Records what your system does today, then fails any new version that silently changes it. Crawl, snapshot, diff, signed report.
neuruh-verify← Top 10 commercial products
Harnesses · verification as a product
We test whether an AI system actually works.
Before you ship an agent, accept a vendor’s AI, or let a release touch production, a harness engagement answers one question with reproducible evidence: does it still do only what it is supposed to do? You get one decision and a signed receipt, not a demo and a promise.
Who it is forTeams shipping AI agents into real workflows, and the people who have to accept them.
What you getA baseline, a case set, the delta and regressions, permission and evidence checks, one decision, a signed receipt.
Starting offer$2,500 pilotOne workflow · test commercial pricing · quoted by email
01
What happens, in order
Every engagement runs these thirteen stages in this order. Nothing is skipped; a stage that cannot run is recorded as blocked.
01
INTAKE — You name the workflow, the version you want to ship, and who gets to say yes or no.
02
SCOPE — One workflow, what it is allowed to do, and what evidence counts, written down.
03
BASELINE — How the current version behaves is captured and frozen. That is the yardstick.
04
GOLDEN CASES — Cases you consider correct. The new version must keep passing them.
05
HELD-OUT CASES — Cases the builder never saw, so the test cannot be gamed.
06
ADVERSARIAL CASES — Injected instructions, secrets, oversized input, boundary crossings. They must be refused.
07
RUN — The new version and the baseline run the same cases under the same conditions.
08
DELTA — What changed, case by case.
09
REGRESSIONS — Anything the old version passed and the new one fails. Listed, never averaged away.
10
AUTHORITY CHECK — Did it stay inside its permissions? Stepping outside is a failure, not a footnote.
11
EVIDENCE CHECK — Is the evidence it cites actually there, and does it say what is claimed?
12
DECISION — One word: PASS, SHADOW, MORE_EVIDENCE, or REJECT, derived from the checks above.
13
SIGNED RECEIPT — Identity, cases, results, and decision bound together and signed by something that is not the thing under test.
02
See it run in your browser
Load the sample pair, press run, and watch the thirteen stages. The sample candidate drops the evidence the baseline had, so the demonstration ends in a REJECT with the regression named. The checks are structural, the digest is unsigned, and nothing touches a production system.
BROWSER DEMONSTRATION — NOT A PRODUCTION ASSURANCE RESULTRuns entirely in your browser on numbers or text you type. Paid work runs in the governed systems described above and returns a receipt; this panel does not.
- 01INTAKE
- 02SCOPE
- 03BASELINE
- 04GOLDEN CASES
- 05HELD-OUT CASES
- 06ADVERSARIAL CASES
- 07RUN
- 08DELTA
- 09REGRESSIONS
- 10AUTHORITY CHECK
- 11EVIDENCE CHECK
- 12DECISION
- 13SIGNED RECEIPT
Ready.
Load the sample pair or paste your own, then press run. Each stage lights up as the browser walks the sequence; the result shows the checks, the delta against the baseline, the decision, and an unsigned digest.
03
The harnesses behind the product
Neuruh runs these on its own systems. A pilot applies the relevant ones to your workflow. None is a self-serve tool today.
Runs a candidate against held-out and adversarial cases, measures the difference from the baseline, and proposes promote or reject for a human to decide.
neuruh-factoryRefuses to accept a result unless the receipt, the execution record, the permission record, and the identity of what ran are all present and bound together.
neuruh-sentinel-courtProves an agent still knows who it is and what it is allowed to do after memory, context, restarts, and a swapped model provider.
context / memory pathChecks health probes, restarts, pinned versions, dependency order, bounded logs, and that the system halts closed when something is wrong.
runtime supervisionFreezes cross-system cases from ingestion through decision and replays them read-only against production to prove the chain still holds.
neuruh-live-integrationChecks that what a system cites actually exists and says what is claimed, and keeps a hash-chained record of every check.
crucibleBounded repair loops that must pass verification before a fix is accepted.
ncellConfirms a running system matches its pinned, courted version before it is trusted.
neuruh-cockpitThe pilot deliverable: your workflow’s zero-regression receipt, packaged for you rather than kept as internal telemetry.
Agent Integrity SuiteStatus meanings. USED ON NEURUH RELEASES: gates this site or a production system today. INTERNAL · COURTED: exists with its own tests and receipts, runs inside Neuruh only. RESEARCH: exists, not part of a release gate. PROPOSED: not built. None of these harnesses is externally callable; a pilot is run by a founder against your workflow. Repository names refer to private repositories.
04
What you buy
- One workflow mapped to explicit permissions and expected outcomes.
- Golden, held-out, and adversarial cases written against it.
- A baseline capture and a candidate run under the same conditions.
- Delta, regressions, permission and evidence checks in one report.
- A decision and a signed receipt naming the exact tested version.
- Optional re-runs on each release, quoted.
05
How you use it
- Bring the agent, workflow, or release candidate and the person who decides.
- Agree the allowed actions and the evidence boundary in writing.
- Neuruh captures the baseline and writes the cases; you review them.
- The run produces the report and the decision.
- Repair, rerun, and seal, or reject and stop.
Evidence that this is how Neuruh already works
This site’s own release court records 0 broken links and 0 serious or critical accessibility violations at the shipped version. The DeedSonar production court records 1,510 tests green at the version production attests. Both are dated and compiled on /proof. No external customer pilot has been run yet; the hard zeros on /refusals apply.