From raw filing to grounded verdict in one pipeline.
Every output is deterministic first, LLM-augmented second. No hallucinations allowed.
Ingest
Continuous BSE/NSE filing pull — PDFs, XMLs, exchange notices. Every document is parsed, chunked, and indexed into pgvector for retrieval.
Signal
14 deterministic heuristics run first: auditor churn, promoter pledge spikes, RPT ratios, earnings quality, dilution tracking, options anomalies, peer divergence.
Investigate
14 specialist AI agents each examine one forensic dimension. They read actual paragraphs from primary filings — not summaries, not pre-processed data.
Cross-check
A DEVILS_ADVOCATE agent specifically tries to refute every finding. Claims that don't survive adversarial scrutiny are discarded before the memo is written.
Verdict
SCRIBE synthesises all surviving signals into a final memo. Every claim gets a paragraph-level citation with filing date and source document reference.
Before any LLM runs, these hard-coded heuristics scan every filing. Results are fully reproducible.
Auditor Churn
Promoter Pledges
RPT Ratios
Earnings Quality
Dilution Tracking
Options Anomalies
Peer Divergence
Cash Flow Divergence
Board Composition
Insider Trading Patterns
Concall Sentiment Drift
Disclosure Lag
Entity Network Mapping
Cross-Filing Inconsistency
Every verdict is grounded. Every paragraph cited. Every score reproducible.
The pipeline is designed so that no claim survives without a traceable source. The DEVILS_ADVOCATE stage actively tries to invalidate findings before they reach the final memo. If it survives challenge, it earns its place.
BlackLedger outputs are AI-generated forensic research for informational purposes only and do not constitute investment advice. SEBI Research Analyst registered · INH000018432