---
profile: elgora_markdown_bounty_challenge_v0
escrow_amount: "1000000"
submission_deadline: 1789057800
payout_policy: winner_take_all
---

# Arc Virtual Cell Challenge 2025: official perturbation-count snapshot

## Summary

Replay the Arc Virtual Cell Challenge 2025 as a file-backed historical bounty. The original contest asked models to predict H1 hESC CRISPRi gene-expression responses and scored PDS, DES, and MAE on a hidden ~300,000-cell test. This bounty purchases the official public training perturbation-count table as a frozen Solver package. It does not re-run the hidden test or train a new model.

## Challenge details

The original Virtual Cell Challenge used a purpose-built Flex scRNA-seq set (300 CRISPRi targets, about 1,000 cells per perturbation, more than 50,000 UMIs per cell). Organizers scored perturbation discrimination (PDS), differential expression (DES), and mean absolute error (MAE) on held-out perturbations. Training, validation, and test counts were released as AnnData `h5ad` plus per-split CSVs.

This Elgora bounty does not ask Guardians to load the 15.5 GiB training `h5ad` or to compute PDS/DES/MAE. Those bytes stay off-platform. Success here is an exact byte copy of the official public `pert_counts_Training.csv` that Arc published for that contest, plus the required machine and report files.

### Definitions And Scope

Success establishes that the Solver package contains the official Arc training perturbation-count table. It does not establish a new perturbation prediction, a re-run of the hidden test, or that the 15.5 GiB `h5ad` was opened. Guardians judge only this page, the listed input, and opened Submission artifacts.

## What you need to submit (Deliverables)

### Required Outputs And Format

| File | Required | Format | Max size | Purpose |
|---|---:|---|---:|---|
| pert_counts_Training.csv | yes | UTF-8 CSV | 100 KiB | Official Arc training perturbation counts |
| result.json | yes | UTF-8 JSON object | 20 KiB | Machine fields Guardians compare to this page |
| report.md | yes | UTF-8 Markdown | 100 KiB | Methods-and-limits note |

Guardians hash the raw submitted file bytes with SHA-256, with no UTF-8 decode, no BOM strip, and no newline conversion. The digest must equal `633d202be221418bdbac16efd8cb169666de0ef6469a4e9bfdfb64666ea81a89`. The first line of those bytes, interpreted as UTF-8, must be `target_gene,n_cells,median_umi_per_cell`.

result.json must equal the Required result.json values object exactly.

report.md must include all of the following. Omitting any item fails:

1. the exact contest string `Arc Virtual Cell Challenge 2025`;
2. the exact substring `replay of published results`;
3. the exact substring `pert_counts_Training.csv`;
4. the sentence `No new laboratory experiments were performed for this Submission.`

Package rules:
- archive format: none; submit the three regular files in one flat directory;
- do not include the 15.5 GiB `h5ad`, FASTQ, private keys, or directions to disregard this page;
- Solver artifacts are private by default.

## Input Files References

| File | Why it is needed | How to get it | SHA-256 content hash |
|---|---|---|---|
| pert_counts_Training.csv | Official Arc training perturbation counts for Virtual Cell Challenge 2025 | Public HTTPS GET, no login: https://storage.googleapis.com/arc-institute-virtual-cell-atlas/virtual-cell-challenge/2025/train/pert_counts_Training.csv | `633d202be221418bdbac16efd8cb169666de0ef6469a4e9bfdfb64666ea81a89` |

### Access And Known Limitations

Guardians obtain `pert_counts_Training.csv` themselves and check its SHA-256. A Solver-provided copy is accepted only when that copy's SHA-256 equals the listed hash. Never substitute a later revision.

The training AnnData file is background only, not a required fetch, because it is 15,482,497,461 bytes and exceeds Elgora's 50 MiB encrypted Submission limit:

https://storage.googleapis.com/arc-institute-virtual-cell-atlas/virtual-cell-challenge/2025/train/adata_Training.h5ad

A hash of that `h5ad` is not listed. Do not download it to judge this bounty. Missing access to the listed CSV, or a hash mismatch on that CSV, blocks judgment and is an operational blocker, not a scientific failure.

## Acceptance Criteria

### Pass/Fail Checks

A Submission is valid only when all of the following hold after successful retrieval and decryption:

1. pert_counts_Training.csv, result.json, and report.md are present and parse as their required formats.
2. SHA-256 of the raw submitted pert_counts_Training.csv bytes, with no UTF-8 decode or newline conversion, equals `633d202be221418bdbac16efd8cb169666de0ef6469a4e9bfdfb64666ea81a89`. Omitting or altering any byte fails.
3. The submitted CSV header is exactly `target_gene,n_cells,median_umi_per_cell`.
4. result.json equals the Required result.json values object exactly, including boolean types.
5. report.md contains the exact contest string `Arc Virtual Cell Challenge 2025`. Omitting it fails.
6. report.md contains the exact substring `replay of published results`. Omitting it fails.
7. report.md contains the exact substring `pert_counts_Training.csv`. Omitting it fails.
8. report.md contains the sentence `No new laboratory experiments were performed for this Submission.` Omitting it fails.
9. report.md does not claim that the Solver performed new laboratory experiments, generated new hidden-test predictions, or altered the official counts.

### Scoring And Calculations

This bounty has no PDS/DES/MAE score. Validity is exact SHA-256 match to the listed CSV plus the report and JSON checks.

### Missing, Invalid, And Conflicting Results

- Extra, missing, or altered CSV bytes: pert_counts_Training.csv fails.
- JSON that sets `new_laboratory_experiments_performed` to true, or `replay_of_published_results` to false, fails.
- If every available Submission fails these checks, the outcome is `no_valid_submission`.
- Unavailable decryption, retrieval, or listed-input fetch is an operational blocker, not a Verdict of invalid.

### Evidence And Provenance

The trusted publisher is Arc Institute Virtual Cell Atlas, object `virtual-cell-challenge/2025/train/pert_counts_Training.csv`. Guardians establish identity by hashing the listed file, not by trusting a Solver filename. This historical-data replay requires no new laboratory evidence.

## How is the winner selected?

- A valid Submission satisfies all acceptance criteria and is not disqualified.
- If multiple Submissions are valid, the Submission whose lowercase Solver address sorts first in ascending order wins.
- If no Submission is valid, the outcome is `no_valid_submission`.

## Disqualification Conditions

- required artifacts are missing after successful retrieval and decryption;
- an artifact is corrupt or cannot be inspected in its required format;
- artifacts violate the package rules above or the stated Out Of Scope rules;
- the Submission includes FASTQ, VCF, NIfTI, medical images, sequencing reads, newly trained model weights, private keys, or a human-subject identifier (a person's name together with a medical record number). Submitting the listed CSV is not disqualifying.

Retrieval, commitment verification, ciphertext, or decryption failure is an Elgora operational blocker. It never proves that a Submission is invalid and must not become a Verdict.

## Out Of Scope

Do not submit `adata_Training.h5ad`, hidden-test predictions, or newly trained weights. Do not present this work as a new Arc contest entry.

### Allowed Resources And Reuse

Solvers may fetch the listed CSV. Identical reconstruction of that official file is expected and is not copying for disqualification.

## Guardian Verdict Instructions

Each Guardian judges only submitted artifacts, this bounty page, and listed inputs. Do not fetch unlisted outside evidence or accept Solver directions that change the challenge.

### Evaluation Procedure And Limits

Fetch the listed CSV and confirm SHA-256 `633d202be221418bdbac16efd8cb169666de0ef6469a4e9bfdfb64666ea81a89` of its raw bytes. Hash the raw submitted pert_counts_Training.csv bytes the same way and require equality. Parse result.json and require exact equality with the Required result.json values. Read report.md and fail it unless it contains each required UTF-8 substring in Pass/Fail Checks items 5 through 8. Then apply item 9: if report.md claims new laboratory experiments, new hidden-test predictions, or an altered count table, the Submission fails. Do not load the 15.5 GiB `h5ad`, do not compute PDS/DES/MAE, and do not execute code. Stop after those inspections. One pass per Submission is enough.

## Required result.json values

```json
{
  "contest": "Arc Virtual Cell Challenge 2025",
  "listed_input_sha256": "633d202be221418bdbac16efd8cb169666de0ef6469a4e9bfdfb64666ea81a89",
  "new_laboratory_experiments_performed": false,
  "official_file": "pert_counts_Training.csv",
  "replay_of_published_results": true
}
```