IFFA Signals · research

Methodology

Evidence before narrative. This page summarises the rules; the full versioned documents live in the repository.

Collection

ARGUS reads only official, documented, unauthenticated public endpoints: Polymarket's Gamma API (discovery, market metadata, rules) and Data API v2 (resolution state, price history). No HTML is scraped, nothing is traded, and no location restriction is bypassed.

Every response — including every failure — is stored with its SHA-256 before it is parsed. All datasets are rebuilt offline from those stored bytes, so any number can be traced back to a specific response.

A host that answers with a bot challenge is recorded as blocked and never retried or worked around (kalshi.com, 2026-09-24).

Settlement records are read from Polygon through public, keyless JSON-RPC endpoints listed in Polygon's own documentation. Each provider's stated limits are honoured as stated, including a 30,000-block ceiling on log searches, even where a request shape happened to slip past the check.

India classification

A market is India-related when its own text (question or outcome label) or its event title names an Indian entity from an explicit, versioned lexicon — places, institutions, parties, people, teams, companies. Ambiguous words (Congress, Stalin, Hyderabad, rupee…) count only alongside an unambiguous India anchor.

Known collisions are excluded before matching and recorded (Indiana, India Walton, Indian Wells, baseball RBI…). Markets reached only through event context or a platform tag go to a manual-review queue.

The classification describes what a market is about. It never implies that anyone who traded it is Indian or in India.

Resolution

An outcome is read from two platform surfaces: Gamma's final prices and Data API v2 /v2/resolutions (which serves two row shapes: UMA price rows and payout rows). Only markets where both readings agree enter calibration; disagreements are kept and flagged, never resolved by guesswork.

A third, independent reading comes from the chain (next section). A resolution card is marked verified only when all three agree.

A Polymarket resolution is an oracle determination under the market's own rules text. It is not an independent finding about the world.

On-chain settlement

Positions are redeemed against the payout record that the Conditional Tokens (CTF) contract writes to Polygon: one ConditionResolution event per market condition. ARGUS takes the contract address from Polymarket's published contract list and computes the event signature itself.

For each resolved India-related market, ARGUS searches for that event with the market's condition id. It searches the exact block when the platform names it; otherwise it estimates the block for the platform-reported resolution time from block headers, then searches bounded windows. The payout numerators map to the market's outcomes in order and are compared with the platform's reading: match, mismatch, not found, query failed or anomaly. These statuses are never merged.

Integrity checks: the condition id must recompute from the event's own oracle, question id and outcome count. On a fixed sample, a second, independent RPC provider must return the identical record and block time.

“Not found” means one provider returned nothing in the blocks searched. It does not mean the market never settled. Settlement time is when payouts were reported on-chain, not when the real-world event happened.

Polymarket's resolution rows carry their own transaction reference. ARGUS does not assume it is the settlement: it records whether it is, and on a sample, which event it actually points at.

Calibration

For each resolved binary market, the price of the first outcome is read at fixed horizons before the earlier of the scheduled end and the actual close. A bucketed price counts only once its bucket has ended; horizons finer than a series' measured spacing are marked unavailable.

Reported: Brier score with a 95% event-cluster bootstrap interval, a base-rate reference, reliability buckets with Wilson intervals, expected calibration error, log loss and the Murphy decomposition. Strata with fewer than 30 markets are reported as insufficient.

The 7-day horizon and the $1,000 volume floor were fixed before results were computed. Each horizon has its own sample, so rows are different populations.

Evidence and corrections

Every published claim maps to an evidence card with a permanent IFFA-EVIDENCE id, a status (verified, partially verified, disputed, unverified, retracted, outdated), a quotation from stored bytes, and a list of what it does not establish.

History is never overwritten: a correction changes a card's status and is logged in the corrections record, with a new dataset version.

Full documents

Current bundle: methodology 0.2.0, classifier v1.1, dataset v2026.09.24, run 20260924T233343Z-08eab6, built from commit cdc4f40917 with uncommitted changes.