How Daybreak works

Data sources

Calibration database

The calibration database answers: historically, when a prediction market in category X priced an event at probability P with T days to resolution, how often did that event actually occur?

Calibration is computed against a three-axis grid:

Total grid: 240 cells (6 × 10 × 4). Each cell requires at least n = 10 resolved markets before it is surfaced — below that it carries a low-confidence flag and is shown with an explicit caveat.

Foundation state: 758 resolved markets ingested, 128 / 240 cells populated. The database fills in over time as markets resolve under the live ingest pipeline. There is no shortcut — depth is the point.

Cell identifiers follow the format “polymarket:politics:0.70-0.80:30d” — source, category, probability band, time bucket. Every calibration annotation in the product traces back to a cell identifier so the claimed count is always verifiable.

Reality Check

Reality Check takes a free-text claim and returns the deepest prediction market that prices it, the market's current implied probability, the 30-day Smart Money vs Crowd trajectory, and — when the calibration cell clears the n = 10 floor — the historical resolution rate for similar markets.

Matching. The claim is embedded with OpenAI's text-embedding-3-small (512 dimensions) and compared via cosine similarity to every active market's embedding. We keep the top 3 matches that clear a 0.55 similarity threshold and additionally have a liquidity score ≥ 0.5. When nothing clears the threshold, the page says so plainly rather than serving a weak match.

Interpretation. The one-sentence interpretation under the hero number is generated by Claude Haiku from a structured snapshot of the matched market. The model is instructed to emit prose only — every number on the page comes from the structured snapshot, not the model. A verifier rejects any output that contains a digit or numeric word and forces a corrective retry; if the verifier fails twice, the page renders the market data without a synthesis line rather than fudging it.

Calibration delta. The historical resolution rate is read from one cell of the calibration database described above — keyed on the market's source, category, implied-probability bucket, and time-to-resolution bucket. The delta displayed (“priced N pts high”) is the difference between the market's current implied probability and the historical realized rate for markets in that same cell. When the cell holds fewer than n = 10 resolved markets, the delta is replaced with an explicit limited-history caveat.

Share-image gate. The branded PNG export is disabled until the calibration database holds at least 200 resolved markets across at least 4 categories. Below that, the page renders normally but the share button is greyed out — we will not put thin numbers on a public share-image footer.

Rate limit. Five queries per IP per UTC day. We hash the IP with a server-side salt before storing the counter and never log the raw address. Anonymous use is the point until accounts arrive.

Anomaly baselines

The Biggest Movers strip ranks markets not by raw point-change, but by how unusual today's move is for that specific market. A 3-point move in a thin geopolitics market may be the first signal of something; a 3-point move in a liquid, high-volume market is often noise. Unusualness is measured as a z-score against each market's own baseline.

The window. Each market's baseline is its 30-dayrolling-vol — the standard deviation of daily log-returns over the prior 30 trading days. A market's z-score on a given day is its log-return divided by that baseline. A score of 2.6σmeans today's move was 2.6 times larger than a typical daily move for that market. The window is 30 days rather than 90 because Polymarket's market-age distribution makes a 90-day floor too exclusionary — it would drop 53% of the universe.

Exclusion floors. Three filters apply before a market appears in any strip:

Excluded markets are counted in the tile header (“N tracked · M excluded”) so the exclusion is transparent rather than silent.

Quiet Movers. The Quiet Movers strip surfaces markets where the z-score is less than 0.5 (“less than half a typical daily move”) but a known scheduled event lands within the next 7 days. The hypothesis: a market that hasn't repriced ahead of a known event is worth watching.

Events are drawn from a curated event_calendar.json— FOMC dates, SCOTUS ruling windows, elections, central bank meetings, and major economic releases, each tagged with the market IDs most likely to be affected. We do not infer event associations algorithmically; only manually tagged markets qualify. The limitation is honesty: if a market isn't in the calendar, it won't appear as a quiet mover, even if it should be.

Cold-start honesty. On the day the widget shipped, event-calendar entries had empty affectedMarketIds. The quiet-mover strip fills in over time as the calendar is manually curated. We track a quiet-mover hit rate — whether quiet markets actually reprice within 7 days of the listed event — as the moat-validation metric. It will not have enough data to be meaningful until 14+ days post-ship.

Anomalies — the six event types

The Anomalies feed answers one question — is anything fishy happening? — by naming what happened, in plain language, rather than collapsing everything into a single mystery score. Each card is one typed event: a specific, named pattern detected on one market on one day. A market can show more than one type on the same day — we keep them as separate stories rather than averaging them into one number.

All six detectors run continuously over every market. There is no single combined “anomaly score”: each type is its own signal, with its own evidence and its own track record.

Categories & fairness. Markets are categorized using the source platform's own tags (politics, macro, geopolitics, crypto, tech, sports, other), not keyword guesses. Each signal is judged within its category — a busy sports market is measured against other sports markets, not against a quiet macro one — so one high-volume category can never crowd out the board.

Choosing what to show

The detectors are deliberately sensitive — on a busy day they flag far more real events than anyone should read. A calm “is anything fishy?” surface cannot open at hundreds of rows. So the feed ranks every event and shows only the most noteworthy few of each type, per category, per day.

How we rank. Two things move an event up the list: how surprising it is relative to what is normal for that type of event in that category, and how much real moneystands behind it. A genuine eight-point move in a liquid market outranks the same move in a near-dead one; a wallet putting thousands on a position outranks one putting tens. Nothing is deleted — thin, low-money events simply rank below better-backed ones. We rank by a market's own volatility and liquidity baselines (which take months of resolved history to mature), so the ranking itself is built on proprietary data, not a generic “most recent” sort.

What this means for you. The feed is a curated short list, not an exhaustive log. Every event we detect is still recorded and still counts toward the track-record figures below; the feed just surfaces the strongest handful so the page stays readable. Markets in their final stretch are routed to the separate Closing now strip and capped the same way.

Liquidity transparency. Because we de-rank rather than delete, a noteworthy move in a genuinely thin market can still appear. When it does, its low liquidity is shown on the card — we never present a number cleanly when the market behind it is thin.

Wallet accuracy

When a wallet address enters a market, Daybreak looks up its lifetime track record: out of all its previous trades, what percentage resolved YES (or NO)? A wallet that has resolved true 72% of the time on similar markets is higher-signal than one at 52%. This lifetime-calibration percentage is shown on the Smart Money Appeared card next to each flagged wallet.

How it's computed. For each wallet, we track every resolved market it has traded in, grouped by category and time-to-resolution bucket. The accuracy is the ratio of YES outcomes to total outcomes. A wallet with fewer than 5 resolved trades is shown with an explicit thin-record caveat — we will not present “5 of 5” as if it were a proven hand.

This metric is one of the three anchors of wallet-level reputation in Daybreak. It decays as new markets resolve — a wallet that was 75% accurate two years ago but has been wrong on the last three trades will trend toward 50%.

How often each signal has been right

For each event type, Daybreak tracks how often a flagged market actually moved in the days after the flag, compared with an ordinary market over the same window. That is the plain-frequency figure on the card — “flagged moves matched the final outcome 8 times in 10, vs 6 in 10 for ordinary moves.” The number is a direct ratio from resolved-outcome records; no language model ever touches it.

Cold-start honesty. An event type with too few resolved flags to quote a stable rate is marked still building its track record rather than shown with a thin, unstable number. The figure firms up as markets resolve. Two types — Resolution Sprint and One-Shot Whale — carry no confidence number at all by design: they are observations of what happened, not predictions.

These figures are computed from historical market resolutions stored in the Daybreak calibration database. They update as markets close, not in real time.

What we never do

Report an inaccuracy

If you spot a number that looks wrong, a market that's misclassified, or a calibration ribbon that seems off — report it. Every report is logged, investigated, and addressed.

Reports logged. Response within 48h.