Fair isn't a promise here. It's a procedure — and you can re-run it.

The landing page says the hackathon is checkable. This one shows the machinery underneath. The three rounds are a governance escalation the rest of the stack already uses; the AI only ever advises; the panel is drawn where anyone can reproduce the draw; and every decision is a signed record that verifies on its own. None of it asks for your trust.

← Back to the hackathon

Don't trust us — verify.

The mechanism

The three rounds aren't three ad-hoc stages. They are the Chamber's three tiers.

The NEO stack has a governance overlay — the Chamber — that settles disputes by escalating them through three tiers: an automatic first pass, a human reviewer, and a full panel. Drazen's hackathon design is those same three tiers, run on submissions instead of disputes. So a hackathon is not a bespoke process we invented for the occasion — it exercises the exact escalation the stack uses when real money is in dispute.

ROUND 01

Tier T1

AI triage

Every entry read with the same attention and ranked, with reasons recorded. The shortlist is a recommendation — a human signs the cut before it is real.

ROUND 02

Tier T2

Mentored review

Assigned mentors and judges work with the teams that got through, alongside the AI, not replaced by it. Who advised whom, and when, is recorded rather than remembered.

ROUND 03

Tier T3

The human panel

A drawn panel of people picks first, second and third. The AI advises; the signature is human, and it is the only thing that counts.

Safety outranks the round. An entry that trips an integrity flag jumps straight to the panel — even in round one. AI-triaged is not finally adjudicated.

The AI advises · the human signs

AI Triage, never AI judge.

An AI reads every submission the same way — the one thing a human judge cannot honestly promise on the two-hundredth pitch of the week. But it decides nothing, and not because we distrust it: a decision needs someone who can be asked to answer for it, and software cannot be asked.

The AI's output is labelled and fingerprintedEach triage note carries a content digest bound to exactly what it judged — not a loose comment, a sealed one.

The human decision references that fingerprintEvery cut points back to the advice it consulted (ADVISORY_REF). The reasoning sits beside the signature, where you can read it.

Tampering breaks the binding, detectablyChange the AI's text after the fact and the fingerprint no longer matches. This is the AI Act Article 22 path — built into the Chamber, not layered on by the hackathon.

Assigned to mentor · drawn to judge

Two different jobs, deliberately two different mechanisms.

Mentors are assignedMatching expertise to a team is a judgement worth making. It is recorded — who worked with which team, in which round — because across hundreds of teams nobody holds that in their head.

The final panel is drawnBy sortition, from the eligible judges. Nobody picks who judges you; the operator cannot tune a panel that is drawn from anchored inputs.

The draw is reproducibleTake the published inputs, re-run the draw yourself, and you get the identical panel. "Nobody chose your judges" is a claim you can check, not one you have to accept.

Every decision is a receipt

A cut is a new signed record — never an edit of an old one.

Co-signed, not decreedThe panel's placing needs a majority of the drawn panel to sign it — k of n, not one organiser's word.

Anchored in timeAnchoring fixes the record's position in the sequence: nothing was backdated or quietly reordered afterwards.

Replayable offlineAnyone can take the record away and re-verify the signatures and the anchor independently. Verify, don't believe.

A change is visible as a changeA later decision is a new record that points at the old one. Nothing is overwritten; a rescore would show up as a rescore.

We don't just describe it — it runs

The whole procedure runs end to end as one program, and every record it produces verifies independently.

This is not a diagram of an intention. The three rounds, the AI binding, the drawn panel and the signed receipts run as a single command against the live tier engine. On synthetic entries — no real venture is shown — the run proves the wiring holds:

Each round routes onto its tier — round 1 → T1, round 2 → T2, round 3 → T3.

A flagged entry escalates to the panel — in any round, safety wins.

Every AI advisory is bound to the human decision — and a tampered advisory is rejected.

The panel is drawn, and the draw reproduces — the same names, every time, from the same inputs.

The drawn panel k-of-n signs the 1/2/3 — and every receipt replay-verifies offline.

$ python -m neo_hackathon.spike
SPIKE OK: R1→T1  R2→T2  R3→T3 — AI advisory bound (tamper caught), panel drawn + k-of-n signed, all replay-verified

A wiring proof on synthetic data — the machinery, exercised, not a record of an event. The next edition's dates, rubric and prizes are set separately.

What this does not claim

The boundaries are part of the design.

We cannot prove the judges were rightOnly that the process was not tampered with. Whether a venture deserved to win is a human judgement, and it stays one.

The AI never decides anythingIt ranks and it explains. Every cut and every placing carries a human signature, and there is no setting that changes this.

Verified is not the same as goodVerification proves a record is authentic and unaltered. It says nothing about the quality of what the record describes.

Being drawn is not being competentA reproducible draw proves nobody hand-picked your judges. It does not prove they were the right ones.

Lose, and still know it was fair.

Dates, team size, criteria and prizes for the next edition are being finalised. Tell us you are interested and we will send them the moment they are set.