MANIFOLD BINARY MANA RESOLUTION AUDIT — OCTOBER 4, 2026 UTC Original research by Prediction Contracts. This is a bounded noncommercial study of play-money market labels, not forecasting accuracy or cash contracts. ARTICLE https://predictioncontracts.com/blog/manifold-binary-resolution-audit COLLECTION One unauthenticated GET on the existing Hetzner research server, using Python 3.14.4 on Linux. Request started 2026-10-04T03:36:12.100Z and completed 2026-10-04T03:36:13.722Z. One attempt, HTTP 200; response Date header Sun, 04 Oct 2026 03:36:13 GMT; CF-Cache-Status BYPASS, Cache-Control private, no-store. Exact request: https://api.manifold.markets/v0/search-markets?term=&filter=resolved&sort=resolve-date&limit=500&contractType=BINARY The response contained 500 unique, resolved BINARY records explicitly labeled token=MANA, ordered by descending resolutionTime. All were eligible; none had missing creator IDs, invalid lifetime volume, or unknown resolution labels. There were no ordering violations. Recorded resolutions span 2026-09-19T13:21:54.153Z through 2026-10-04T02:47:55.632Z (14.56 days). One record occupies the oldest timestamp. This verifies the returned order, not completeness of the platform's visible, deleted, private or revised records. The protocol was written before this collection, following a two-record feasibility probe that showed two cancellations by one creator. It was not a preregistered experiment. No extra pages or market details were fetched to enlarge this sample after seeing its results. ELIGIBILITY The frozen analyzer rejects, in order: non-object records; missing market ID; duplicate ID after its first occurrence; non-BINARY; token not explicitly MANA; isResolved not literal true; missing resolution label; invalid resolution time; or resolution more than 60 seconds after local collection completion. Unknown nonempty labels would remain as OTHER. No such exclusions or unknown labels occurred in this response. Missing numeric values are not imputed as zero. RESULTS YES: 189 (37.8%); NO: 278 (55.6%); CANCEL: 25 (5.0%); MKT: 8 (1.6%). Outside YES/NO: 33 of 500 (6.6%). YES/NO retained: 467 (93.4%). Creators represented: 238 before the label filter, 227 after it (11 lost). All eight MKT resolutionProbability values are strictly between 0 and 1, ranging from 0.26 to 0.72; no MKT endpoint values appeared. The largest creator overall supplied 35 records, all YES/NO. Removing those leaves 33 of 465 outside YES/NO (7.0968%). The creator with the most records outside YES/NO supplied 11 cancellations and no other sampled records. Removing those leaves 22 of 489 (4.4990%). Creator ties use lexical original creator ID. Public grouping codes preserve that order. Giving each of the 238 creators equal weight averages their individual outside-YES/NO fractions to 5.6246%. These are descriptive sensitivity checks, not confidence intervals. The 33 records outside YES/NO account for 97,121.15349424044 of the sample's 2,431,996.54555009 reported lifetime Mana volume (3.99347%). Volume is virtual activity reported by the provider, not dollars, payout amounts, losses or refunds. All 500 volume fields were finite and nonnegative. Age is (resolutionTime - createdTime) / 86,400,000 milliseconds per day. Median age: YES 30.0638 days (189 records), NO 33.2059 days (278), CANCEL 312.0413 days (25), MKT 4.8505 days (8). All creation timestamps were valid. These are ages at the currently recorded resolution, not overdue intervals or settlement delays. No historical scheduled deadlines were collected. VERIFY OFFLINE Save evidence.json, summary.json and verify.py from this same directory into one local folder. With Python 3.10 or newer, run: python3 verify.py The verifier uses only the standard library, makes no network requests, checks the actual sample's eligibility and independently recounts every value in summary.json. It does not import the original research analyzer. The raw archive and original analyzer also have a separate independent verification script in the repository research directory. SHA256SUMS lists hashes of the four public files. In the downloaded directory, run `shasum -a 256 -c SHA256SUMS` on macOS or `sha256sum -c SHA256SUMS` on Linux. Hashes check file consistency, not authenticity of the provider's claims. PUBLIC FILES (relative to this README) evidence.json — all 500 records, source links, grouping codes and required fields summary.json — statistics recomputed from that extract verify.py — independent standard-library recount SHA256SUMS — public file consistency checks The extract omits questions, descriptions and avatars and replaces creator IDs with stable grouping codes. Source URLs remain public, so this is not a promise of anonymity. It preserves the facts required to reproduce all article numbers. The original API response and receipt remain in the repository research archive. SHA-256 of the original uncompressed response (517,521 bytes): 06d4bd479adcccda1afa1a12a39d71d6c9ac00f033d509bb5141827348efd36d LIMITS This is a sample of recent visible resolution records, not all Manifold history or a cohort of newly created questions. It cannot estimate a new market's chance of cancellation, future platform rates, or comparative platform quality. Questions can be correlated across and within creators. One creator supplied 11 of the 25 cancellations. Removing a creator is not an independence adjustment. Cancellation and probability-based resolution have different meanings; neither label establishes an error, dispute, loss or inaccurate prediction. The data do not explain why any market received its recorded resolution. No pre-event fixed-horizon forecasts or real-world outcomes were independently verified. Excluding records changes sample coverage; this does not establish the direction or magnitude of predictive bias. The API probability field and the platform's internal stored probability follow different code paths; neither field, merely observed after resolution, establishes an advance forecast. Fresh API requests produce new samples and may reflect re-resolutions. Use the saved evidence to reproduce these results. The work is an original dated measurement, not a claim to have discovered non-YES/NO outcomes or to be the first research on selection effects. No inference to cash markets is supported. PRIMARY SOURCES (checked October 4, 2026 UTC) https://docs.manifold.markets/api https://docs.manifold.markets/terms https://manifold.markets/calibration https://github.com/Expertium/manifold-calibration/blob/25aa16cfc9699f7ae03a3ce9b1b84a91e16d50ec/scrape.py#L116-L131 Official Manifold source revision inspected: 63ee7115e3570ac955b98173f4dcf6e842ce4cf1 Search ordering: https://github.com/manifoldmarkets/manifold/blob/63ee7115e3570ac955b98173f4dcf6e842ce4cf1/backend/shared/src/supabase/search-contracts.ts#L789 Resolution handling: https://github.com/manifoldmarkets/manifold/blob/63ee7115e3570ac955b98173f4dcf6e842ce4cf1/backend/shared/src/resolve-market-helpers.ts#L96 MKT payout: https://github.com/manifoldmarkets/manifold/blob/63ee7115e3570ac955b98173f4dcf6e842ce4cf1/common/src/payouts-fixed.ts#L189 The inspected public revision is implementation evidence, not proof of the deployed revision or of how every historical record was produced.