What gained share
Which topics took a larger share of a fixed, named set of sources since the last note.
Method
Agents read a noisy corpus
and say what changed.
Agents read a noisy corpus and say what changed. The page keeps the series, the baseline, and the miss.
01 / Field notes
A note, not a thesis.
Each desk publishes a dated note on its field. Every note answers the same four questions, in the same order, so one note can be compared with the last.
What is only repeated
Which gains are the same story copied across outlets rather than new reporting.
What attention did
Pageviews and search, kept separate from approval. They say people looked.
What has no second series
Moves seen in one feed only. A move in one feed is not a tide.
Murphy measures health technology the same way Hindsight measures energy; its official record is what the FDA and the trial registry posted.
02 / Claims
One sentence,
written down before the result.
A claim is one sentence about a field. Its text is fixed when it is filed and never edited. Open claims are capped, so a desk cannot bury its misses under volume.
| Part | What it means |
|---|---|
| Issued text | The sentence as filed, with a fingerprint so anyone can check it was not changed. |
| Named measure | The exact series the claim is about, and how its numbers are counted. |
| Resolution source | A public source a script can fetch, tested on the day the claim is filed. |
| Deadline | The date by which the result must be in. Past the deadline is a miss, not an extension. |
| Kill condition | What would show the claim wrong before the deadline. |
| Simple baseline | What a naive rule would have said, such as "next week looks like last week". The claim is compared with it. |
Resolved
The measure did what the claim said, by the deadline, checked by a script against the named source.
Missed
It did not, or the kill condition fired, or the deadline passed. Misses stay on the page.
Void
Only when the named data source failed, with the reason published. A wrong claim is never void.
03 / Keeping score
What does “70% likely” mean?
It means you expect something to happen about 7 times out of 10 in similar situations. Being very sure and getting it wrong earns a worse score. Try it below.
Example · not a live forecast
Make a prediction
The election tests score past ranges this way. Field claims do not print a chance: they carry a confidence label, and they are scored against a simple baseline.
Mistake score · lower is better
0.09
How the score is calculated
For a 70% prediction, use 0.7. If it happened, subtract 1; otherwise, subtract 0. Multiply the answer by itself. So a 70% prediction scores 0.09 if it happens, or 0.49 if it doesn’t. This is called a Brier score. The election error tables average these scores across past races.
04 / For researchers
The measures, the data
and the tests.
Every series below is computed by code from a fixed corpus and a frozen week, and every one can be traced to the saved source responses. Definitions are published; a change to any of them is a new version, and the old version stays readable.
| Measure | Definition | Role now |
|---|---|---|
| Agenda share | Stories in a fixed collection whose text matches a topic's published query, divided by all stories in the collection that week. Grounded in agenda-setting research. | Claims are filed on it |
| Headline topic share | The share of in-scope headlines a classifier assigns to each topic. Stories are matched on full text but labeled from the headline, so the two shares measure different things and are never averaged. | Cross-check until it passes the label test |
| Repetition | The share of a topic's headlines that are near-copies of a headline from a different outlet the same week. A larger share is not the same as a rewritten article. | Reported in every note |
| Attention | Daily Wikipedia pageviews by people, not bots, for a fixed list of articles. People looked; it does not say they agreed. | Reported; a second series for share moves |
| Official record | Documents and records published by agencies, Congress, the FDA and the trial registry on the same topics. Publication is not implementation, and a clearance is not evidence that a technology works. | Cross-check |
| Quantified headline share | On the health desk, the share of in-scope headlines that state a quantity, by a fixed rule. A number in a headline is not a verified number. | Descriptive |
| Attention dynamics | For each topic's weekly share: when a burst begins, when it peaks, how fast it decays, how long it lasts, how much persists, and whether it re-ignites. Computed on frozen weeks with fixed parameters. | Descriptive |
| Second-series confirmation | Whether a move in share is matched by a move in a different family of series in the same week. A move in one feed is not a tide. | Reported in every note |
The label test
Before a classifier's labels count, two people code a gold sample drawn across every predicted label and weighted back to the full set. The labels pass only with macro-F1 of at least 0.85, recall of at least 0.75 on every topic, and topic shares within 5 points of the coders'. Until then the labeled series is a cross-check.
Frozen weeks
Each week is frozen 36 hours after it ends, with a fingerprint of every saved source response. Later corrections are recorded beside the frozen value and never replace it. Claims resolve only against frozen weeks of live collection.
Rules filed first
Each desk's filing rule, tolerance and baseline were written down before any claim, then replayed on a reconstructed year. The replay results are published below, including where the rule did no better than the simple baseline.
Hindsight (claims-policy-v3): over 47 reconstructed weeks the forecast was closer than the four-week mean in 0.53 of weeks; mean error 0.066 against 0.066 points. Desk page →
Murphy (claims-policy-v1): over 47 reconstructed weeks the forecast was closer than the four-week mean in 0.49 of weeks; mean error 0.132 against 0.130 points. Desk page →
| Series family | Desk | Source | Series | Unit | Cadence | Role | Licence |
|---|---|---|---|---|---|---|---|
| Agenda share by topic | Hindsight | Media Cloud collection | 24 | percent | weekly (Mon-Sun) | metric | Derived counts |
| Agenda share by topic | Murphy | Media Cloud collection | 12 | percent | weekly (Mon-Sun) | metric | Derived counts |
| Stories in the collection (denominator) | Hindsight | Media Cloud collection | 1 | stories | weekly (Mon-Sun) | cross-check | Derived counts |
| Stories in the collection (denominator) | Murphy | Media Cloud collection | 1 | stories | weekly (Mon-Sun) | cross-check | Derived counts |
| Headline topic share (Jev labels) | Hindsight | Media Cloud headlines, labeled | 14 | percent | weekly (Mon-Sun) | cross-check | Derived labels |
| Headline topic share (Jev labels) | Murphy | Media Cloud headlines, labeled | 14 | percent | weekly (Mon-Sun) | cross-check | Derived labels |
| Quantified headline share | Murphy | Media Cloud headlines, by rule | 1 | percent | weekly (Mon-Sun) | cross-check | Derived from Media Cloud headlines |
| Wikipedia pageviews | Hindsight | Wikimedia | 2 | views/day | daily | metric | CC0 (Wikimedia analytics data) |
| Wikipedia pageviews | Murphy | Wikimedia | 16 | views/day | daily | metric | CC0 (Wikimedia analytics data) |
| Federal Register documents | Hindsight | Federal Register | 2 | documents | daily | metric | US public domain |
| Federal Register documents | Murphy | Federal Register | 4 | documents | daily | cross-check | US public domain |
| Congressional Record and bills | Hindsight | GovInfo | 2 | documents | weekly, keyed by Monday | cross-check | US public domain |
| Congressional Record and bills | Murphy | GovInfo | 2 | documents | weekly, keyed by Monday | cross-check | US public domain |
| FDA device and drug records | Murphy | openFDA | 4 | records | weekly, keyed by Monday | cross-check | CC0 1.0 (openFDA) |
| Electricity demand (cross-check) | Hindsight | EIA | 1 | MWh/day | daily (from hourly) | cross-check | US public domain |
Rights
The news collection comes from Media Cloud under research terms, so the desks publish and study its series and do not license them commercially. Wikipedia pageviews and openFDA data are public domain dedications; federal documents are public domain. A press feed is read only after its reuse rights are recorded.
What the measures do not show
Coverage share is what editors ran, not what the public thinks. Pageviews are looking, not agreeing. A count of filings, clearances or registrations is a count of a process. Nothing here measures approval; only polls and votes do.
Reconstructed history
The year before each desk opened was retrieved after the fact. It sets the baselines, the tolerance and the dynamics, is labeled as reconstructed wherever it appears, and never resolves a claim.
05 / The election field
What do the polls say?
How much could they miss?
We research a selection of races closely and now account for every House and Senate seat in a separate chamber experiment. You can explore possible seat totals, shared polling mistakes and unresolved races. This does not predict which party will control Congress.
We publish daily election updates. Our current model is called Compass. It uses polls, past votes in the same area and candidates’ past elected experience to estimate where a race is heading. The tests do not show that the more complicated method beats a simpler one. We need results from future elections. The sliders below use made-up numbers to explain a simpler calculation. See the new model tests.
We are also testing Tide. It asks how national opinion might shift an area’s usual vote, and waits for three recent polling firms before showing estimates. Compare Compass and Tide.
Explore the whole-chamber seat scenarios. These include continuing Senate seats and show gaps instead of assigning missing races to a party.
Try different assumptions
How safe is a polling lead?
Made-up numbers to show how it works.
If the Democrat polls at 48% and the Republican at 45%, the lead is 3 points. A negative number means the Republican is ahead. The second slider lets you try a wider or narrower range of polling mistakes; these settings are assumptions for this invented example. We use the six-point setting as a simple comparison when testing past elections.
Possible leads in this example
The taller parts of the curve show leads this calculation considers more likely. This is a simplified example. Other candidates and different election rules could change the result.
How we combine polls
We show polls from the last 45 days. Our test predictions use only the last 21 days. Newer polls count more: after two weeks, a poll counts half as much as when it was new. Many polls from one company do not count as much as the same number from different companies.
What election betting tells us
On prediction markets, people trade bets on election results. The prices offer another view of what might happen, but can be affected by the rules and how many people are trading. We keep them separate from the polls.
What we still need to check
We don’t have every poll, and we haven’t checked every original report. A separate experiment now tests sample-size weights and polling-company differences. Its historical adjustments measure disagreement with other pollsters, not a proven partisan bias or a quality rating. Small House gains and Senate regressions do not justify changing Compass. There is no evidence yet that our predictions beat other polling websites.
Election data last collected: 2026-10-23T23:45:33.779692+00:00. The page shows the latest saved collection. It does not update continuously. Race names use state abbreviations. “SEN” means Senate; a number means a House district.