# Wishing Well Bench

**Version:** `wwbench-corpus/1` · **Content address:** `0x98daa6b258cef9ba8da6084dbb55dde32d6b4605b2d6a3c403183fc288d74947` · **License:** CC-BY-4.0

A dated, capital-weighted, signature-provable record of what AI models can and can
not do — assembled from the wishes people actually funded. Each row is one
capability check (K independent pass@k trials, Wilson-gated, three-consecutive-sweeps
verdict) recorded the moment it happened.

## Provenance & what's proven
Every row is a run-log entry **signed by the oracle** when the sweep ran, and each
verdict was **attested on chain** with the row's keccak as the evidence hash — so a
row's content and timestamp are anchored, not asserted, and cannot be fabricated
without the oracle key. In the *full* run log each entry is also **hash-chained to its
predecessor**, so nothing can be inserted, edited, or reordered undetected (verify with
the in-browser verifier over the complete log). NOTE: this corpus is a per-capability
*slice* of that log, so it proves each row is oracle-signed + on-chain-attested, but by
itself it does **not** prove no rows were dropped — run the whole-log chain verifier for
completeness. The `contentHash`/`corpusHash` commit to exactly the rows + totals shown.

## How to verify a row
1. Recompute `keccak256` of the canonical entry → must equal `hash`.
2. `ecrecover(hash, signature)` → must equal `signer` (the oracle).
3. Look up the on-chain `attestCapability` event whose evidence hash == `hash`.

## Contents
- **1** capabilities · **7** checks (3 pass / 4 fail)
- **5** models across **4** providers: anthropic/claude-opus-4-8, anthropic/claude-sonnet-5, deepseek/deepseek-v4, google/gemini-3-pro, openai/gpt-5
- Window: 2026-07-05T06:46:06.786Z → 2026-07-25T06:46:06.786Z
- Compute spent proving the frontier: **$0.02** (2884 in / 363 out tokens)

## Files
- `corpus.jsonl` — one verifiable row per line
- `corpus.csv` — same, columnar (`import pandas; pandas.read_csv("corpus.csv").to_parquet("corpus.parquet")`)
- `datasheet.md` — this file
- `/api/bench/<capabilityId>` — the full per-capability dataset (transcripts + content hash)

## Row schema
- `capabilityId`
- `capabilitySummary`
- `model`
- `provider`
- `seq`
- `ts`
- `pass`
- `passRate`
- `wilsonLower`
- `cleared`
- `streak`
- `satisfied`
- `inputTokens`
- `outputTokens`
- `costUsd`
- `hash`
- `signature`
- `signer`
- `prevHash`