How signals are detected, scored, connected and red-teamed. Deterministic where it matters; models only write and review prose, never decide what gets flagged.
Radar (0–100) — how much is going on at a company and how much it matters. It is an attention score, not a direction: it combines the company's signal scores so that the single most consequential signal dominates and each additional signal adds a little (a "noisy-OR"), then it is scaled up when independent filings form a setup. Read it as: <40 little of note · 40–69 elevated · 70+ a lot is happening, look at the setups.
Support ▲ (0–100) — the same combination restricted to signals where insiders or large holders put their own money in: open-market insider buys, insider-buying clusters, 10% holder purchases, a new Schedule 13D. A dash means none in the last 30 days.
Dilution pressure (0–100) — Radar restricted to financing and balance-sheet signals: shelves, offerings, ATMs, equity lines, converts, warrant inducements, short runway, negative equity, share-count jumps. High = more shares are likely to be sold.
Listing risk (0–100) — Radar restricted to exchange-standard signals: deficiency notices, bid-price / equity / MVLS deficiencies, hearings panels, delisting language, reverse-split references, proximity to Nasdaq's $5M floor.
Governance risk (0–100) — Radar restricted to reporting-integrity and going-concern signals: auditor change, non-reliance (restatement), late filings, executive-departure clusters, defaults, bankruptcy, going-concern language.
Direction — Pressure ▼ / Support ▲ / Mixed ◆ / Context · — stated for the existing shareholder. Pressure: more shares coming, survival or listing at risk, or weaker reporting. Support: insiders or holders buying. Mixed: both present (e.g. insiders buying while a raise is coming). Context: informative alone but not directional. This is a characterization of what the filings say, not a forecast or a recommendation.
Runway (months) — cash and short-term investments divided by monthly operating cash burn at the latest reported period. "Cash-flow positive" means operating cash flow is positive. Every ticker page shows the balance-sheet date and how stale it is; a financing after that date is the main reason a figure can be out of date.
Signal score (0–100) — importance × novelty × confidence × magnitude, per signal (details below). Hover a score on a ticker page to see the four components.
Setup and strength (0–1) — a setup ("chain") is a rule that stitches independent filings into a thesis, e.g. Forced raise = cash need + live selling machinery. Strength is how completely the rule's ingredients are present and how confident the evidence is. Every setup lists what would confirm and what would prove it wrong.
Bottom line — the plain-English takeaway on each ticker page is generated from the strongest setup (or, absent one, the most consequential signal); it names the next filing to watch. It is written to be specific and falsifiable, never a buy/sell view.
Red-team verdict — analyst notes are drafted by one model and attacked by a separate model; publish = every claim traced to the evidence packet; publish with caveats = supported but the listed caveats must be read alongside; reject = not shown.
Base rate — how common a signal type is across the whole universe in the last 30 days (shown in the glossary). Rare events score higher on novelty; a signal that a quarter of companies show barely moves the score.
Exchange-listed (Nasdaq / NYSE / NYSE American) US filers whose 10-K cover-page public float is below $300M (≈1,400 companies as of Aug 2026; nano-cap < $50M ≈ 40% of them). Sources: SEC EDGAR only — submissions metadata (forms, 8-K item codes, acceptance time), full-text search (EFTS), XBRL frames and companyfacts. Everything is public-domain and free; the SEC fair-access limit (10 req/s) is respected.
A dated, evidence-linked observation about one company. Every signal links to the exact filing that produced it. Signals are detected deterministically (no LLM in the loop) and carry four sub-scores in [0,1]:
| Sub-score | Question | How v0 sets it |
|---|---|---|
| importance | How consequential is this type of event for a microcap? | Fixed prior per type (e.g. non-reliance 0.90, MVLS deficiency 0.90, ATM 0.75, 424B takedown 0.70, officer departure 0.40). Priors come from the dilution / delisting / restatement literature and practitioner behaviour (what Dilution Tracker, Bear Cave, short-sellers actually watch). |
| novelty | How unusual is it — for this company and for the universe? | Company: first occurrence of this form/item in the last 12 months = 0.9, 2nd 0.6, 3rd 0.4, 4+ 0.25. Universe: blended 40% with (1 − 4·base_rate)^0.5 so a type that 25% of companies show scores ~0 on this component and a 1% type ~0.98. Going-concern novelty is upgraded when a per-company EFTS lookback shows no going-concern language in the prior 12 months (a new GC). |
| confidence | How strong is the evidence? | Structured metadata (form type, 8-K item code) 0.90–0.95. Phrase match in a filing 0.40–0.75, upgraded when the phrase sits in an attached agreement (EX-10 → 0.85), an Item 3.01 8-K (0.95), or ASC 205-40 going-concern note language (0.90); downgraded for proxy proposals (0.50), S-3 boilerplate (0.50), auditor consents. Fundamentals: decays with balance-sheet staleness (≤75d 0.75, ≤140d 0.60, else 0.40). |
| magnitude | How big relative to the company? | Runway months, count of prior 424Bs, share-count growth, cluster size, float distance to the $5M MVLS floor. |
score = 100 · importance · (0.5 + 0.5·novelty) · (0.4 + 0.6·confidence) · (0.7 + 0.3·magnitude), ×1.08 for after-hours / Friday-evening 8-Ks.
Multiplicative on purpose: a weak dimension pulls the score down; a boilerplate phrase match can never outrank a hard 8-K item.
radar_score = noisy-OR of the company's signal scores (top signal dominates; each additional signal adds with 0.8^i decay),
× (0.75 + 0.25 · strongest chain strength), × 0.9 if the top signal is low-confidence text. Dilution pressure, listing risk and
governance risk are the same noisy-OR restricted to their categories.
Rules that stitch independent filings into a setup, each with confirmation and invalidation criteria: Forced-raise (cash need + live selling machinery), Listing cliff (deficiency + structural reason cure is hard, incl. Nasdaq's 2026 one-reverse-split-per-year rule and $5M MVLS no-cure floor), Death-spiral watch (variable-priced paper + reverse split / notice), Reporting-integrity stress (auditor change or non-reliance + late filing / departures), Shelf loaded (first S-3 in 12 months with <12 months runway), Activist meets distress (new 13D + cash pressure).
liquidity = latest cash (+ short-term investments at the same date) from companyfacts (frames as fallback); burn = latest fiscal-year operating cash outflow, or annualized YTD outflow when ≥170 days are available, else annualized quarter flagged low-confidence; runway = liquidity / monthly burn. Banks, insurers, REITs and SPACs are excluded from runway logic. Every figure carries its basis and balance-sheet date; the main failure mode (a raise after the balance-sheet date) is why staleness lowers confidence and why the analyst layer must check for financing 8-Ks/424Bs since that date.
Deterministic ranking picks at most N companies per run for LLM analysis. The Analyst writes: what happened, what most
people will notice, what we noticed, why it may matter, evidence (with URLs), what could prove us wrong, confirmation and
invalidation. A separate Skeptic call tries to disprove every claim (unsupported facts, runway arithmetic, boilerplate vs.
real going concern, novelty, advice-like or defamatory language, base rate). Only publish / publish_with_caveats
verdicts are released. Model cost per published item is recorded.
Every run persists signals, chains and company snapshots in SQLite. An outcomes job records, at 30/90/180 days after each signal: subsequent 424B / offering 8-K, reverse split, delisting (Form 25), NT filing, non-reliance, share-count change, and (once a price feed is added) drawdown. Importance priors are then replaced by observed hit-rates per type and per chain, and novelty by observed conditional frequencies ("what happened historically when X and Y co-occurred").
Public float is a lower bound for market cap and is as-of the last fiscal Q2; newly-IPO'd companies (no 10-K) are missing; Form 4 direction (buy/sell) is not parsed yet; no price data; EFTS phrase matches can be boilerplate — treated as low-confidence until a refinement query or document classification upgrades them.