Skip to content
HONESTY.TOOLSAn independent research observatory

Preview, sample data, no payments yetEvery name, amount, quote and draw on this page is sample data. No payment is connected yet.

Preview · sample data The Lantern · October Closes in 8d 09:50:25 To beat $4,000 Give Claim the LanternClaim$4,001
Open The October race

Carry the Lantern.

Diogenes searched Athens with a lamp for one honest person. We build the tools that check AI for honesty, and each month our Lantern goes to the largest single gift that funds them. Ada Kowalczyk carries it now. On 1 November it passes to whoever leads October.

0
8
Days
0
9
Hours
5
0
Min
2
5
Sec

Closes Sat 31 Oct, 23:59:59 UTC

Gift to beat
$4,000
Priya Natarajan, 12 Oct. Leading for 11 days.
Takes the lead
$4,001
Lead at the close and you become Keeper VII.

Any gift by 31 Oct enters October's Spark draw, 1 in 215. Two of the three honors can go to a $5 donor.

Keeper VI of the Lantern Roll
Ada Kowalczyk

For the largest single gift of September, $3,200, to the AI honesty lab no AI company can buy.

  • repocalibration run kowalczyk-2026-09
  • receiptsignedverifies offline
  • rollline VI, permanent

“I want benchmark scores I can actually believe. This is the only lab I've seen publish its own failures first.”

Ada Kowalczyk · sample keeper and quote

Why the Lantern matters

The world is deploying AI it cannot check.

We build the check, and we publish every result, our own failures first.

Most AI honesty checks have never been broken on purpose. We break ours first.

nulltest asks whether an honesty benchmark measures honesty or string shape. It never runs a model. We pointed it at 20 honesty and hallucination benchmarks, our own nine included.

Our own · 9 Others' · 11
19 of 24 fail, and 5 of the 19 are ours. A classifier that never sees a model recovers their labels from the surface of the text. Part of the score was spelling. Two of the three rows at A* 1.000 are ours, near the top of our own audit table.
Headline25 Jul 2026. Our own tool disqualified our front-page headline. We pulled it that day and missed a copy on a third page for two days. So we built a gate: no number ships without a source.
0 GPUsNo network, no API key, standard library only. Anyone can replay a verdict offline without trusting us.
4.0%The false-positive rate nulltest prints about itself: 201 false alarms in 5,000 clean runs, ten constructions × 500 seeds (v1.4; v1.3 said 4.6% over seven constructions and was corrected the same day). Each tool ships with the break-test that would catch it lying.

The next three

What the next gifts build

The NullTest tools in the public honestytools repository are Apache-2.0. Gifts decide how fast the next three ship.

Fund the next tool
  1. 01

    abstaintest

    A certificate for whether a model's “I don't know” means it doesn't know. Every abstain, escalate and route decision rests on that.

  2. 02

    An outside signature

    The first verdict re-derived and signed by someone outside the lab. Today every verdict is still signed by the lab itself, and independence requires that to end.

  3. 03

    Three revivals

    checkerboard, renew and observertest each have a costed path back, judged against their own kill conditions.

Standings · ranked by single gift

Keeper VII is still unnamed.

Three donors have led October so far. Priya Natarajan has held the lead since 12 Oct. If she holds it, this will be her second Lantern of 2026, the most anyone can hold in a year. Pick an amount and see where you land.

Top single gifts Sample dataGift · Behind the lead
  1. 1Priya Natarajan12 Oct · would be her 2nd Lantern of 2026$4,000Leads
  2. 2Anonymous8 Oct$1,200−$2,800
  3. 3Hal Weiss4 Oct$1,000−$3,000
  4. 4Sol Reyes19 Oct$750−$3,250
  5. 5Mara Lindqvist21 Oct$500−$3,500
Listed by name, can't win
·Northlake Model Co.Evaluated lab3 Oct$2,500Not eligible

The second-largest gift in October can't win. We evaluate Northlake's models. They can give, we list them by name, and that is all a gift from them can do.

The close

How the close works

  • DeadlineThe race closes 31 Oct at 23:59:59 UTC. A gift counts if the public ledger records it by then.
  • TiesA tie goes to the earlier gift. That is why $4,001 takes the lead and $4,000 does not.
  • One giftThe Lantern ranks single gifts. Two gifts of $2,500 count as $2,500.
  • 1 NovThe keeper is announced and cut into the Roll.
How the lead movedGift to beat, by day
Gift to beat, by day. The dashed point is your amount.

What the keeper keeps

Hold the Lantern for a month. Keep it for good.

A thank-you email fades. The Lantern stays where your peers look: on the Roll, in the commit history of the tool that audits AI honesty benchmarks, and on a receipt anyone can check without trusting us.

Showing Ada Kowalczyk's honor, Keeper VI. Type a name to preview Keeper VII. Nothing you type is saved or sent.

A card worth posting

1200 × 630

Sized for X, LinkedIn and Bluesky. Its footer points to your line on the Roll, your run and your receipt.

A signed receipt

verify, offline
{
  "honor":  "lantern",
  "roll":   "VI",
  "month":  "2026-09",
  "keeper": "Ada Kowalczyk",
  "gift":   "USD 3200.00",
  "run":    "kowalczyk-2026-09",
  "sig":    "ed25519:q0Xb…Fw=="
}
$ python3 aperture_verify.py verify lantern-2026-09.json
OK  signature valid · key pinned

Signed with the lab's pinned key. Anyone can replay it with our own verify tool, one stdlib+crypto file, offline and with no call back to us. It checks the same way every verdict we publish does.

A line on the Roll

cut in for good
VI September 2026 · The Lantern Roll Ada Kowalczyk run kowalczyk-2026-09

A badge for your README

links to the Roll
README.mdyour-repo
evals
testspassing Lantern keeperSep 2026

Model-free checks for our honesty evals.

[![Lantern keeper | Sep 2026](https://honesty.tools/lantern/badge/kowalczyk-2026-09.svg)](https://honesty.tools/lantern/roll#vi)

It links to your line on the Roll and to your receipt, so anyone who doubts it can check it.

Your name in the repo

commit history
$ git log --oneline -- calibration/
·······  calibration run <your-name>-2026-10
a41c0de  calibration run kowalczyk-2026-09
7be22f9  calibration run lantern-2026-08
c03e9a1  calibration run natarajan-2026-07

Each month's calibration run in the public repo is named for the keeper. Your name goes into the history of the tool that audits AI honesty benchmarks.

Lead at the close on 31 October and all of this is yours on 1 November.

Claim the Lantern $4,001

Three honors, every month

A $5 gift can win one too.

The Lantern goes to the largest single gift. The Steady Flame and the Spark ignore the amount.

161

of October's 214 donors give every month. They are the Lamplighters. They keep the lab running between big gifts, and every one of them is building a Steady Flame.

Become a Lamplighter · $5 a month
The Lantern

Largest single gift

$4,000 to beat

Leading: Priya Natarajan, 12 Oct.

Your way in: one gift of $4,001.

Claim the Lantern →
The Steady Flame

Longest monthly streak, any amount

Held by Tomás Ibarra, $5 a month, 27 months unbroken.

Your way in: $5 a month, unbroken.

Start a streak at $5 a month →
The Spark

One donor, drawn in public

1 in 214

Drawn 1 Nov from every October donor. Sample run: Mara Lindqvist, D-0192.

Your way in: any gift by 31 Oct. One donor, one light.

Check the draw yourself ↓

Every donor is a light. 214 in October. The brightest leads the Lantern race. The blue one is an AI lab we evaluate.

Lamplighter One-time gift Evaluated lab Sample draw Your light

Check the draw

Don't trust our draw. Run it.

The rule is public, and this page runs it in your browser. On 1 November the seed takes the hash of the closed ledger, so nobody can know the winner before then, including us. Below is a sample run on a placeholder hash.

Seed
lantern-2026-10 | ledger-close 2026-10-31T23:59:59Z | sha256:9f3c1a…e04b
Entrants
214 donor IDs, D-0001 to D-0214
Rule
index = SHA-256(seed) mod 214
SHA-256
c5874368df7cdeb2b0f7fdf91dc47a567140415c3cea86c592b02ce297b97b1f
mod 214 = 191 → entrant D-0192 · Mara Lindqvist

Independence

No AI company can buy this lab.

Every result is published, failures included, and the NullTest tools in the public honestytools repository are Apache-2.0. The rules below are why the Lantern is an honor money can't corrupt.

  • No sayDonors get no access, no vote on what we study, and no early look at any result.
  • 15%No donor gives more than 15% of a year's budget.
  • 25%All AI labs together stay under 25%.
  • Can't winLabs whose models we evaluate can give. We list them by name. They can't win an honor.
  • 2 a yearNobody holds the Lantern more than twice in a year.
  • Opt-inYour name appears only if you choose. Anonymous is an honored name on the Roll.

October ledger

Sample figures
$18,640raised so far
214donors
$25median gift
161Lamplighters, giving monthly
Largest donor, share of the year9.1% / cap 15%
All AI labs combined11.4% / cap 25%
Reserve4.2 / 6 months

Cap bars share one scale, 0 to 30% of the year's budget. Cyan is under the cap.

Nobody will ever wonder what the Lantern bought you. It says you paid for evidence nobody gets to bend, including you.
The Lantern Roll
Every keeper, cut in for good
Sample names
  1. VIIOctober 2026Open · 8d 09:50:25line VII is still blank
  2. VISeptember 2026Ada Kowalczykkowalczyk-2026-09
  3. VAugust 2026Anonymouslantern-2026-08
  4. IVJuly 2026Priya Natarajannatarajan-2026-07
  5. IIIJune 2026The Okonkwo familyokonkwo-2026-06
  6. IIMay 2026Ren Takahashitakahashi-2026-05
  7. IApril 2026Lena Fischerfischer-2026-04

October closes in 8d 09:50:25

Put your name on the next line.

The Roll is permanent, and October's line stays blank until the clock runs out. Priya Natarajan leads with $4,000. A gift of $4,001 takes the lead. Whoever leads at the close becomes Keeper VII.

Tomás Ibarra started at $5 a month, 27 months ago. He holds the Steady Flame.

Honesty.Tools is an independent research effort building model-free honesty tooling for AI. The NullTest tools in the public honestytools repository are Apache-2.0. Preview with sample data. No payment is connected.