Passey Terminal
POLL
How it works

Help

News Terminal runs every headline through six numbered pipeline steps. Steps 3–4 use Loop 2 (existing factors help interpret news). Steps 5–6 use Loop 1 (observed price moves update factors and thesis). The Logs page uses the same step numbers.

Pipeline flowchart

Click a step to open Logs filtered to that step. Step 4 is multi-agent (LLM_MODEL list × LLM_RUNS).

Step 4 · Factor impact analysisLoop 2 · Estimatemulti-agent

Per matched ticker: multi-agent LLM runs (LLM_MODEL × LLM_RUNS). Average directional impact, majority bullish/bearish/neutral, then a merge pass for factor state.

logs →

Multi-agent substeps

  1. 4a

    Multi-run query

    Same classify prompt run LLM_RUNS times, cycling models from LLM_MODEL.

  2. 4b

    Vote + average

    Directional impact averaged; assessment / marketMoving by majority.

  3. 4c

    Merge factor state

    One follow-up LLM call merges notes from all runs into a single factor-state note.

View step 4 logs (runs + merge)

The two loops

Loops are conceptual — logs show which step ran and which loop it belongs to.

Loop 1 · Learn

Steps 5–6 · Price → learning

After price windows finish, each factor the LLM mapped is scored against the move. ≥1% → add relevance (size of the move). Flat/small move → nudge relevance down. Separately, the raw factors catalog may refresh. The living thesis updates when a headline was classified market-moving.

  • · Step 5 · Measure — Hyperliquid entry + 10m / 4h / 1d moves
  • · Step 6 · Learn — score each mapped factor vs that move
  • · ≥1% → add relevance; flat → −1; then the factors catalog may refresh
Loop 2 · Estimate

Steps 3–4 · Factors → interpretation

Uses existing factor drivers and manual context to interpret a new headline — map which themes apply, then run factor impact analysis per ticker (catalyst vs noise).

  • · Step 3 · Map factors — load catalog, LLM maps drivers + impact 1–100 (once)
  • · Strong impact (≥70 by default) → TG notify chat (news + analysis)
  • · Step 4 · Factor impact analysis — multi-agent: average impact, majority stance, merge factor-state note
  • · Uses mapped factors and manual context (Admin → Contexts)

Loop 1 · How relevance is updated

Applies only to factor_impacts the LLM mapped for this headline (step 3), after price windows finish (step 5).

Threshold

A move counts if any of 10m / 4h / 1d is ≥1% absolute on the factor’s target (market ticker, or sector proxies). Below that, the story did not confirm the theme.

Confirmed (≥1%)

Relevance is increased by +round(|move%|), at least +1. Examples: 1.2% → +1, 2.4% → +2, 4.9% → +5. Scores stay in 0–100.

Miss (mapped, but no ≥1% move)

Relevance −1 (floor 1). Themes that keep showing up without moving price slowly lose weight — so Loop 2 leans less on them next time.

Daily decay

Independently, all factor_impacts lose −1 relevance per UTC day (same floor). Headline miss penalties are per story; decay is background aging.

Data model (short)

news — one row per headline: matched tickers/proxies, price impacts, factorMappings (which drivers the LLM linked).

factors — raw driver catalog (macro / sector / market). One tradable risk dial per theme — a bypass or mine exposure is an update to Hormuz shipping risk, not a new factor. Edited on Thesis → Raw factors.

factor_impacts — links a factor to a target (e.g. GOLD, metals, macro) with its own relevance score; top drivers highlighted on Thesis tab.

relevantTickers — not the same as matched tickers: only symbols that moved ≥1% after the headline (used in step 6).

Factor catalog · one dial per theme

The factors LLM (step 6) and the mapping LLM (step 3) share this rule. Prompts live in Admin → Prompts.

A factor is a structural risk dial in a working model (stocks / flows / expected flow changes) — not a headline and not a datapoint. News that changes how that dial should be read is an update. Expected confirming prints update evidence; they do not become the primary dial.

Example: if Hormuz shipping risk is the working dial, do not add hormuz-bypass-shipping or treat “Saudi production −23%” as a separate top factor — that print is evidence under the dial. China raising domestic fuel prices is usually a different mechanism, not a map onto Hormuz geopolitics. Prefer empty mappings over false positives (false high-impact maps trigger TG alerts). After a true strong change, the dial’s situation (factors.trace) is updated immediately so republished headlines do not spam alerts.

Each factor carries why it matters and an invalidation. Theses must state the working model (primary dials by relevance), not lead with the latest news paraphrase.

Where to click

  • Daily brief — trader cockpit: material factor changes (≥50): Macro first, then sectors with markets nested, open positions on the side with thesis, then desk chat. Logs stay under More.
  • Home — pick a ticker (default WTI): thesis, factor impacts (click a row for why / invalidation / related news), and matched news for that market.
  • Trading — hot news, chart dots, assessment, and trade in one place.
  • More → News — all matched headlines; click one for impacts, factor mappings, and the pipeline trace.
  • More → Logs — all pipeline steps; filter by step at the top; expand a row for details; click a news ID to filter to one headline.
  • More → Alerts — TG thesis-change alerts, plus an Explore tab to browse factor-change news by impact threshold.
  • More → Thesis — factor_impacts + synthesized thesis per macro/sector/ticker.
  • Tickers — history, chart, living thesis, and manual context for one market.
  • More → Admin → Contexts — manual regime notes fed into step 4 factor impact analysis (admin wallets only).