Stage 1 — Triage to a shortlist

Stage 0 produced a universe. Stage 2 can afford about ten deep dossiers. This stage
picks which ten, on two axes at once:

• change_expectation — how much could this print move the stock?
• ai_edge — can careful research actually say something useful about it?

Both must be high. A biotech binary readout scores 95 on the first and 15 on the second:
it will move enormously and nobody can forecast it, so it is a waste of a dossier. Being
willing to say that is the point of this stage.

1. Load state

python3 scripts/run_paths.py --json

Read 00-universe.json. If it is missing, stage 0 did not run or did not publish —
run the earnings-universe skill first, then come back.

2. Decide whether to run at all

Read config/pipeline.yaml. If eligible names ≤ triage.skip_if_universe_at_or_below
(default 10), skip the screen entirely. Write all eligible names straight to
01-shortlist.json with change_expectation and ai_edge set to null,
"triage_mode": "skipped_small_universe", and a selection_rationale of
"universe at or below the triage threshold — all eligible names carried forward".
Then jump to step 5. Screening eight names down to eight costs tokens and decides
nothing.

3. Scout the universe

Split the eligible names into batches of triage.batch_size (default 15) and spawn one
earnings-triage-scout subagent per batch, up to triage.max_scouts. Run them in
parallel — they are independent.

These scouts are deliberately cheap (Sonnet, medium effort). Do not substitute the
deep researcher here; depth at this stage buys almost nothing, because most of these
names will be discarded within the hour.

Give each scout the tickers, company names, sessions, event dates, and market caps from
the universe file. Nothing else.

If the universe was rebuilt by web search (status: partial), tell the scouts so and
ask them to confirm timing more carefully than usual.

4. Rank and cut

Collect the scout JSON. For each name:

priority_score = weight_change_expectation * change_expectation
               + weight_ai_edge           * ai_edge

Then apply, in order:

1. Drop anything with timing_confirmed: false or tradeable: false.
2. Drop anything below min_change_expectation or below min_ai_edge. These are
   hard floors, not tiebreakers — a name that fails either is not worth a dossier
   regardless of how well it scores on the other axis.
3. Sort by priority_score descending and take triage.shortlist_size (default 10).

If fewer than ten names clear the floors, take fewer. Do not backfill from below the
floor to reach a round number. A six-name shortlist of genuinely researchable setups
beats a ten-name list padded with four names nobody can call. Record the shortfall and
why.

Prefer a mix of AMC and BMO reporters when the scores are close — a shortlist that is
all after-close names leaves the next morning's window unexamined. Note the tilt if one
session dominates.

5. Write the shortlist

01-shortlist.json:

{
  "schema_version": 1,
  "run_date": "2026-08-10",
  "window_covered": "<copied from the universe file>",
  "triage_mode": "scouted | skipped_small_universe",
  "universe_total": 61,
  "universe_eligible": 34,
  "scouts_used": 3,
  "floors": {"min_change_expectation": 35, "min_ai_edge": 30},
  "dropped_for_floors": 12,
  "shortlist": [
    {
      "ticker": "TTWO",
      "company": "Take-Two Interactive",
      "session": "amc",
      "event_date": "2026-08-10",
      "market_cap_usd": 41200000000,
      "change_expectation": 78,
      "ai_edge": 64,
      "priority_score": 71.7,
      "expected_move_hint": "9.2%",
      "selection_rationale": "<one line: the catalyst, and why research helps>",
      "evidence": "good"
    }
  ]
}

Validate before publishing:

python3 scripts/validate_stage.py shortlist <run_dir>/01-shortlist.json

Fix anything it reports. Do not publish an invalid shortlist — stage 2 reads this file
unattended and will happily research a malformed entry.

Also write 01-shortlist.md: the window, the funnel counts (universe → eligible →
cleared floors → shortlisted), a table of the shortlist with both scores and the
rationale, and a short list of the most notable names that were dropped and why. That
last part is what makes the archive auditable — if the pipeline keeps discarding names
that then move 15%, the dropped list is where you would see it.

6. Log and publish

Append to _run-log.md:

## Stage 1 — triage (<HH:MM> CET)
- Mode: <scouted | skipped (universe <= threshold)>
- Funnel: <n> universe -> <e> eligible -> <f> cleared floors -> <s> shortlisted
- Scouts: <k> subagents (sonnet/medium)
- Session mix: <a> AMC / <b> BMO
- Notable drops: <ticker (reason)>, ...

python3 scripts/update_index.py
scripts/publish.sh "stage 1: shortlist for <YYYY-MM-DD> (<s> names)"

7. Report

The funnel in one line, the shortlist as a compact table (ticker, session, both scores,
one-line reason), and any warning stage 2 needs — thin evidence, an unbalanced session
mix, or a short list.