Stratum's live signals are deterministic: peer-reviewed forensic-accounting models computed from regulator-published XBRL facts. No model weights, no black boxes — every flag exposes its inputs, and every input traces to a filed document on EDGAR.
Beneish M-Score (Beneish, Financial Analysts Journal, 1999): an 8-ratio probit model of earnings manipulation; M > −1.78 matches the manipulator profile in the original sample. Supplemented by the Sloan accrual ratio (Sloan, The Accounting Review, 1996): high accruals relative to assets predict earnings reversals.
Dechow–Ge–Larson–Sloan F-Score (Contemporary Accounting Research, 2011), Model 1: a logistic model of AAER misstatement firms. F is the ratio of predicted to unconditional misstatement probability; F > 1 is above-normal risk, > 1.85 substantial. RSST accruals are approximated where full components are not tagged; the evidence block records the construction used.
BIS zombie definition (Banerjee & Hofmann, BIS Quarterly Review, 2018): interest coverage below 1.0× for three consecutive fiscal years. Context from Altman Z″ (Altman, 1995 four-variable, book-equity variant) — no market data required.
Defined-benefit funded ratio from tagged plan assets and benefit obligations; severity scales the deficit against shareholder equity.
Persistent effective-tax-rate gap versus the statutory rate across three years. A low ETR is frequently legitimate; the signal triages persistence and magnitude for review of the tax footnote.
Loughran–McDonald dictionary densities (Journal of Finance, 2011) — uncertainty, negative, and litigious language per 1,000 words in the latest annual report versus the prior one. Note: the full LM dictionary requires a commercial licence for commercial use; Stratum ships a research subset and loads the licensed CSV if provided.
Clusters of Form 4 filings in the 14 days preceding 8-K events over the trailing year. A timing pattern is a reason to read the filings — not an accusation.
Subsidiary jurisdictions parsed from Exhibit 21 of the latest 10-K, screened against a configurable watchlist. Mention-count proxy; ownership and revenue attribution belong to the LLM-enrichment roadmap.
Three signals use a language model (DeepSeek V4 by default; any OpenAI-compatible endpoint) for what deterministic math cannot do: reading claims, strategic language, and asset descriptions out of prose. The AI layer operates under three hard rules:
1. Quote or it didn't happen. Every extracted claim must include a verbatim passage; Stratum verifies the passage exists in the filing text and silently discards anything unverifiable. The discard count is reported in each signal's evidence. 2. Metered. Every call is logged to a cost ledger (/api/costs: spend, tokens, cache-hit rate, cost per company) and stops at a configurable daily budget. 3. Deterministic before probabilistic. Where a signal can be split (P-10's 13D counts, P-11's PP&E intensity), the deterministic part runs first and stands alone if no LLM key is configured.
Bounded, budget-aware — not open-ended autonomy. Triage ranks the universe for $0 (pure arithmetic, no model call); investigation runs a hard-coded 4-step sequence capped at 5 steps / 2 LLM calls, enforced by a counter in code, not a prompt; the watch agent attaches a memo to a triggered alert only when budget allows, never blocking delivery. Full rationale in stratum/agents/README.md.
60% weight on the strongest signal, 40% on the mean of available signals; the level is the maximum across signals, and coverage is always reported. One triggered signal is enough to surface a company — by design.
Tag coverage varies by filer; sub-ratios that cannot be computed default to their neutral value and are listed in the evidence. These are screening models with meaningful false-positive rates — Stratum ranks where to look first; the filings remain the ground truth. Nothing here is investment advice.