How the dispatch is written

How the Daily Dispatch turns archived widget documents into four checked passages, and how the Options Dispatch screens and curates its option cards.

The question

The Daily Dispatch and the Options Dispatch are written each morning from the previous session's data. The Daily Dispatch describes in prose what the previous session's numbers showed. The Options Dispatch shows the Options passage and the option cards: the option structures that pass a fixed set of rules, with their prices.

The method

Segments and evidence

The Daily Dispatch has four segments, defined in a closed registry: Fed Watch, Rates Desk, Volatility and Options. Each segment lists the archived widget documents it uses, by widget, symbol and window. The archive is rebuilt each morning from the site's own widget data after the daily poll.

The Daily Dispatch job copies a segment's documents into a temporary directory and adds four files, which together form the evidence pack:

  • Glossary: definitions of the terms.
  • Persistence file: streak counts for the segment's main series.
  • Deltas file: today's values compared with the prior session and with five sessions back.
  • Recent passages: the segment's last seven published passages.

Long time-series tables are cut to their recent rows. The cut is recorded in the pack. When a pack exceeds its size budget, the segment fails; no documents are dropped to fit.

Write, check, judge

A language model drafts each passage in a fresh context, from a brief of trader questions and the pack alone. Three mechanical checks then test the draft.

  • Numerals: every number in the headline and body must trace to a value in the pack, allowing declared unit scalings. Dates are exempt.
  • Style: the target length is 150 to 300 words. The check fails a draft under 150 words or over 450, one and a half times the top of the target. It also enforces a paragraph band per segment. It bans headings, lists, typographic dashes, the first person and a list of stock phrases. A forecast word ("will", "likely") passes only when the same sentence attributes it to a model or a board. The Options passage also bans ranking words and stated reasons for curation.
  • Headline: a separate check on the headline line.

A second model, the judge, then compares the draft with the same pack. It lists every number, direction or streak claim that the pack does not support. It scores each brief question from 1 to 5, then reviews readability. Its verdict is strict JSON. Output that does not parse counts as a grounding failure.

A draft publishes when the checks pass, the unsupported list is empty and every score is at least 4. Otherwise one revision round runs, so a segment costs at most four model calls. Of the drafts that passed the checks, the Daily Dispatch job keeps the best. It prefers an empty unsupported list, then the higher total score, then the later round.

When no draft clears

When the best draft still contains unsupported sentences, the Daily Dispatch job cuts exactly those sentences. The remaining text publishes only if it passes the same checks, and the cut sentences are recorded. Otherwise the segment publishes a numbers note, built without a model. The note gives each main series' latest value and one-day change. It ends with "No narrative cleared grounding today." A numbers note never replaces a passage already published for the same day.

Cards

The card pipeline builds one card per symbol, expiry and structure from the previous session's chains. Only expiries at least 7 days away are used. The straddle uses the strike nearest spot; a tie goes to the lower strike. The strangle uses the 25-delta wings. Each leg needs a positive bid and ask, an uncrossed quote and a positive smoothed volatility. A card with a leg that fails any of these is not built.

Theo prices each leg by Black-Scholes on the forward, using that leg's own ORATS smoothed volatility. The fill ladder and close targets are built from the mid. The status is set by the first rule that matches:

  • Dead: the VRP z-score is below +0.10, or the VRP points are zero or less.
  • Borderline: the z-score is exactly +0.10, or a leg fails the spread or open interest check.
  • Live: all other cards.

The screen keeps live and borderline cards with 28 days or more to expiry.

Curated

Curated is a set of pass rules. A card is curated when all five rules hold:

  1. Its symbol is in the curated class, a configured ticker list.
  2. No estimated earnings date falls inside the hold. Entry is the first session after the trade date. The hold ends 14 calendar days later or at expiry, whichever comes first. Every ETF passes this rule.
  3. Its status is live or borderline.
  4. Its DTE is 28 to 60.
  5. It is a straddle.

The Curated view lists cards by symbol, then by expiry. When no card passes and an earnings check returned unknown, the Curated view displays a notice that the earnings calendar is stale.

Inputs and lag

The poll starts at 00:15 UTC, Tuesday to Saturday, and builds the cards from the previous session's close. The archive rebuild runs at 07:00 UTC and the Daily Dispatch at 07:45 UTC on the same days. The Daily Dispatch date is the trading date of the Volatility Board documents. The Daily Dispatch does not backfill. A missed run leaves a gap. Every input has the lag described in the daily run and the lag rule.

How it was validated

The prose checks test whether each passage matches its pack. They do not test whether a passage later proved right. In a 2026-09-05 dry run, every sentence the judge rejected contained a real defect. Four of the five failed drafts failed on form: three exceeded the word limit and one put text before its headline. After that run, the word limit was set above the target. A 2026-09-16 review found that the judge caught five grounding defects the older checks had passed.

Each run appends a record with the code version. For each stage the record stores a hash of the role prompt, a hash of the input, the output, the duration and the exit code. A correction is a new version of the same segment and day.

The two Curated gates come from a naked replay of short cards, 2022 to 2026, described in how the models were tested. On that replay, cards that passed the class gate beat the pool of all cards by 0.050 of entry credit, 90% interval 0.027 to 0.076. Cards that passed the earnings check beat the pool by 0.007, interval 0.002 to 0.013.

Limits

  • The numeral check does not test a qualitative statement that contains no number. Only the judge checks such statements.
  • A passage covers only the documents its segment lists.
  • Older passages keep the segment labels in use when they were written.
  • Every card price is a past close. The system uses no live quotes.
  • The Curated gates use the symbol class and the earnings calendar only. They do not use greeks, direction or the Volatility Board.
  • The replay behind the Curated gates books no commissions and fills at end-of-day prices.