Independent verification for AI agents

Prove your agent actually did the job.

Agents report their own success. av9n checks whether it's true — against evidence the agent doesn't control, sealed in a record it can't rewrite. It catches the expensive failure everything else misses: an agent confidently reporting work that never happened.

  refund-agent · batch RB-2213
AGENT CLAIMED
"All 50 refunds processed successfully."
av9n VERDICT · checked against Stripe
REFUTED
7 of 50 failed or duplicated — the report was wrong
Checked against independent evidence
Sealed, tamper-evident record
See it run — no signup

Watch av9n catch an agent that lied about the job.

Pick a scenario. The agent reported success. av9n checks it against independent evidence — the payment processor, the CI system, the ticket log — and shows you what the report hid. This is the real pipeline, running live.

Create a free account Try the full demo →
The gap

Observability shows what an agent did. av9n proves what actually happened.

Traces, dashboards, and eval tools all read what the agent reports about itself. A green "done" just means the agent said it worked.

av9n checks that claim against systems the agent doesn't control — the payment processor, the CI pipeline, the system of record — and seals the result in a tamper-evident record. It catches the agent that confidently reports success on work that never happened.claim → checked against independent evidence → verdict

The method

Every score survives the four questions an auditor asks.

Not a verdict on activity — a defensible verdict on value. Each score decomposes into the four things a skeptic will challenge, answered from sealed evidence.

Q1
Compared to what?
Value is measured against a baseline captured before the agent ran, or a live holdout. No counterfactual, no credit.
counterfactual
Q2
Says who?
Captured at execution and sealed with an asymmetric signature. Anyone can verify it — no access to av9n required.
tamper-evident
Q3
Which agent?
In a multi-agent chain, credit splits along the causal path. Two agents can never each claim the same dollar.
no double-counting
Q4
Net of what cost?
Tokens, tools, infra, and the human review time everyone hides. Work that costs more to check than it saves scores negative.
fully loaded
Beyond the score

Verification that keeps working after go-live.

A one-time check is a snapshot. Real trust means catching the moment an agent starts to slip, weighting evidence by how independent it actually is, and a record anyone can verify without trusting us.

Drift detection
Statistical process control watches each agent's verified success rate and flags degradation the moment it starts — before failures compound. It tells you when it began, not just that it happened.
early warning
Evidence independence
Not all evidence is equal. A verdict backed only by the agent's own logs is discounted; a verdict confirmed by systems the agent can't touch carries full weight. Independence is scored, not assumed.
on by design
Portable credentials
An agent's verified track record travels with it — a signed credential anyone can verify on any platform, without calling av9n and without trusting the agent's owner.
verify anywhere
Anchored to Bitcoin
The verdict history is committed to an append-only ledger and anchored to Bitcoin, so not even av9n can rewrite a past result undetected. Built to the RFC 6962 standard.
tamper-proof
Integrate

Wrap an agent. Get a provable score.

One decorator around your agent's tool calls. av9n seals every action, checks the output, and returns a health score you read in code or watch on the dashboard.

agent.py
from av9n import Av9n, Baseline

k = Av9n(workflow="invoice-run")
k.baseline(Baseline(unit_cost=18.00, method="holdout"))

# wrap any agent tool — av9n seals + scores every call
@k.track(agent="extractor")
def extract_line_items(pdf):
    return model.run(pdf)

score = k.health()
# HealthScore(grade="A", value=+787, assurance=0.91, flags=[])
The sealed record

Every action is sealed the moment it happens.

av9n records each agent action at the execution gate and hash-chains it into a signed ledger. Change any past entry and the chain breaks — anyone can catch it with the public key alone. This is the record every score is built from.

ledger · invoice-run · 4 events independently verified
01extract_line_items · extractor-agenta3f1…9d2e
02validate_against_po · validator-agent7c04…b1a8
03human_review · ap-clerk-jane · 2.0 mine5d9…4f37
04outcome · +$18.00 value3e72…3c63
sealed head 3e72c5f9b688…3c63 · ed25519 tamper any row → chain breaks → verify() = false
The score

One number, and you can see every part of it.

The health score isn't a black box — it's a formula over evidence. Cheap work that's wrong scores negative, because a wrong answer is a liability, not a zero. That's the part activity dashboards can't show.

gross value × assurance
cost (tokens + tools + infra + review)
inflation penalty = health
Correct agent
verified · assurance 0.91
+$787
Unsupported agent
plausible · unverified · flagged
+$400
Confidently wrong agent
refuted · same $1k claim, same cost
−$20
The guarantee

An agent can't inflate a number it doesn't author.

av9n rejects any score where the agent being valued controls its own inputs. The value comes from a baseline someone else set, costs measured at the gate, correctness checked against outside sources, and a sealed trace. Try to self-author it, and the computation refuses.

rejected.py
# agent tries to set its own baseline + value
k.health(subject="agent-X",
         baseline_by="agent-X",
         value_by="agent-X")

InflationViolation: rejected — inputs are self-authored
  · baseline authored by the subject
  · gross value self-asserted by the subject
  · costs self-reported, not measured at the gate
1Baseline set by someone else. The counterfactual can't come from the agent whose savings it measures.
2Value derived, not declared. Worth is attributed from the sealed trace — the agent never types its own number.
3Cost measured at the gate. Tokens, tools, and human review are recorded, not self-reported.
4Correctness checked outside. Assurance comes from authoritative sources, not the agent's own say-so.
Calibration

The score knows how much to trust itself.

av9n measures its own verifier and marks a score board-grade or directional. Uncertain claims flow to review, get labeled, and the score sharpens automatically.

01
Flag
Low-assurance claims surfaced
02
Label
Humans confirm, kappa-checked
03
Grow
Gold set expands
04
Recalibrate
Miss rate drops
05
Certify
Score turns board-grade
For agent builders

Prove what your agent is worth — to someone who won't take your word for it.

Shipping or selling an agent? av9n turns "trust me, it works" into an evidence bundle a buyer can verify themselves. You produce the proof. You just can't rig it — which is exactly why it's believed.

YOU — the builder
Wrap your agent with one decorator and run it on real work.
av9n seals every action and scores the output as it happens.
Export a signed evidence bundle — score, sealed trace, and its full four-question defense.
signed
evidence
bundle
THE BUYER
Verifies the seal with the public key — no access to you or av9n required.
Sees the health score net of every cost, not a cherry-picked demo.
Trusts it because you couldn't author the inputs that set the number.
why it counts
The baseline, the correctness check, and the cost all come from outside your control — the same reason a company can't audit itself. av9n is the proof you can hand over precisely because you couldn't have faked it.
Enterprise

Built for teams that get audited.

av9n produces evidence, so it holds itself to an evidence standard. Here's what's live today, and what's on the roadmap for enterprise deployments — stated plainly, because a verification company shouldn't claim controls it hasn't shipped.

Ed25519 signed records Tamper-evident sealed log Independently verifiable — public seals Zero training on your data Adversarial suite — resistance to gaming

On the enterprise roadmap, delivered with design partners: SOC 2 Type II, SSO / SAML & SCIM, role-based access with audit logs, data residency, and a self-host option. We'll say when each ships — not before.

Pricing

Free to prove it to yourself. Paid to prove it to others.

You meter on verified scores, not seats or traces — price tracks the thing that's actually valuable. Certification, the board-grade seal and the exportable evidence bundle, is the paid unlock.

Developer
$0 / mo
Prove worth to yourself, in code.
250 verified scores / month
Full SDK & sealed ledger
Health score + four-question breakdown
Directional grade only
Community support
Start free — create account
Team
$199 / mo
Certify agents and hand proof to buyers.
25,000 verified scores / month
Verification badge unlocked
Exportable evidence bundles
Overwatch dashboard, unlimited seats
Calibration flywheel + gold sets
Overage $8 / 1,000 scores
Get started with Team
Enterprise
Custom
Volume, control, and compliance.
Unlimited verified scores
Private verifier tuning
Adversarial / red-team suite
SOC 2, SSO/SCIM, audit logs (roadmap)
Self-host & data residency (roadmap)
SLA & dedicated support
Talk to us

Prices are early illustrative figures — the real numbers come from talking to buyers. Verified-score meter, not per-seat or per-trace.

av9n Verified

A live trust badge you put everywhere your agent is.

Certify your agent and get an embeddable badge — served live by av9n, so it's always current. Put it on your site, your repo, your marketplace listing. Every badge links to a public verification page showing the agent's real, tamper-evident score right now.

The badge, live on your agent: av9 av9n Verified A verified today · score current Drop it in with one line:
<a href="https://av9n.ai/verify/your-agent"> <img src="https://av9n.ai/v1/badges/your-agent/svg" /></a>
Served live, never a saved image. av9n controls what the badge says — if your agent degrades, the badge updates everywhere at once. That's why buyers trust it.
Links to a public verification page. One click shows the current score and the sealed evidence behind it — proof, not a claim.
Goes anywhere your agent is listed. Website, GitHub README, agent marketplaces, directories, your product UI. Each placement is a trust signal buyers learn to look for.
Revocable. Stop maintaining a passing score and the badge stops verifying. Because it's served live and re-checked on a cadence, a copied or stale image can't keep claiming a status that's no longer current.
Verified — Individual
$99 / yr · per agent
The on-ramp. Get listed, get trusted.
Live embeddable badge
Public verification page
Continuous re-scoring
Auto-revocation if it degrades
Get Verified
Verified — Pro
$299 / yr · per agent
For agents you sell or ship to buyers.
Everything in Individual
Priority re-scoring
Custom badge styling
Full historical proof log
Exportable buyer evidence bundle
Get Verified Pro
Marketplace
Volume
Directories that verify their listings.
Bulk verification of listed agents
"av9n Verified" required-for-listing
Directory-wide verification API
Co-branded trust page
Talk to us

Recurring, not one-time — a live badge only stays trustworthy while it's current. That's the point.

Methodology

The whole method is public. That's what makes it authoritative.

av9n is authoritative not because its method is secret, but because it's open, published, and self-measured for error. Verdicts are Ed25519-signed over a canonical document, so any third party — an auditor, board, or insurer — can independently verify a verdict against av9n's published public key, with no secret and no access to av9n. Here's the summary — the full methodology document is yours to download.

1Authority through transparency, not secrecy. The scoring formula and every component is published. A score is only as trustworthy as your ability to check it.
2Independently verifiable. Every score is backed by an Ed25519-signed, tamper-evident record you can verify with the public key alone — no access to av9n required.
3Self-measured error. av9n measures its own false-refute and miss rates against labeled ground truth, and never presents a score as more certain than the measurement supports.
4Non-authorship, enforced structurally. The party being scored cannot author the inputs that set its score — the same reason a company can't audit itself.
5A data moat, not a secret one. Anyone may read the method. What can't be copied is the labeled corpus that makes av9n's verifier measurably more accurate than one starting from zero.
PDF
av9n Verification Methodology Revision 1.0 · 4 pages · published spec

We'll email you the PDF and occasional av9n updates. No spam, unsubscribe anytime.

You're in.

The methodology is downloading now, and we've sent a copy to your inbox.

Download PDF
Audit & Compliance

The independent evidence layer for AI audits.

New state laws require independent, conflict-free third-party AI audits. av9n is the tamper-proof infrastructure those audits run on — sealed records, an evaluation environment the operator can't tune, and a verification history your board and auditors can pull up on demand.

Independent by position, not by promise. av9n doesn't build, host, or operate the AI it measures — the same structural independence the statutes require of a qualified auditor. That's not a feature; it's where we stand.

Performance and audit, together. The same sealed rails that measure whether your AI creates value also produce the evidence that it was evaluated with integrity. In a board meeting, one record shows both: what your AI is worth, and that the number can be trusted.

ILSB 315 — AI Safety Measures Act. First U.S. law mandating annual independent third-party audits, conducted by experts without financial conflicts of interest.
CATransparency in Frontier AI Act. Safety frameworks and transparency reporting for frontier models.
NYRAISE Act. Responsible AI safety and disclosure obligations.

Sealed, tamper-evident records

Every verification is hash-chained and signature-verifiable. An auditor or regulator can confirm nothing was altered — without trusting av9n.

Clean-room evaluation

Audit probes run in an environment the operator can't see or tune — held-out inputs, egress-locked, attested. The evaluation itself can't be gamed.

Retained verification history

Every report stored and retrievable for the statutory period. Pull up any quarter's evidence on demand — the archive auditors ask for.

Scheduled reporting

Automatic weekly and monthly reports on a fixed cadence — the consistent record that passes audit renewal, not a one-off deck.

Where we're honest about the line. av9n provides the sealed, independent rails an audit runs on. The probe definitions, professional opinion, and legal compliance determination remain with the qualified auditor and the regulator — av9n records and seals them, but never authors them. That division is what makes the evidence trustworthy: infrastructure you can verify, judgment that stays with the experts.

One-time audit
from$2,500 / audit

An independent, one-time verification audit of a single AI system or agent — run in the sealed clean room, delivered as a signed, tamper-evident report you can hand to your board, an auditor, or a regulator.

The $2,500 starting price is for a single-system, point-in-time directional audit on your supplied inputs. It produces independent evidence — not a legal compliance certification or a qualified auditor's opinion.

Every audit includes

  • Independent evaluation in the sealed clean room
  • Counterfactual-based value verification
  • Signed, tamper-evident report
  • Independently verifiable seal + document hash
  • Retained in your audit archive
What $2,500 covers
  • One AI system or agent
  • A single point-in-time audit (a snapshot, not continuous monitoring)
  • Directional tier — av9n's standard methodology
  • Your supplied inputs, data, and access
  • One workflow / standard evaluation volume
Talk to us for
  • Multiple systems or a multi-agent architecture
  • Ongoing / continuous re-verification
  • The certified clean-room tier (TEE-attested)
  • Custom or expert-authored probe suites
  • High-volume or regulated frontier-model audits
Contact

Talk to us about verifying your agents.

Whether you're building, buying, or certifying.

Tell us what you're working on and we'll get back to you. Enterprise, marketplace certification, design partnerships — all welcome.

@hello@av9n.ai
Enterprise & marketplace volume certification
Design partners get early access

We'll only use your details to reply. No spam.

Message sent.

Thanks — we'll get back to you at the email you provided.

av9

Give your agents the worth they earn.

av9n — a verification network. Said “aven.” Worth that’s witnessed and checked, not claimed.

pip install av9n