AssetFlow IntelligenceResearch
Coverage and data quality
Research · Scorecard · Coverage & data quality · Names
What each publication covers, what it measured in its latest issue, the rules its tables apply, the known defects of the stores it reads, and what has been frozen point-in-time so later scoring can be checked against what was visible at the time.
Publications
The Attention Board
daily · audience public · What is the world paying attention to today, and which listed names does that attention reach?
universeThe public encyclopedia's front-page news items for the day, the latest fully traced month's stories and the listed names their readers reached, plus the day's news-intensity board and 50-day trend crossers across the published attention universe.floorsTraced stories need a non-zero share of front-page audience; the news board lists the 15 highest intensity readings; crossers up to 12.cutoffsStories = posted items; traced month = the newest month with navigation data; crossers within the last 5 sessions.latencyDaily; audience tracing lags posting by one to two months because navigation data arrives monthly. The day's front page falls back up to 10 days when the record is empty.revision policyFront-page records and monthly navigation data are not revised once published; the news-intensity partition the issue reads is frozen point-in-time per date from 2026-09-02.tracked newsFront-page items and a news-intensity board (article volume against each name's own trailing 30 days).
No published issue yet.
Table rules: today: top 8; dedupe story_text · traced: kind eq blurb; mp_share gte 1e-06; top 8; dedupe story_id · reached: top 15; dedupe Ticker · pulse: top 10 · crossers: top 12; dedupe ticker_primary
Editorial Attention
monthly · audience public · Which company entries in the public encyclopedia are being rewritten most heavily against their own history, and what changed?
universeA fixed watch list of 46 large companies, each pinned by (ticker, page title) to the company's own public-encyclopedia entry; subsidiary, product and brand pages are excluded.floorsAt least 5 edits in the month to appear; contested = at least a fifth of the month's edits reverted; substantive revisions need at least 200 bytes.cutoffsThe latest full month against 24 months of the entry's own history; the top quintile of own history for the revision log.latencyMonthly, from a pinned snapshot of the edit history (currently the 2026-07 dump); months after the last full month are not covered.revision policyThe snapshot is pinned and does not refresh; an issue publishes only when the covered month has moved. No point-in-time dump is taken (the source is already frozen) and no scored claim is logged until the store is deliberately re-pinned.tracked newsNot used.
latest issue2026-09-01data as of2026-07-31stale33 day(s) behind the build daterowsEditing intensity in the latest full month, against each entry's own history: 30, Contested editing in the latest full month (share of the month's edits that were reverted): 6, Largest net changes in the month: 8, Substantive changes on entries in the top quintile of their own editing history (verbatim edit summaries): 0, Snapshot: 1names shown30
Table rules: intensity: months[-1].ym eq (the run's own value); months[-1].edits gte 5; top 46; dedupe query_page · contested: months[-1].ym eq (the run's own value); months[-1].edits gte 5; months[-1].revert_share gte 0.2; top 15; dedupe query_page · bytes: months[-1].ym eq (the run's own value); months[-1].edits gte 5; top 8; dedupe query_page · summaries: housekeeping falsy; summary truthy; section truthy; e.query_page truthy; top 25 · snapshot: top 1
Fed Language Tracker
event · audience public · What is the Federal Reserve saying more of, and who is saying it?
universePublic-record transcripts of Federal Reserve press conferences (chair-era speech, 2012 to date) and prepared speeches by governors and presidents; six fixed concepts in press conferences, one in speeches.floorsA rate on fewer than 5 mentions carries its mention count; the speeches table lists only months with at least one mention. Every row prints the number of press conferences (or speeches) behind it, because the event count, not the segment count, is the sample: Warsh's 2026 rows rest on two press conferences, and a rate from two events is two observations whatever the segment count says.cutoffsTerms are matched as plain substrings inside the speaker's own segments, per 100,000 characters; concept comparisons under a 25% normalised gap are called similar.latencyEvent-driven: a new issue when a press conference or a new month of speeches enters the record; otherwise held unchanged. Data date = the current chair's latest press conference.revision policyTranscripts are public record and are not revised once ingested; per-press-conference rates are rebuilt from the segment store for each base rate, and a claim resolves against the speaker's next press conference (void after 120 days without one).tracked newsNot used.next eventNext press conference: 2026-09-16
latest issue2026-09-01data as of2026-07-29stale35 day(s) behind the build daterows'Inflation expectations' — press-conference rate per 100k characters, by chair and year: 10, 'Labor market' or 'employment': 10, 'Balance sheet': 10, 'Tariff': 10, 'Financial conditions': 10, 'Productivity': 10, 'Tariff' in prepared speeches, by speaker and month (last 12 buckets): 12, Latest press-conference remarks by the current chair: 6names shown0
Table rules: inflation_expectations: top 12 · labor: top 12 · balance_sheet: top 12 · tariffs: top 12 · financial_conditions: top 12 · productivity: top 12 · speeches_tariffs: top 12 · remarks: top 6
Regime Watch
weekly · audience public · Whose web attention just changed state, and how persistent is the new state?
universeA fixed watch list of 24 anchor names plus every name that crossed its 50-day attention trend in the last 7 sessions (up to 54 names), one regime model per name on its own attention series.floorsA name appears only when the model returns a current state; the long-dwell table needs 40 or more sessions in one state.cutoffsSwitch = current state began within 7 sessions; long dwell = 40 or more sessions; Days ÷ expected dwell above 5 is a fit flag, not persistence. The state, its days and its persistence are the FORWARD-FILTERED decode from 2026-09-02: each day is labelled with data to that day only, so a switch on the board is one that could have been seen that day. The smoothed full-series decode, which labels a day with later data, is kept as a research view (hidden columns, the chart shading) and is not the official label.latencyWeekly on Wednesday; the model is refit daily on the last 500 sessions and the issue reads the fit as of the latest session.revision policyThe official state path is the forward-filtered decode of a model refit daily on the full history, so the label for a past day cannot change with later data (the fitted parameters can, and a refit can therefore relabel history through the parameters, not through the decode). Vintages fit before 2026-09-02 hold only the smoothed decode, whose labels use later data and carry fewer short runs; the persistence base rate says how many of its names are filtered, and leans toward persistence while smoothed vintages remain. The board and the daily fit files are frozen point-in-time per issue date from 2026-09-02, and a persistence claim is void if the model's state count changes or the name leaves the board.tracked newsNot used in the tables; front-page story and regime-switch marks appear on the charts.
latest issue2026-09-02data as of2026-08-31stale2 day(s) behind the build daterowsAttention regime board (anchor names + this week's trend crossers): 40, Switched state in the last 7 days: 10, Long-dwelling states (40 or more sessions in one state): 12names shown48
Table rules: board: top 40; dedupe ticker_primary · switched: current_state_rank truthy; days_in_current_state lte 7; top 20; dedupe ticker_primary · long: current_state_rank truthy; days_in_current_state gte 40; top 12; dedupe ticker_primary
Sector Attention Breadth
weekly · audience public · Which sectors are gaining web attention broadly, and which gains rest on a few names?
universeOperating companies in the published attention universe with a sector label (about 2,100 names); sectors with fewer than 20 names, the unclassified residual and the digital-asset bucket are excluded from the breadth table.floorsAt least 20 names with a prior-session reading per sector; the crossover tables list up to 15 names each side. The breadth table lists every sector that clears this floor — eleven as of 2026-09-02, under a row cap of 20 that has never bound — so a count of its rows is a count of the eligible sectors in the universe, not of a ranked cut.cutoffsAdvancing = attention value above the previous session, declining = below; flat names are out of the denominator. Crossers = names crossing their own 50-day average within the last 7 days. Top-5 share = the last 7 calendar days of positive day-over-day log changes, equal-weighted per name.latencyWeekly on Monday for the latest session; attention values are the previous session's close.revision policyThe attention store is append-only and machine-tested not to move; the breadth table and the crossers are frozen point-in-time per issue date from 2026-09-02. Universe membership and sector labels are the current map applied to history (a dated copy is frozen daily). A sector's history before 2026-09-02 is therefore under today's label: a company reclassified since then sits in its new sector for its whole past, so a multi-year sector base rate knows which companies were reclassified. From 2026-09-02 the daily frozen map lets history be rebuilt under the labels of the day.tracked newsNot used in the tables; front-page story marks appear on the charts only.
latest issue2026-09-02data as of2026-08-31stale2 day(s) behind the build daterowsSector advance/decline in web attention (latest session, 90-day index): 11, Names crossing above their 50-day attention trend this week: 14, Names crossing below their 50-day attention trend this week: 15names shown29
Table rules: breadth: n_names gte 20; sector ne Unclassified; sector ne crypto; top 20 · up: top 15; dedupe ticker_primary · down: top 15; dedupe ticker_primary
Data-quality register
14 known defect(s), one row per store; register built 2026-09-02.
| Store | Defect | Window | Affects | Effect | Detected | Status |
|---|
| web attention (value panel) | Digital-asset market caps are present but stale: the upstream cap feed froze silently. | 2026-02-17 to 2026-08-04 | any digital-asset row in attention tables | Attention-value ratios for those names are computed on a frozen denominator; equities are unaffected. Filter by asset class before comparing. | 2026-08 | open |
| web attention (value panel) | Two builds of the same pipeline meet at one date: the daily build owns everything from 2023-01-01, an older build covers what precedes it. | boundary 2023-01-01 | base-rate page, any multi-year attention history | Measured on both sides in the panel's validation table; no known level break, but any series crossing the boundary should be read with that in mind. | 2026-08 | documented |
| web attention (universe) | A name whose attention falls below the publication floor silently leaves the published universe; the exit is indistinguishable from a delisting. | always | sector breadth, regime watch, base-rate page (coverage counts) | Universe membership is a function of attention, so counts of names are not counts of companies. | 2026-08 | open |
| web attention (history) | The historical cross-section is today's universe backfilled: a name onboarded later carries rows for earlier dates. 115 of 2,135 names end before 2026-06 and 303 have gaps; there is no dated onboarding record before 2026-08-17. | 2015-05 to 2026-08 | base-rate page, any historical screen reconstruction | Values themselves are frozen once published (append-only, machine-tested), but the set of names visible on a past date is not the set that was visible then. | 2026-09-02 | open |
| sector labels | Each name carries one sector label, the most recently observed; there is no dated sector history. | all history | sector breadth, top-5 share of gains, sector-conditioned base rates | Every historical sector figure is computed with today's sector map. A dated copy of the label file is frozen daily from 2026-09-02. | 2026-09-02 | mitigated from 2026-09-02 |
| price mirror | The mirror's universe begins in May 2018; only 9 to 21 names per year are priced before that. | before 2018-05 | base-rate page | The study starts 2018-05-01 by default and reports dropped, unpriced events per screen. The mirror does retain delisted names (138 of 2,312 end before 2026-06), so it is not survivorship-trimmed. | 2026-09-02 | documented |
| retail order flow | Each date is fetched once and never re-fetched; a vendor correction issued after the first fetch is never observed. Seven rolling full-history snapshots are retained. | always | retail flow ledger, retail base rates, claim resolution | Point-in-time by accident rather than design: values do not move, but neither do corrections arrive. A fingerprint of the trailing 60 dates is frozen daily from 2026-09-02 and checked before any claim is scored. | 2026-09-02 | mitigated from 2026-09-02 |
| retail order flow | About 8,700 of 10,800 names round to zero at the feed's four-decimal reporting floor on a given day. | always | retail flow ledger | Absence from a retail table means unmeasured, not inactive; percentiles are computed against non-zero days only. | 2026-08 | documented |
| encyclopedia edit history | The edit store is a pinned monthly dump (last full month 2026-07) and does not refresh on its own; mixing dump vintages produces contradictory title chains, so re-pinning is a deliberate act. | 2001-02 to 2026-07 | editorial attention | Claims about the next month cannot resolve until a re-pin, so that publication carries no scored claim. Revert share counts edits that ARE reverts (the reverting action), not edits that were reverted. | 2026-08-25 | documented |
| attention regime model | The state path is a full-series decode of a model refit daily (capped at the last 500 observations; the number of states is chosen per name). From 2026-09-02 the official label is the forward-filtered decode (each day labelled with data to that day only); the smoothed path is kept as a research view under smoothed_* and the persistence base rate says how many vintages are filtered. Vintages fit before that date remain smoothed. | always | regime watch, state-persistence base rates | A smoothed path has fewer short runs than the live view, so run-length base rates lean long until daily model vintages accrue (archive from 2026-08-20, frozen daily from 2026-09-02). Expected dwell is a model quantity, not a measured spell length. | 2026-09-02 | mitigated from 2026-09-02 |
| central-bank transcript corpus | Press-conference transcripts run 2012-01 to date; the current chair has two press conferences in the corpus as of 2026-09-02. Prepared speeches run 2018 to date. Meeting transcripts end 2020-12 (five-year release lag). | varies by document kind | Fed language tracker | Any per-speaker rate with few documents carries its count; a rate change across document kinds mixes role and era. | 2026-09-02 | documented |
| tracked news | The front-page news feed changed source on 2026-06-19; the two feeds are not subsets of each other. | boundary 2026-06-18/19 | news intensity (tracked news split), attention board | Intensity baselines are 30-day, so readings after 2026-07-19 are internally consistent; no comparison should cross 2026-06-18. | 2026-08-20 | documented |
| front-page audience traces | Audience-flow data arrives monthly; traces of where a front-page story sent readers lag posting by one to two months. | always | attention board | The daily board separates stories on the page now from where the latest traced month sent readers. | 2026-09-01 | documented |
| backtest ledger | The status tool returns the newest 15 requests only. Paged from 2026-09-02: limit (to 200), offset, status and since filters; n_total and truncated are returned. | always | what we tested | The page is a window, not the full ledger; base rates over the ledger use the underlying files. | 2026-09-02 | mitigated from 2026-09-02 |
Point-in-time inventory
Each publication run is frozen once per issue date (first write wins; a failed or back-filled run never claims the slot). Input snapshots are frozen once per date beside them. Note the inputs record what the morning run READ — a store refreshed later the same day is the next date's snapshot.
| Publication | Frozen issue dates | First | Last | Reconstructed |
|---|
| The Attention Board | 1 | 2026-09-02 | 2026-09-02 | 0 |
| Editorial Attention | 1 | 2026-09-01 | 2026-09-01 | 0 |
| Fed Language Tracker | 2 | 2026-09-01 | 2026-09-02 | 0 |
| Regime Watch | 1 | 2026-09-02 | 2026-09-02 | 0 |
| Sector Attention Breadth | 1 | 2026-09-02 | 2026-09-02 | 0 |
| Input snapshot | Dates | First | Last |
|---|
| retail cohort event study (first-write cache, one per data date) | 1 | 2026-09-01 | 2026-09-01 |
| regime-model cache listing (vintages) | 1 | 2026-09-02 | 2026-09-02 |
| tracked-news partition the run read (copy) | 1 | 2026-09-02 | 2026-09-02 |
| retail order-flow store fingerprint (trailing 60 dates) | 1 | 2026-09-02 | 2026-09-02 |
| universe and sector map (copy) | 1 | 2026-09-02 | 2026-09-02 |
Base-rate study coverage
The base-rate study's coverage table is published at the standard tier (issue 2026-09-02); it is not reproduced on the public page.
Universe membership and sector labels are the current map applied to history; a dated copy is frozen daily from 2026-09-02.
Latest-issue counts are read from the issue as published; 'measured names' is the denominator the issue itself reports, when it reports one.
Built 2026-09-02T22:33:26