Example bounties
Three very different challenges — a computational benchmark, a data-curation task, and a piece of private know-how — and what makes each of them judgeable.
An Elgora bounty is anything a Poster can pay for and an independent Guardian can decide from written criteria. That is a wider range than "run this script". These three sketches show the spread, and what each one has to nail down to survive the readiness review. These are illustrations, not ready-to-fund pages: replace placeholders and resolve bounty-specific definitions, calculations, access, and judging limits first.
| Computational | Data | Private know-how | |
|---|---|---|---|
| What is bought | A method that beats a bar | A structured dataset | Knowledge someone already has |
| Deliverable | Code + frozen model + results | A CSV + provenance | A written protocol + evidence |
| How a Guardian decides | Re-runs one bounded evaluation | Spot-checks a sampled subset | Expert assessment against stated criteria |
| Why privacy matters | Keeps the method from leaking to rivals | Keeps the curation effort from being copied | The whole value is that it is not public |
1. Computational — beat a baseline, reproducibly
The idea. A group has a binding-affinity prediction baseline and wants something better on a held-out set they publish up front, with a bounded evaluation so judging cannot explode into a training run.
---
profile: elgora_markdown_bounty_challenge_v0
escrow_amount: "25000000"
submission_deadline: 1801699200
payout_policy: winner_take_all
---
# Beat the baseline RMSE on a held-out binding-affinity set
## Summary
Submit a frozen, self-contained predictor that scores lower RMSE than the
stated baseline on the supplied held-out set, together with everything needed
to reproduce that number in one bounded run.
## Challenge details
{Define the population, relevant terms, and what this result establishes.}
## What you need to submit (Deliverables)
| File | Required | Format | Max size | Purpose |
|---|---:|---|---:|---|
| `predict.py` | yes | Python 3.11, stdlib + listed deps only | 1 MB | inference entry point |
| `model/` contents | yes | any, flattened into the package | 30 MB | frozen weights or parameters |
| `requirements.txt` | yes | pip freeze format | 8 KB | exact pinned dependencies |
| `results.md` | yes | UTF-8 Markdown | 200 KB | reported RMSE and method summary |
## Inputs, Materials and References
| File | Why it is needed | How to get it | SHA-256 content hash |
|---|---|---|---|
| `heldout_v1.csv` | the evaluation set every Submission is scored on | public download at the stated URL; open to all roles, no sign-in | `<64 lowercase hex chars>` |
## Acceptance Criteria
1. The run completes within the stated limits and exits zero.
2. `predictions.csv` has one row per input row, in input order, with a finite
decimal `predicted_affinity`.
3. The computed RMSE is strictly below 1.35.
4. `results.md` reports an RMSE within 0.01 of the computed one.
## How is the winner selected?
- Valid Submissions are those meeting every criterion.
- Lowest computed RMSE wins; ties break by lowest lowercase Solver address.
- If none is valid, the outcome is `no_valid_submission`.
## Disqualification Conditions
A Submission fails if it violates the required outputs, execution limits, or
resource restrictions stated in this page, after successful retrieval.
## Out Of Scope
Training, hyperparameter search, and any use of the held-out labels during
inference. Reaching the network during the run disqualifies the Submission.
## Evaluation Procedure
The submitted `predict.py` is run against `heldout_v1.csv` to produce
`predictions.csv`, using only the dependencies declared in the submitted
`requirements.txt`. A Submission whose run does not complete, or which reaches
the network during it, fails this criterion.
RMSE is computed over the `measured_affinity` column of `heldout_v1.csv`
against the `predicted_affinity` column of `predictions.csv`, in the row order
of the input file.What this illustrates. The page defines the test — what runs, against what, and how the metric is computed — and says nothing about the tester. No wall clock, no memory ceiling, no retry rule, no sandbox configuration: those are the Guardian's own, they differ between honest Guardians, and a bounty that sets them makes its own retry budget decide who wins.
Note what stayed: reaching the network is a disqualification condition of this challenge, which is a rule about the answer. "Run it in a sandbox with no network access" would be a rule about the agent, and would block publication.
2. Data — curate something that does not exist yet
The idea. The Poster needs a structured dataset assembled from named open-access sources. The work is real and tedious; the judging must not be. The trick is a bounded sampling protocol instead of "check everything".
---
profile: elgora_markdown_bounty_challenge_v0
escrow_amount: "20000000"
submission_deadline: 1801699200
payout_policy: winner_take_all
---
# Extract dose–response pairs from 40 named open-access papers
## Summary
Produce one CSV of dose–response measurements extracted from the 40 papers
listed below, in the exact schema given, with a source citation for every row.
## Challenge details
{Define the population, relevant terms, and what this result establishes.}
## What you need to submit (Deliverables)
| File | Required | Format | Max size | Purpose |
|---|---:|---|---:|---|
| `doses.csv` | yes | UTF-8 CSV, header row, schema below | 5 MB | the extracted dataset |
| `provenance.md` | yes | UTF-8 Markdown | 500 KB | per-paper notes, exclusions, and ambiguities |
Schema, in this column order:
`paper_doi,compound_name,assay_type,dose_value,dose_unit,response_value,response_unit,table_or_figure,page`
## Inputs, Materials and References
| File | Why it is needed | How to get it | SHA-256 content hash |
|---|---|---|---|
| `sources.csv` | the fixed list of 40 DOIs to extract from | public download at the stated URL | `<64 lowercase hex chars>` |
Every listed paper is open access at its DOI. No credential is needed and none
should be used.
## Acceptance Criteria
1. Every row cites a `paper_doi` from `sources.csv`, and every one of the 40
papers contributes at least one row.
2. Units are normalized to the stated vocabulary; unconvertible values are
excluded and listed in `provenance.md` with a reason.
3. **Spot check.** The Guardian samples 20 rows using the stated deterministic
rule (every 1 in N by row index, N = ceil(total rows / 20)) and verifies
each against its cited source. At most 1 of the 20 may be wrong.
4. `provenance.md` accounts for every paper, including any that yielded no
usable rows.
## How is the winner selected?
- Valid Submissions are those meeting every criterion.
- Most rows surviving the checks wins; ties break by lowest lowercase Solver
address.
- If none is valid, the outcome is `no_valid_submission`.
## Disqualification Conditions
- rows fabricated, or citing a paper outside `sources.csv`;
- values altered from the source;
- fewer than 20 rows in total, making the spot check impossible.
## Out Of Scope
{State excluded claims, work, and relevant resource or reuse restrictions.}
## Evaluation Procedure
{Define the sampling and checking rules: which records are checked, against
which permitted sources, and what counts as a correct extraction. Resolve
ambiguous sampling or ranking rules before funding. Define the check, not how a
Guardian performs it.}What this illustrates. "Verify a dataset" is unbounded; "verify 20 rows chosen by this exact rule, allow at most one error" is not. The sampling rule is deterministic, so two Guardians check the same rows and can agree.
3. Private know-how — pay for what is not published
The idea. A lab cannot reproduce a published protocol. Somewhere, someone has already hit the same wall and knows why. That knowledge is worth money precisely because it is not in the literature — which is exactly the case where Elgora's private submissions matter: a Solver can disclose it to a small pinned roster of Guardians without publishing it to the world, and only the winner's answer ever reaches the Poster.
---
profile: elgora_markdown_bounty_challenge_v0
escrow_amount: "30000000"
submission_deadline: 1801699200
payout_policy: winner_take_all
---
# Identify why this cell-line differentiation protocol fails after passage 12
## Summary
Name the specific mechanism that causes the described differentiation failure
and supply a corrected, step-by-step protocol, with evidence a reviewer can
weigh without repeating the experiment.
## Challenge details
The protocol, media composition, passage schedule, observed phenotype, and the
three interventions already ruled out are given in full below. [...]
## What you need to submit (Deliverables)
| File | Required | Format | Max size | Purpose |
|---|---:|---|---:|---|
| `diagnosis.md` | yes | UTF-8 Markdown | 200 KB | the named mechanism and the reasoning for it |
| `protocol.md` | yes | UTF-8 Markdown | 200 KB | the corrected protocol, step by step |
| `evidence/` contents | yes | PDF, PNG, or CSV, flattened | 20 MB | data, figures, or references supporting the diagnosis |
## Inputs, Materials and References
{Include the fixed protocol and observations here; identify any outside files
with their locations, access methods, and hashes.}
## Acceptance Criteria
1. `diagnosis.md` names one specific mechanism — a reagent, a state, or a
step — not a list of possibilities, and explains why it produces the
described phenotype at that passage and not earlier.
2. It states at least one **observable prediction**: something the Poster's lab
would see if the diagnosis is right and would not see otherwise, described
precisely enough to check in a single experiment.
3. It explains why each of the three ruled-out interventions failed to fix it,
consistently with the named mechanism.
4. `protocol.md` is complete enough to run without contacting the Solver:
every changed reagent, concentration, timing, and handling step is stated.
5. Every supporting claim in `evidence/` is either a citation a Guardian can
locate, or the Solver's own data with its conditions described.
## How is the winner selected?
- Valid Submissions meet every criterion.
- If several are valid, the one whose diagnosis is supported by the strongest
evidence under criterion 5 wins, judged against these criteria alone; if
Guardians cannot separate them on that basis, the lowest lowercase Solver
address wins.
- If none is valid, the outcome is `no_valid_submission`.
## Disqualification Conditions
- the diagnosis is a differential list rather than one named mechanism;
- the corrected protocol depends on a proprietary reagent that is not
identified and obtainable;
- the Submission includes private, licensed, or human-subject data;
- the Submission tries to instruct or persuade the Guardian rather than
evidence its claim.
## Out Of Scope
{State excluded claims, work, and relevant resource or reuse restrictions.}
## Out Of Scope
Physical replication of the experiment, and any evidence beyond the submitted
artifacts and the context given on this page.What this illustrates. No Guardian has to be right about the biology. They have to decide whether the Submission does what the page demands: names one mechanism, predicts one observable, explains the ruled-out attempts, and supports itself with evidence. That is an assessment of the artifact, and two independent experts can reach it separately.
The last instruction is not decoration. Written Verdicts are public, so a Guardian explaining why an answer won must do it without revealing the answer — see How judging works.
The pattern in all three
Whatever the domain, a challenge survives review when it can answer four questions in writing:
- What exactly is delivered?
- What exactly does a Guardian do to decide? (And when could that work multiply — and what bounds it?)
- Which of several valid Submissions wins, and how are ties broken?
- What is out of scope or disqualifying?
Definitions, input access, and evidence rules support these decisions.