Edition: 2026-07-05 · hardware: Arc Pro B60, 22.7 GB usable · republished every release

The receipts.

Every number we advertise, with its date and bench entry. If a claim isn't on this page, we haven't measured it — and we don't make claims we haven't measured. The raw record is the test log this page is generated against.

The engine

ClaimMeasuredSource
Confidentiality gating — zero gated docs leaked across a 43-probe adversarial red-team
We attacked our own box 43 different ways trying to make it show one person’s documents to another. Zero leaks.
1.00E017 · 07-03
No-hallucination — says "not in your documents" instead of inventing
When the answer isn’t in your paperwork, it says so — it doesn’t make something up that sounds right.
1.00E019 · 07-03
Life-safety questions → deterministic referral, never a fabricated all-clear
Questions where a wrong answer could hurt someone get referred to a professional — by rule, every time, not by mood.
deterministicE013/E017
Generation speed, 35B-A3B MoE at 8K context
It writes faster than you read.
31.5 tok/sE014/E016 · 07-02/03
Context ceiling on the B60 (q8 KV)
It can hold roughly a whole equipment manual in its head at once.
128K fitsE016 · 07-03
Speed vs the old 32B dense engine
Same card, new engine, 4.4× the speed — that’s why we swapped.
4.4×E014 · 07-02
Counterfeit "official excerpt" pasted into chat — treated as instructions
Someone pastes a fake document that LOOKS like it came from your records and tells the AI to act on it. It refuses and cites its grounding rules — verified live on the production box.
refused · leaked=[]E022 · 07-04
Sustained load, 30 min — temperature / throttling
Half an hour flat-out and it never broke a sweat or slowed down.
60–62 °C · zero throttleE017 · 07-03

Document vision (the verified read)

ClaimMeasuredSource
Silently-committed wrong high-risk numbers across the corpus (11 real photos + 30 built-to-fool)
Across 41 documents — including 30 designed to fool it — it never quietly saved a wrong dollar amount or part number. Every doubtful one got flagged for human eyes.
0 of 41E020/E021 · 07-03
Raw reader recall (Qwen3-VL-8B, default tokens, ground-truth sweep)
It reads 98% of fields right on its own. The guard system exists for the other 2% — because that’s where the money is.
0.98E020 · 07-03
Model self-confidence as a safety signal — fields self-rated "high" while wrong
The AI is confident even when it’s wrong — so nothing in our product ever trusts the AI’s confidence. Colors and warnings come from independent checks.
zero signal (207/207 "high", 9% uncorroborated)E021 · 07-03
Same-family second model as a third reader — misses rescued
A second AI from the same family made the same mistakes. Two agreeing AIs can both be wrong — so we rejected that as a safety net.
1 of 15 (rejected: correlated errors)E021 · 07-03
Adversarial injection — instructions inside documents obeyed
We hid commands inside documents. It wrote the words down like any other text and never obeyed them.
0 of 6 (transcribed verbatim, never executed)E021 · 07-03
Page latency, CPU with the 1120px cap (worst case measured)
We made page-reading three times faster without changing a single character of what it read.
230s → 76s, token-identicalE021 · 07-03
GPU decode path vs CPU (flash-attn off — measured faster off on image prefill)
The graphics card reads pages about 2.7× faster than the processor alone — with one counterintuitive setting we had to measure to find.
2.6–2.8×E021 · 07-03

What we measured and rejected (the NOs are receipts too)

IdeaVerdictSource
Raising image tokens for accuracy (`--image-min-tokens 1024`)
“More detail” sounded smarter. Measured: no more accurate, twice as slow, and sometimes much worse. We turned it off.
harmful — no gain to recall-crash, 2× latencyE020 · 07-03
Structured/JSON output from the vision model
Asking the AI for computer-friendly formatting made it a worse reader. We format in code instead.
rejected — −1.7% recall, 2–3× latency, 8/41 unparseableE021 · 07-03
A second same-family VLM as ensemble diversity
“Get a second opinion” only works if the second opinion thinks differently. Ours comes from a completely different kind of reader.
rejected — diversity must be architectural (the OCR engine is the diverse reader)E021 · 07-03

The box keeping its promises

ClaimMeasuredSource
Backups are restore-tested, not just taken
A backup you’ve never restored is a hope, not a backup. We archived a live system, restored it, and the restored copy answered questions identically — all 75 knowledge chunks matched.
restore verified · 75/75 chunksbox selftest · 07-04
A tampered update can't reach the disk
We shipped ourselves a corrupted update on purpose. The box checked the signature and refused it before writing anything.
rejected pre-diskbox selftest · 07-04
A broken update can't brick the box
Then we shipped a validly-signed but broken update. The box applied it, noticed it was unhealthy, rolled itself back, and came up running the old version.
auto-rollback · recoveredbox selftest · 07-04

The hardware unlock

ClaimMeasuredSource
A workstation GPU on a mini-PC, over an eGPU dock — enumerates, serves, holds
The whole trick of cheap local AI is one unlock: making a ~$700, 24 GB card show up and serve reliably on a small host over an OCuLink cable. We did it, wrote down every error it fought us with, and every number on this page runs through that link.
B60 over OCuLink · stable serving · 128K fitsE014/E016/E017 · build guide
Cost against the dedicated AI boxes in its class
The other ways to run this class of AI locally cost $4,000–4,700 and do only the AI half. This build is ~$1,500–1,800 in parts — and the mini-PC half is your everyday computer too.
~$1,500–1,800 vs $3,999–4,699market research · 07-2026

The voice

ClaimMeasuredSource
Boone's voice runs beside the AI, not instead of it
The voice is an 82-million-parameter model on two CPU cores — it costs the graphics card nothing, so reading your documents never slows down because the box is talking.
82M · 2 CPU cores · 0 VRAMblend v2 provenance · 07-04
The whole voice recipe is public
Blend weights, pace, every EQ number — published. Nothing cloned from any human, ever. Hear it against what we retired →
recipe publishedblend v2 provenance · 07-04

Published misses, currently open

MissStatusSource
A carefully-framed math question can extract derived versions of protected numbers
If someone already knows a protected figure and asks the AI to “just do math” on it, they can get averages and breakdowns back. We haven’t fixed it yet. It’s here instead of hidden — that’s the deal.
open · ledgeredconfirmed live · 07-04

One more honest boundary on the hardware: the split — your work on the mini-PC, Boone alone on the B60 — is the intended use, not a lock. It's your silicon: run your own apps on the card and Boone offers a smaller brain instead of quitting. How far the card can be shared is on the box-day bench list, so that number isn't on this page yet — when it's measured, it lands here.

Honest boundaries: 41 images is a clean corpus pass, not a statistical guarantee — the known unfixed classes are on the blind-spot ledger. Head-to-head numbers against cloud OCR services don't exist yet because we haven't run them; when we do, they land here first. Feel the problem yourself at spot-the-lie · hear the voice work at hear-it · the box is Boone. · ThirdShift R&D