Three layers. Three audiences. One stack.
OloLand is the depth layer for Anthropic-powered finance work — the verifier stack and persistent system of record that turns Claude’s generic agent runtime into IC-defensible private-equity and M&A underwriting.
An IC-ready memo.
Not search results. Not a chat transcript. The terminal artifact your partner names by hand at the meeting — with source-linked numbers, every assumption ledgered, and citations defensible in front of LPs.
A verifiable intelligence layer.
Hebbia and AlphaSense are intelligence layers — better search over your data room. Verifiable intelligence is a different category: inputs reconciled against the source hierarchy (CPA > tax > management > AI), outputs cited to source pages, conclusions blocked from the memo when evidence is missing.
Deterministic engines and a persistent deal record.
Claude Cowork can read documents and call tools. It cannot run a real DCF against typed financial values, run the Beneish/Benford/EBITDA-bridge forensic battery, or remember what your associate flagged on deal #14 when you start deal #201. OloLand ships both — and they compose with the Claude agents you already use.
Three Teams. One Conclusion.
Karpathy AutoResearch
700 autonomous experiments improved eval accuracy by 11%. No model was changed. The harness did all the work.
LangChain Terminal Bench
Structured tool orchestration moved agentic coding scores from 52.8% to 66.5%. Same underlying models.
M&A Agentic Benchmark
OloLand scored 124/125 vs. Claude Pro at 114/125 on the same Sonnet 4.6 model. Infrastructure changes shipped in 24 hours.
Anthropic Ships Breadth. OloLand Ships Depth.
We’re not a competitor to Anthropic’s finance plugins. We’re the deterministic, persistent, citation-enforced depth layer that runs underneath them. Both compose. The PE associate already running Anthropic’s private-equity plugin invokes ours when the IC pushes back on the number.
The workflow surface
- Five vertical plugins (private-equity, financial-analysis, investment-banking, equity-research, wealth-management)
- Ten named finance agents (Pitch Agent, Valuation Reviewer, GL Reconciler, KYC Screener, Earnings Reviewer…)
- Eleven read-only data connectors (Daloopa, Morningstar, S&P Kensho, FactSet, Moody’s, PitchBook, LSEG, Aiera, Chronograph, Egnyte, MT Newswires)
- Microsoft 365 add-ins for Excel, PowerPoint, Word, Outlook
- hooks/hooks.json shipped as [] across all five vertical plugins
Underwriting-grade computation
- Deterministic DCF, LBO, Monte Carlo, Real Options, Comps engines (strict unit enforcement)
- Forensic QoE primitives: Beneish, Benford, EBITDA bridge, journal-entry, lapping detection, covenant cascade
- Structured risk taxonomy (300+ leaf-level checks) with industry overlays + ensemble scoring
- Cross-document reconciliation (CPA > tax > structured-feed > management > AI hierarchy)
- Middle-office assumption controls + server-side IC approval gate + approval evidence snapshot (new, May 2026)
- Cross-deal flywheel: similar deals, outcome calibration, analyst-correction retraining
- MaskablePPO war-game RL engine (16-quarter competitive simulation)
- Three OloLand plugins: ololand-dd, ololand-forensic-qoe, ololand-compliance-hooks
The Underwriting Control System
Each pillar is necessary. None is sufficient alone. Together, they are the verifier stack and persistent system of record that lives across the deal portfolio — not the conversation. Claude captures the session. OloLand captures the institution.
System of Record
Deterministic computation, structured extraction, provenance chain, analytical chain, workflow orchestration, and institutional-grade security.
Underwriting Controls
Tracked assumption ledger, evidence-quality classification (none/weak/partial/supported), server-side IC approval gate, and approval evidence snapshot for line-by-line audit defensibility. Shipped May 2026.
Institutional Memory
Deal outcomes, cross-deal patterns, analyst corrections, calibration tracking, and firm-specific playbooks that compound with every deal.
Predictive AI
War game simulations, correlated Monte Carlo distributions, and autonomous experiment loops that turn probabilistic reasoning into quantified decision support.
The Fusion Table
A language model has the right column. OloLand has both.
| Domain | Ground Truth (Structured) | Probabilistic (LLM) | Why Both |
|---|---|---|---|
| Financials | DCF engine computes EV | LLM extracts assumptions from 10-K | LLM cannot compute; engine cannot read |
| Risk | Structured risk taxonomy (300+ checks) quantifies | LLM identifies risks from documents | Taxonomy without identification is empty |
| Compliance | OFAC/CFIUS/HSR gates screen | LLM contextualizes regulatory exposure | Gates catch deterministic violations |
| Provenance | Citation anchors (page/cell) | LLM produces grounded analysis | Analysis without anchors is unverifiable |
| Valuation | Monte Carlo simulates distributions | LLM parameterizes from DD findings | Parameters without simulation are point estimates |
| Decision | Playbook rule engine fires verdicts | LLM synthesizes IC narrative | Rules without narrative are unexplained |
M&A Agentic Benchmark: AI vs AI
Real M&A due diligence deal (SigmaTron International take-private, FY2025 10-K). Identical prompts, no coaching, no re-prompting.
| System | Q1 (/25) | Q2 (/25) | Q3 (/25) | Q4 (/25) | Q5 (/25) | Total |
|---|---|---|---|---|---|---|
| ChatGPT Pro (GPT-4.1) | 17 | 18 | 15 | -- | -- | 50/75* |
| Claude Pro (Sonnet 4.6) | 21 | 23 | 23 | 24 | 23 | 114/125 |
| OloLand (Sonnet 4.6) | 24 | 25 | 25 | 25 | 25 | 124/125 |
*ChatGPT session expired before Q4-Q5. GPT-4.1 was the active model at time of evaluation. This is a reproducible eval — not a formal benchmark — and we invite independent replication.
General-Purpose AI vs. OloLand Harness
| Capability | General-Purpose AI | OloLand Harness |
|---|---|---|
| Frozen metrics | None | DCF/LBO/Monte Carlo deterministic engines |
| Persistent state | Lost on session close | Full deal lifecycle in PostgreSQL |
| Cross-deal learning | None | Outcome database + calibration |
| Auditability | None | Page/cell-level provenance chain |
| Risk quantification | Essays | 300+ leaf-level checks, dollar impacts, correlation |
| Analytical chain | Independent essays | Q3 risk -> Q4 WACC -> Q5 coverage ratios |
| Security | Consumer-grade | SOC 2 Type II (in progress), TEE-ready |
| Compliance screening | None | OFAC, CFIUS, HSR automated gates |
| Scenario simulation | Three point estimates | Monte Carlo with correlated distributions |
| Strategic modeling | Strategy essays | War game with 10,000+ simulations |
13 Points in 24 Hours. Zero Model Upgrades.
Every improvement came from harness changes — the model was identical throughout.
Initial eval (risk prefetch broken)
2026-03-12
Risk data prefetch fix
2026-03-12
Risk category expansion + correlation logic
2026-03-13