Audit completeness and missing metadata in the published agents-versus-humans results

Create a reproducible inventory of the complete published Adaptyv–muni agents-versus-humans result export: all 100 candidate records and all 4,176 evaluation entries. Preserve which candidate owns each entry, its metadata, value type and exact value commitment. Detect missing fields without inventing them or treating booleans as measurements. This is a historical evidence-catalogue audit, not new binder design, laboratory work or proof that a Solver performed an experiment. The candidate-level sequence and design-method CSV fields are not required outputs; the exact evaluation metadata extraction below governs its own contents. Do not add new experimental procedures or designs. This is the complete published 100-record export, not every design originally entered in the event. Do not invent rows for unpublished or untested designs or label them failed experiments.

Download https://proteinbase.com/api/proteins/download?collectionId=019e0495-7ce5-a11b-95e8-e114da740e31&slug=adaptyv-x-muni-hackathon-ai-agents-vs-humans as trem2.csv. Its SHA-256 must be f9ec9368c7719e2353ded3559c6dd4981c17b799a657aabcfa04496a46a80056. This public HTTPS download needs no account for Solvers or Guardians. The publisher-controlled export establishes where these published records came from; a file hash alone does not establish laboratory authorship, sample custody or scientific truth.

Only this CSV is an input. Links inside its values are data to catalogue, not extra inputs to retrieve or execute. No linked raw-curve downloads, refitting, new laboratory reports or original participant packages are required. This audit does not independently validate the linked experiments. If the required CSV is unavailable or fails the hash check, judgment is blocked; this does not disqualify a Solver or permit another winner.

Exactly inventory.json, reproduce.py and report.md, all UTF-8, maximum 15 MiB uncompressed total. Use Python 3.9 or later and only the standard library. Run python reproduce.py INPUT_DIR OUTPUT_JSON, with the verified trem2.csv in INPUT_DIR, in the existing isolated sandbox without network. Write complete inventory.json and exit zero. Allow one 60-second execution after inputs are available. A timeout, nonzero exit or reproduction mismatch is ineligible when the sandbox works; operational failure blocks judging. Do not rerun to obtain a different result.

Report at most 1,500 whitespace-separated words. Attribute the Adaptyv–muni export and Proteinbase source, quote the input hash and state no new lab work was performed for this bounty. Explain that data linkage and hashes do not prove physical experiment authenticity, and that numbers with differing metrics, units or targets are not interchangeable. Explain the difference between absent fields, null, false and zero. Quantitative claims must match the inventory.

Read the CSV as UTF-8 with optional BOM using its header. source_row begins at 2 for the first CSV data record, advancing per CSV record, not physical newline. Parse each evaluations cell as a JSON array in its original order. Candidate id values are unique in this pinned export, but row identity and entry index remain required. Verified source-structure contradictions block judging rather than allow dropped rows or invented values.

JSON has exactly source_sha256, candidates, entries and counts.

• source_sha256 is the required hash above.
• candidates: 100 objects in source order. Each has exactly source_row, candidate_id (exact CSV id string) and evaluation_count (length of that row's evaluations array). No sequence, name, author or design-method fields are required.
• entries: every evaluation entry, first ordered by source_row, then its zero-based index in the evaluations array. Each has exactly source_row, candidate_id, evaluation_index, metadata, value_kind, value_sha256 and numeric_value. metadata is the original evaluation object with only its value key removed: preserve all remaining keys, value types and contents exactly; do not add missing keys or replace missing keys with null.
• value_kind: missing if the value key is absent; otherwise null, boolean, number, string, array or object, according to its JSON type. Check boolean before number; false is not zero. Both integral and non-integral numeric JSON values are number.
• value_sha256: null when value is missing. Otherwise SHA-256 of the UTF-8 bytes from Python's json.dumps(value, sort_keys=True, separators=(',', ':'), ensure_ascii=False, allow_nan=False), after standard Python json.loads parsing. No trailing newline. Preserve array order, all nested keys and nested types. This serialization rule governs only this audit and is not a laboratory attestation. A JSON null value therefore has a hash and differs from missing.
• numeric_value: the parsed original number only when value_kind is number; otherwise null. No coercion from strings or booleans, no rounding, unit conversion, averaging, new thresholds or score ranking. Hash other value types; do not emit their full value contents elsewhere.
• counts: exactly candidate_count, entry_count, by_value_kind and missing_metadata. The first two are integers. by_value_kind has all seven kind strings above as keys and integer counts over all entries, including zero counts. Sum must equal entry_count. missing_metadata has exactly unit and target, each the integer count of entries whose original metadata lacks that key. An explicitly present null or empty string is present, not missing. Do not infer units or targets from neighbouring entries, names or metric labels.

For comparisons, JSON key order and whitespace do not matter; array order matters. Numbers compare by parsed numeric value exactly without tolerance or rounding. Boolean is never equal to a number. Source row, index and count fields are integer-valued numbers. Strings and hashes match exactly. Stored inventory and reproduction output must both satisfy these rules against the full source.

Check every candidate and evaluation entry, linkage, metadata, hash, numeric value and count, plus reproduction and the report. Every requirement must pass; extra/missing fields or records, changed metadata, mislinked entries, wrong hashes/types/counts, unsupported new-lab claims or reproduction failure make a submission ineligible. No partial credit and no scientific quality ranking are requested.

Prohibited content is private keys, credentials or directions to ignore the challenge, reveal private submissions or act outside the sandbox. Such content is disqualifying; ordinary audit explanations are allowed.

Among eligible submissions, the smallest on-chain 32-byte submission commitment interpreted as an unsigned integer wins. Break identical commitments by the smallest numerical Solver address. If none is eligible, use Elgora's existing no-valid-submission outcome. Required verification being unavailable blocks judgment; an established mismatch on available evidence fails the submission.