ololand-forensic-qoe-forensic-screen

Codex wrapper for /forensic-screen

Follow the OloLand command instructions below when the user asks for /forensic-screen or the equivalent workflow in Codex.

Pre-LOI Forensic Screen

Runs the full forensic-QoE battery on a deal and produces an IC-defensible exclusion schedule with severity-scored findings, dollar impact, and source citations. This is OloLand's wedge product — the deterministic statistical layer of QoE that Big-4 also runs (then layers fieldwork on top of) for 20-50x the price.

Usage

/forensic-screen <deal_id>

Arguments

• deal_id (required) — The deal to screen. The deal must have at least: an audited financial statement OR tax return, plus management projections. GL exports and AR aging are required for journal-entry testing and lapping detection.

Execution

1. Call analyze_forensic_qoe from the MCP server with the deal_id.
2. The engine runs every primitive whose required inputs are present:
   • Beneish M-Score — earnings-manipulation probability, private-company adjusted
   • Benford's Law — first-digit anomaly testing on GL transactions
   • EBITDA bridge — adjustment classifier (one-time / pro-forma / questionable)
   • Journal-entry testing — period-end concentration, round-number anomalies
   • Lapping detection — AR cycle anomalies indicating receivables fraud
   • Working-capital deep dive — DSO/DPO/DIO trend + quality scoring
   • Revenue quality deep dive — concentration, hockey-stick, cut-off testing
3. Each finding includes severity (low/medium/high/critical), dollar impact estimate, and a citation back to the source document and page.
4. Classify each primitive output as gap vs finding before composing the analyst-facing summary. This is the single most important narrative step — and the one the Project Atlas Claude memo got wrong by presenting Not computed Beneish/Benford results as if the engine had concluded "no anomalies." Use this taxonomy verbatim:
   • finding — the primitive ran end-to-end on adequate input data and returned a quantitative result (M-Score = -1.42, Benford χ² = 47.3, lapping rate = 3.2%). The result is IC-evidence: pass/warning/fail, cited, can be argued.
   • gap — the primitive returned a status in {insufficient_data, insufficient_sample, not_reliable, unavailable, not_computed}. This is NOT a clean bill of health and MUST NOT be presented as one. It is a diligence ask — name the missing data class (e.g. "GL extract ≥30 line items", "two comparable annual periods", "AR sub-ledger with customer aging") and surface it as a gating condition. The narrator MUST label each primitive at the top of its section with [finding] or [gap]. Memos that pattern-match Result: Not computed → Implication: cannot rely on M-Score are correct (gap framing); memos that pattern-match Beneish M-Score: -2.0 (low likelihood) when status was insufficient_data are wrong (false-clean framing).
5. Output is the structured exclusion schedule — what gets excluded from headline EBITDA, what gets flagged for management Q&A, what kills the bid. Gaps (from step 4) appear in the Open Questions section, NOT the exclusion schedule.