten ran and all ten came back clean on the data you entered. Not that your agent is safe. Every one of the ten is explicit about its own limit — none of them reads your code, queries your database, or tests a live system — and stacking ten honest checks does not turn them into a guarantee. They grade models, contracts, actions, paths and decisions. Never people.
The whole line, in the order it makes sense.
Ten complete checks and the rollup that tells you what is worst and what to run next. $629 together, $820 apart.
These Systems grade models, contracts, actions, paths and decisions you describe. Working aids, not legal, security, or compliance advice, and they never score or rank people.
Sold by RedHub AI LLC · Secured by Stripe · redhub.ai