Apply: sponsorships.prezenti.xyz
We back builders on evidence of what they have actually shipped — not on where they studied, who they know, or how many stars a repository has.
This repository is the whole thing: the programme's terms, the rubric your application is scored against, and the running code that does the scoring. It is public so that anyone we assess can check our work, and so that anyone who wants to run a programme like this can take it.
It is a fork of P-U-C/talent-engine, which is the general engine. This fork is Prezenti's live deployment of it.
A trial: 5 places, 4 months, $7,000 total. Applications close when the fifth place is filled, or 2026-12-29, whichever comes first. Each place is $1,400 of tooling that removes the main constraint on someone building alone.
| Claude Max 20x | $200/month × 4 |
| ChatGPT Pro 5x | $100/month × 4 |
| Flexible allowance | $200 one-off |
We are looking for people building agent infrastructure, protocol engineering and on-chain tooling — deliberately including people who have never touched Celo. Scoring rewards what you have shipped anywhere; a credible Celo plan is required at selection, and real Celo work at the month-two checkpoint. Scoring on Celo-nativeness would only find people already inside the ecosystem we are trying to bring people into.
A good-faith pledge of 2% to Prezenti, and to nobody else, covering revenue actually received by the product through Celo, and any grant, prize and retro-funding income won with the sponsored work. Prezenti separately routes half of what it receives onward to the Celo Community Fund, equivalent to 1% of covered income. Bounded in every direction:
- Capped at $14,000 — ten times the sponsorship. It cannot exceed that.
- Expires 36 months after the programme ends.
- Pro-rated by the months you actually take. Leave at the month-two checkpoint and you owe proportionally less.
- No equity. We ask for a right of first offer on a future round — the right to be offered participation, with 14 days' notice. Never an obligation on you.
Continuing support also requires monthly receipts, one monthly public update, and qualifying activity. Thirty days without qualifying activity triggers review; you then have seven days to provide evidence or cure it. Months three and four require a verifiable Celo deployment, integration or material ecosystem contribution by the month-two checkpoint.
Enforcement is reputational. The pledge is recorded publicly as an on-chain attestation that names your GitHub handle, and we are straight about what that means: it is a promise, not a security interest, and nothing here gives us a claim on your company. Acceptance is blocked until counsel clears the exact current terms digest.
Stated because terms that run one way are not a relationship, and because at this cheque size the relationship is the entire return:
- You keep all IP and all equity. We take none.
- No exclusivity. Take other funding, other grants, other work.
- Withdraw at any time without penalty, with the give-back scaling down.
- A public record that we backed you through the mandatory pledge. Any dossier, announcement or evidence package beyond that requires separate consent.
- Introductions where we can make them.
- Your score and every piece of evidence behind it, shared with you.
- Feedback if you are not selected. Everyone, not just the shortlist.
These are enforced in code, not just written here: ProgramOverlay refuses to
load a policy that takes equity, leaves the give-back uncapped, lets it run
forever, or fails to state what the programme owes. See
talent_engine/programs/policy.py.
Four signals carry most of the weight. The complete model is in RUBRIC.md, published so you can reproduce your own number.
- Origination — you create things that did not exist. Forking is free; originating is not.
- Finishing — releases, documentation, a licence, a deployed URL. The last ten percent, the part with no dopamine in it.
- Cadence — distinct weeks of activity, which is far harder to inflate than a commit count.
- Acceptance — other people merged your work, across several projects.
Stars and follower counts are not scored at any weight. They measure access, and access is what this is built to look past. The configuration refuses to load if insider-network signals are weighted above measured shipping.
git clone https://github.com/prezenti/talent-engine
cd talent-engine
python3 -m compileall -q talent_engine tests # dependency-free syntax check
python3 -m pip install pytest # development dependency for full suite
python3 -m pytest tests/ -q # full local suite, no network needed
export GITHUB_TOKEN=...
python3 -m talent_engine.cli score \
--program prezenti-sponsorship-trial --handles YOUR_HANDLEEvery component is evidence-linked — a score with no evidence behind it cannot be constructed, structurally. If a number looks wrong, open an issue and point at the line. That is what publishing it is for.
This matters more than the parts that flatter us.
A score never decides anything. It produces a shortlist; people read the evidence and decide. The code enforces this — a policy that permits automated final selection will not load.
The scoring can be gamed by a patient adversary, and we can prove it. docs/STRATEGY.md contains measured results showing a manufactured profile outranking a genuine builder, kept as failing-if-fixed regression tests in tests/test_redteam.py. We publish this because a rubric everyone can read is a rubric someone will farm, and pretending otherwise would be the dishonest part.
Our defence against that is not clever, it is corroboration. Cross-applicant ring detection requires several independent signals to agree before flagging anything, and it never flags pairs — two people who validate only each other is the normal shape of a founding team. Two external model reviews broke the first version in both directions; both attacks are now permanent tests.
Every score carries a caveat sentence saying what to be sceptical about, and it is never reassuring. When nothing specific stands out it says what the score does not establish, because silence would be a claim we cannot support.
you fill in the form ─┐
├─→ Tally signs a webhook
frontier-lab and │ │
outreach traffic ─────┘ ↓
clawd verifies the signature
↓
collect public GitHub activity
↓
score against the rubric ──→ caveat sentence
↓ │
cross-applicant ring check ────────┤
↓ ↓
emailed to the programme operator, with evidence
↓
humans decide ──→ accept: terms, Safe payment route, public pledge
└─→ decline: feedback, tracked
Alongside it, a daily sourcing run finds people who never applied, scores the best of them, and emails a ranked few.
| Piece | Where |
|---|---|
| Public page and signed intake | talent_engine/server/ |
| Form parsing, PII quarantine | talent_engine/ingest/tally.py |
| Rubric, flags, caveat sentence | talent_engine/scoring/ |
| Ring detection | talent_engine/modes/rings.py |
| Acceptance, decline, feedback | talent_engine/modes/ |
| Programme terms, validated | policies/, talent_engine/programs/ |
| Form questions (source of truth) | forms/sponsorship-application.json |
Some properties worth knowing, each of which replaces a failure that is hard to see after it has happened:
- Signature verification is mandatory. There is no unauthenticated mode.
- Scoring happens off the request path. It takes tens of seconds; form platforms retry anything that is not a fast 2xx.
- Redelivery is a no-op, keyed on the form's own submission id.
- An incomplete collection is never scored. A partial snapshot is a floor, not a measurement, and a low number with a soft warning beside it is worse than no number.
- Contact details are quarantined in their own table, stripped from the application by label, field type and value shape, and never present in an assessment record or an evidence dossier.
- Question labels are the wire format. They live in
forms/*.jsonand a test round-trips them through the parser, so renaming a question in the form builder fails the suite instead of silently dropping a field.
Zero dependencies for scoring — Python 3.10+ standard library only, so that reproducing your score requires no install step.
The engine is program-agnostic. A programme is two files: a scoring config in
programs/ (weights, ecosystem taxonomy, window) and a policy in policies/
(seats, budget, terms, what you owe recipients). Both are validated on load.
python3 -m talent_engine.cli serve --program <your-program>
python3 -m talent_engine.cli rings --program <your-program> # before selection
python3 -m talent_engine.cli feedback-queue --program <your-program>Fixes and improvements to the engine belong upstream at P-U-C/talent-engine. Prezenti-specific configuration lives here.
| RUBRIC.md | the complete scoring model |
| docs/STRATEGY.md | why these signals, and what an adversary can do to them |
| docs/SPONSORSHIP_TERMS.md | current terms index |
| docs/terms/prezenti-sponsorship-trial-2026-08-14-v3.md | canonical terms release agreed by applicants |
| docs/PREZENTI_SPONSORSHIP_TRIAL.md | the trial in full |
| docs/PROGRAM_LAYER.md | how programmes sit above the score |
| docs/ENGINE.md | the engine's own README |