How the findings are produced — and how far to trust them

Every result is labelled by evidence class. This page defines those classes and explains, module by module, exactly what we read, compute, and infer.

Evidence classes

You should trust a finding in proportion to its class:

• Deterministic — read directly from the market's own rule text, resolution source, metadata, or order book. Exact and reproducible (e.g. an exact rule fragment, a best bid/ask).
• Derived — computed from deterministic inputs (executable price at a size, extracted deadlines, market relationships, slippage from the visible book).
• Heuristic / modelled — involve interpretation or classification (the resolution-risk score, anomaly labels, Smart Flow inference). Estimates, not facts.

In the product these classes appear as five evidence chips, which map onto the three classes above: Observed fact and Deterministic finding are deterministic; Historical association and Modelled are heuristic/modelled; Unknown means the evidence was insufficient to classify — treated as no finding, never as a low-risk one.

Sources & precedence

Market structure, rules text and metadata come from Polymarket's public APIs; order books from the public CLOB; resolution outcomes reference the market's own stated source and, where relevant, UMA's optimistic oracle; on-chain reads use public RPC. When sources disagree, the market's own published rule text governs every rule-level finding, and the order book governs every price-level finding — we never substitute a secondary source for either. Canonical identity resolves URL, slug, market id and condition id to one tracked market before any analysis runs.

Deadline interpretation

Deadlines are extracted from the rule text verbatim (a deterministic finding). When the text states a clock time without a timezone, or a date whose calendar day depends on timezone, we flag it as timezone-ambiguous rather than silently assuming one — the flag is the finding.

Resolution Watch

Reads the market's full published rules and checks the ways the text can void or dispute a position even when you're right about the world:

• Rule-title mismatch — the title implies a condition the rules never state (a deterministic finding, with the exact fragment).
• Ambiguous resolution condition — subjective or underdetermined language.
• Authoritative-source dependency & source substitution — resolution hinges on a named source, or permits "or another credible source".
• Deadline ambiguity — timezone or calendar-vs-fiscal edge cases in the deadline.

Each flag quotes the exact rule fragment (deterministic). The composite dispute-risk percentage is the model — see below.

The resolution-risk model

The composite dispute-risk score is a logistic model fitted on the rule features of historically resolved markets and graded out-of-sample on a held-out test split. It currently reaches held-out Brier 0.119 vs 0.149 for a constant base-rate baseline — and beats every naive baseline the engine grades it against (constant, category, segment, rules-only at 0.164) — which is why it is labelled MODEL_BEATS_BASELINES (calibrated). It is still a probability over an uncertain future, not a guarantee. If a future dataset revoked that verdict we would relabel it MODEL_NOT_CALIBRATED and say so on this page. Historical pattern statistics are shown with sample sizes and confidence intervals; patterns people assume are risky can measure neutral-to-protective (explicit cancellation clause 0.2×, generic "ambiguous language" 0.5×) — we publish those inverse results too.

Market Truth

One market on one screen: canonical identity, the rules and resolution source, the live order book, the anomalies it participates in, and a deterministic tradeability read. Identity and rules are deterministic; book depth is a live read at the shown timestamp.

Event Map

Surfaces validated logical relationships between markets — negative-risk membership (prices should sum to ~1), duplicates, deadline/threshold ordering, and direct contradictions. Every edge is a deterministic finding derived from the markets' own rules and structure, not a correlation.

Executable Basket & execution pricing

Prices apparent arbitrage baskets at real depth: executable buy/sell price from the order book, slippage modelled by walking visible levels, partial-fill exposure, and the maximum notional that actually clears. Executable prices and slippage are derived findings from the visible book — they are not a promise of a real fill.

Smart Flow

Pseudonymous wallet-cohort intelligence: it groups on-chain participants into behavioural archetypes and reports out-of-sample follow studies. Cohorts, never named persons; a heuristic context signal, never a trade recommendation. What it does not prove: causation, that any cohort will repeat, or that following it is profitable.

Freshness, caching & timestamps

Results can be live, cached, or historical. Each carries an evidence timestamp, and on-chain reads a block context where used. See Coverage for the current index freshness. We never fabricate an analysis when upstream data is unavailable — we return "insufficient evidence".

What this is not

Scanverity Intelligence is research, not advice. It does not execute trades, hold funds, or predict outcomes with certainty. See the Risk & Methodology Disclaimer.