Skip to content
Deal Intelligence

02How it works

Eight steps. Two of them use an LLM, and neither can change the facts.

Mint marks work done by rules you can read. Violet marks text written by the model. The model only ever sees candidates the engine has already confirmed.

  1. 01Deterministic

    Collect

    SEC EDGAR, NGX, NSE Kenya and JSE SENS disclosures, African and global news, in parallel. Each source fails on its own without blocking the run.

    filings + news inside the lookback window

  2. 02Deterministic

    Detect

    115 patterns across 8 signal types. A source must name the company, and the event must be in the same clause as the name.

    signal candidates, or none

  3. 03Deterministic

    Calibrate

    Rule strength sets the severity band; CRITICAL needs two independent sources. Confidence is measured: how often past flags of the same type and severity were confirmed in a walk-forward backtest. Live outcomes take over once 10 exist.

    severity + measured hit rate

  4. 04LLM · explain only

    Explain

    The LLM writes the headline, evidence list and reasoning for each confirmed candidate. It cannot add, remove, retype or re-grade a signal; it can only flag doubt.

    analyst-language explanation

  5. 05Deterministic

    Score

    Alpha score from severity, source credibility, corroboration, recency and liquidity, with the measured move that followed similar flags and an evidence-quality breakdown.

    0–100 alpha, direction, review flags

  6. 06Deterministic

    Interact

    17 compound-event rules: credit risk followed by distress escalates; insider selling before a miss is flagged, buying into a beat is not. Boosts re-check the human-review threshold.

    compound events, systemic risk level

  7. 07LLM · explain only

    Synthesise

    The LLM drafts the executive summary. It cannot rate the brief below its most severe signal. If the LLM is down, the brief is built from the findings and flagged.

    the brief

  8. 08Deterministic

    Remember

    Warehouse, company risk memory, entity graph and the hash-chained audit log. Re-runs do not recount the same event.

    history for backtests, replay and monitoring

06Scoring

Every score decomposes into parts you can check.

The alpha score ranks how actionable a signal is. The evidence-quality breakdown shows how well supported it is. The move shown with each signal is what actually followed similar flags, with its sample size.

Alpha score, 0 to 100

Weights
  • Severity35%

    rule-strength band

  • Source credibility25%

    SEC and exchanges 1.0, Reuters 0.95

  • Corroboration20%

    independent sources

  • Recency10%

    14-day half-life from publication

  • Liquidity10%

    live market cap in USD

The weights are design choices, not fitted values. The backtest tests whether a higher alpha means more confirmed events: so far confirmations rise from 8% in the lowest third of scores to 24% and 36% in the highest. Scores of 70 or more, and every CRITICAL signal, require human review.

Evidence quality

Intel · Form 4 · live run

67%

  • Source1.00
  • Filing strength0.88
  • Historical precision0.53
  • Entities0.55
  • Corroboration0.47
  • Extraction0.65
  • Recency0.35

Reasoning shown in the brief: “Confidence reduced by: single-source (1 corroboration), signal age reducing temporal relevance.”

08Built for review

Made for teams that have to defend the call.

Compliance, risk and investment committees ask the same question: how do you know? Every part of the system is designed to answer it.

Hash-chained audit log

Every pipeline step is written to an append-only log. Each entry hashes all of its stored fields and links to the previous one, so edited and deleted entries are detected.

$ python main.py audit --verify

Compliance mode

Suppresses low-confidence signals before the LLM sees them and records what was suppressed. Evidence remains linked to its source; attribution doubts and signals requiring further review are explicitly flagged.

$ analyse --compliance

Deterministic replay

Re-runs a stored analysis as of its original date through the same rules, calibration and compound-event logic, and records the feedback data it calibrated with.

$ replay --company Intel --date 2026-08-01

Alerts on change, not on schedule

Monitoring alerts on a new event or an escalation, records whether the webhook was delivered, and never re-counts an event it has already seen.

$ monitor --run --interval 3600

Backtests against real outcomes

Outcomes are labelled automatically: the move net of a market benchmark, and whether a later SEC filing confirmed the event. Each prediction is scored once, with precision, recall against SEC-confirmed events, and hit rate.

$ outcomes --auto-label

Portfolio risk that adds up

Fractional weights are normalised, holdings without a risk profile are reported as coverage rather than counted as zero risk, and contagion pairs are surfaced.

$ portfolio --name fund --risk