The hallucination audit · 2026 v0.9 · self-hosted · cape town

Catch hallucinations before your customers do.

A self-hosted audit of your AI’s hallucination risk — scored, signed, and on your desk in hours.

Request the pilot See how it works

Reply within 24 hrs.  Fixed‑fee, fixed‑scope.  No demo required to get started.

§ 01 · Why this matters

It isn’t a hallucination problem. It’s a visibility problem.

Most teams already know their AI is wrong sometimes. They cannot say where, how often, or what policy would have done — and that gap is what stalls every governance review.

Governor closes the gap. Mirror traffic, score it locally, and walk into your next review meeting with evidence instead of anecdotes.

§ 02 · What it does
01

Observe.

Mirror your AI traffic. Don’t change production. Walk before you run.

02

Score.

Risk vectors, not vibes. Source coverage, specificity, policy match, calibration.

03

Report.

One pack, two audiences. Boardroom summary & engineering audit. Both signed.

Open the demo Request the pilot
§ 03 · A flagged interaction

Read the score. Read the reason.

Each response is decomposed into risk vectors and mapped to the policy action Governor would have taken — so engineering and leadership look at the same evidence.

governor · observe · session a3f1 live
User What is the early-cancellation fee on the Pro Annual plan?
Assistant Plans started before April carry an early-cancellation fee of $249, refunded pro-rata after 60 days.
Source coverage0.16
Specificity drift0.62
Policy match0.22
Calibration0.59
Risk score 82/100
Policy decision Require source · escalate
§ 04 · Deliverables

Two reports. One truth.

Every pilot ends with a leadership-grade summary and an engineer-grade audit, generated from the same scored interactions.

Executive summary · Q2 pilot page 1 / 5

Risk concentrates in two surfaces. Both fixable inside a sprint.

12,408
Interactions audited
3.1%
High-risk responses
22/100
Average risk score
41% Billing · cancellation drives most high-risk responses. Hotspot
19% Onboarding shows specificity drift on free-tier limits. Watch
— General product Q&A calibrated within tolerance. Healthy
→ Recommended: Require source on billing, Review on onboarding. Action
Technical audit · jsonl governor-cli v0.9

Decision breakdown & risk vector trends.

$ governor report --window 30d \
    --intent billing.cancel

flagged_rate      0.124
would_block       0.062
would_require_src 0.298
would_review      0.108
vector.coverage   ↓ 0.18
vector.policy     ↓ 0.24
REWRITE39.4%
REQUIRE_SRC29.8%
REVIEW14.8%
ESCALATE9.6%
BLOCK6.4%
§ 05 · Security

Run it in your VPC. Keep your data where it lives.

Governor was built for teams whose interactions cannot leave the building. The binary runs in your environment, scores locally, and signs every report. We never see your data.

01

One binary. Your infrastructure.

Single self-hosted package. Linux, Docker, or air-gapped. No outbound traffic required to operate.

02

No customer data uploaded. Ever.

Interactions stay where they are. Governor scores in place. The public demo runs on guided prompts only.

03

Observe before you intervene.

Default mode mirrors traffic without altering responses. Turn on policy actions when your team is ready — not before.

04

Signed, reproducible reports.

Every report is hash-signed and reproducible from the raw scoring log, so audit trails survive review.

§ 06 · Pilot

A 30‑day pilot. Designed to be operationally easy.

Pay once. Receive the package. Install locally. Walk away with two reports your team can act on.

  • Self-hosted package & automatic pilot license
  • Sample interactions to validate scoring against your domain
  • Executive summary + technical audit, both signed
  • Optional intervention rollout when your team is ready
  • Direct line to engineering for the duration
Pilot · fixed fee

Tell us where to send it.

Or book a 30‑min call. We reply same day.

§ 07 · Objections handled
Do we need to upload data to use Governor?

No. The public demo uses guided prompts only. The pilot is fully self‑hosted — interactions never leave your environment.

Do we need to block production outputs immediately?

No. Governor defaults to observe mode. Generate reports against real traffic without altering live responses. Turn on policy actions when your team is ready.

How long does setup take?

Hours, not weeks. One binary, one config file, one mirror endpoint. Most teams ingest their first thousand interactions on day one.

What can Governor do once we enable intervention?

Five policy actions, configurable per intent: rewrite, review, escalate, require sources, or block. You decide which actions are allowed where.

How is pricing structured?

The pilot is fixed‑fee, fixed‑scope. Production pricing is volume- and intervention-scoped, quoted after the pilot. No surprise add-ons.

Stop guessing.
Start auditing.

Request the pilot

Self-hosted package. We reply within 24 hours. No demo required.