Anthropic shipped generation. We shipped verification.
Anthropic’s 2026 finance lineup validated the category — and explicitly defined its own scope. The space they left for someone else is the verification layer the IC defends a number against. That’s where OloLand lives.
Anthropic doesn’t claim to be the system of record.
Their finance lineup is framed as analyst tooling — drafts produced for review by a qualified professional. That framing rules out the buyer who has to defend a number to an LP, a credit committee, or a regulator. That buyer needs a persistent deal record, role-gated replay, and a verifier stack underneath. That buyer is the OloLand ICP.
The verification stack is shipped code.
Closed-loop analyst corrections (typed AGENT_CLAIM_CORRECTION events captured as firm-owned learning evidence; retraining fails closed until authorized). Role-gated replay of harness runs that snapshots root prompt, skill pack, subagent definitions, and the exact model provider + version per run — a 2027 regulator can reconstruct the 2026 inference. Continuous eval: a nightly judge sampler with Sentry-alerted drift detection, plus an hourly LLM-as-judge backfill so verifier signal arrives before human correction volume. Four deterministic verifier tools wired through the lead orchestrator — reconcile_documents, check_citation_coverage (semantic, not bracket-counting), run_beneish, run_lbo_model. Per-deal cross-session memory (remember_deal_fact, recall_deal_facts) so the associate’s flag on deal #14 survives into the next session. All in production today.
Same Cowork session. We compose with Anthropic.
Anthropic’s finance plugins provide analyst-seat workflows. OloLand’s Pre-LOI Forensic Screen can run from the same Cowork session. Full QoE is a governed report capability included in Pro and above and starts in the OloLand deal workspace, where OloLand preserves the evidence, calculations, and approvals in the acquisition record.
Three plugins. One Cowork session. End-to-end deal lifecycle.
A deal-lifecycle core, a forensic QoE wedge that fits pre-LOI, and a compliance layer that audits every action. Quantitative claims are bound to deterministic engines and cited sources — ready for partner sign-off, RWI underwriters, and regulator review.
ololand-ddDeal Lifecycle & IC Memo
From pre-screen to IC, with institutional memory.
- Pre-screen & new-deal: seed any public or private target with 10-Ks and 5-yr financials
- Valuation engines: DCF, LBO, Monte Carlo, comps — risk-adjusted, unit-enforced
- IC memo (skeptical): tile-stitched, with gap-vs-finding framing partners trust
- War-game sim: 16-quarter MaskablePPO competitive dynamics
- Sector packs: SaaS, healthcare, industrial, real estate KPI discipline
- Firm playbook + deal memory: WACC ranges, multiples ceilings, and analyst-confirmed facts persist across sessions
Also includes similar-deals, calibrate-vs-history, verify, ic-approve-readiness, meeting-prep, talk-to-deal, dd-merger-analyze (3rd-party arb / antitrust), plus deal-sourcing with OpenClaw outreach drafting and CRM dedupe.
ololand-forensic-qoePre-LOI Forensic QoE
The Big-4 QoE alternative that fits the LOI window.
- Beneish M-Score (private-company adjusted) — earnings manipulation probability
- Benford’s Law GL first-digit testing — numeric fabrication
- EBITDA bridge — classifies every add-back: one-time, pro-forma, or questionable
- Journal-entry anomalies — period-end, round-number, weekend posts
- Lapping detection — AR-cycle fraud patterns
- Working-capital deep dive — cash conversion & earnings quality
Cited forensic deliverable inside the LOI window. Every flag ties back to a source chunk in the underlying PDF.
ololand-compliance-hooksCompliance & Audit Trail
Defensible by construction.
- MNPI guard (PreToolUse) — blocks material non-public information leakage before any tool call
- Citation enforcer (PostToolUse) — flag or deny any quantitative claim without an inline source marker
- Provenance ledger — every action written to
~/.ololand/provenance/ - Audit log mirrorable to the OloLand audit API for regulator export
- Harness-run replay — re-run OloLand harness runs against their prompt and skill snapshots; managed-agent runs link to their platform session instead
Arms automatically at session start. Set OLOLAND_CITATION_BLOCK=1 to switch the citation enforcer from warn-mode to deny-mode. Set OLOLAND_AGENT_KEY to mirror audit events to the OloLand cloud audit API.
Five layers, cheapest first.
A claim runs the cheapest verifier first and escalates only when it fails. Anthropic Managed Agents and Google Agent Engine ship L1 and parts of L5 — the contract validator and the human approval gate. OloLand owns L2–L4: deterministic engines, LLM graders, and adversarial reviewers, bound to a persistent deal record.
Output contract validation
Malformed JSON · missing citation or confidence fields · type errors
Deterministic verifiers
Arithmetic errors · fabricated citations (existence check) · forensic manipulation · cross-document inconsistency
LLM graders, continuous
Hallucinated claims · unsupported attribution · missing risk categories · weak reasoning · drift over time (nightly judge sampler with Sentry alerts + hourly LLM-as-judge backfill so signal arrives before human correction volume)
Adversarial subagents
Motivated reasoning · optimism bias · stale-state failures · narrative ↔ evidence gaps
Human-in-the-loop, role-gated
Novel risks · judgment calls · calibration drift · anything requiring institutional context
Foundation-model platforms ship the cheapest and most expensive layers. OloLand owns the load-bearing middle — and binds the whole stack to a persistent, auditable deal record that the foundation-model platforms structurally cannot ship.
Institutional memory, not session memory
The firm playbook + per-deal session memory means WACC ranges, IRR floors, multiple ceilings, and analyst-confirmed facts compound across deals — not lost when the tab closes.
Deterministic numbers, not LLM guesses
Every DCF, LBO, Beneish score, and Benford test runs in a deterministic engine. The model orchestrates — it doesn't invent numbers. Unit-enforced, reproducible, reviewable.
Defensible by construction
Inline [N] citations, four atomic-claim verifiers, source-hierarchy ranking (CPA > tax > mgmt > AI), and full span-tree replay. Built for partner review, RWI underwriters, and SEC/FCA defense.
Anthropic ships the workflow surface.
Five vertical plugins (private-equity, financial-analysis, investment-banking, equity-research, wealth-management), ten named agents, eleven read-only data connectors. They cover the writing of memos and decks beautifully.
OloLand ships the deterministic computation.
~13 financial engines, 7 forensic primitives, a structured risk taxonomy (300+ leaf-level checks), a structured deal graph, and a 45-tool MCP server — all in production. Anthropic’s plugins call out to them.
They compose. They don’t compete.
Same Cowork session. Anthropic handles the workflow; OloLand handles the math, the structured taxonomy, the persistent memory, and the provenance. A PE associate already running Anthropic’s private-equity plugin invokes OloLand when the IC pushes back on the number.
OloLand fills the empty hooks/.
Anthropic’s vertical finance plugins ship hooks/hooks.json as []. No compliance hook, no citation enforcement, no audit log. ololand-compliance-hooks drops in as the regulated-workflow layer.
Anthropic’s ten finance agents — with an honest verdict for each
Three verdicts. No marketing fudge. For each agent in Anthropic’s May 2026 finance launch we tell you exactly one of: OloLand ships stricter, outside our scope (use Anthropic’s first-party), or we don’t ship that, here’s why we shouldn’t.
Anthropic drafts slide content. OloLand generates a 14-section CIM where multiples, EBITDA figures, and comps tie to source pages — defensible at IC.
mcp__ololand__talk_to_deal (live deal memory + risk surface)Anthropic preps from public docs. OloLand opens the persistent deal record so the meeting touches the open assumptions, the unresolved risks, and the analyst corrections from prior turns.
deep_market_research + deep_precedent_research (multi-hop with KG)Anthropic does single-pass web research. OloLand multi-hops through the knowledge graph and reconciles comparables against the cross-deal database, so precedents are ranked, not just listed.
Anthropic surfaces narrative highlights. OloLand runs the deterministic forensic battery — Beneish M-Score, Benford on the GL, EBITDA bridge classifier — with $-quantified impact and a severity score.
Anthropic prompt-drives Excel. OloLand runs typed financial values through deterministic engines with strict unit enforcement — kills the entire thousands-vs-millions class of bugs that prompts cannot.
Anthropic checks marks against firm policy. OloLand reconciles values across the source hierarchy (CPA audited > tax return > management model > AI extracted) and blocks the IC memo when evidence is missing.
GL reconciliation is the controller’s office, not pre-close diligence. Use Anthropic’s first-party agent here. OloLand stays in the underwriting lane on purpose.
Month-end close is operational accounting. OloLand is buy-side M&A underwriting and pre-LOI forensic screening. Use Anthropic’s first-party agent here.
Anthropic flags anomalies. OloLand runs the named forensic primitives a CPA reviewer would name on cross-examination — lapping cycle, journal-entry stratification, working-capital decomposition — each with a severity threshold and a $-impact estimate.
KYC is regulatory identity verification of a counterparty. That’s a different category from buy-side diligence. OloLand does NOT ship KYC and shouldn’t — use Anthropic’s here. The adjacent forensic-screener answers a different question (“is this target’s P&L manipulated?”).
We picked “outside our scope” or “we don’t ship that” on three of ten on purpose. The wedge is buy-side underwriting; the controller’s office and counterparty KYC are real categories — just not ours. Use Anthropic’s first-party there.
OloLand vs. status quo
Same partner, same deal, same week. Where the work moves when the verification layer is shipped code.
Capability matrix
What each side ships out of the box. Anthropic’s vertical plugins are markdown templates over read-only data; OloLand is deterministic computation, structured taxonomy, persistent state, and policy hooks.
A typical workflow
Source & screen
Discover targets, dedupe vs. CRM, run pre-LOI screen with bear/base/bull SOTP
Forensic wedge
Pre-LOI screen on management financials — Beneish, Benford, EBITDA bridge
Deep DD
Sector pack + DCF/LBO/Monte Carlo with firm playbook constraints applied
Verify & IC
Atomic-claim verifiers run; skeptical IC memo composed; partner sign-off gate
Replay & archive
Full audit trail; regulator export on demand; outcome metadata feeds the playbook
The empty hooks/hooks.json
Every vertical plugin in Anthropic’s finance lineup — private-equity, financial-analysis, investment-banking, equity-research, wealth-management — ships hooks/hooks.json as []. There is no compliance hook, no citation enforcement, no MNPI guard, no audit log. For a regulated buy-side workflow this is unacceptable: you cannot defend a number to the IC if you cannot prove what ran, what it was sourced from, or that no MNPI leaked into a prompt.
ololand-compliance-hooks is the drop-in plugin that fills this scaffold. PreToolUse MNPI guard. PostToolUse citation enforcer. Provenance writeback to ~/.ololand/provenance/ and the OloLand audit API. Audit log for every mcp__ololand__* call. Composes additively with Anthropic’s plugins and any other hook plugin.
$ claude plugin install ololand-compliance-hooksWhen the partner asks “where did this number come from?”
Same Cowork session. Same partner challenge. Different depth underneath.
“This GP mark looks aggressive versus comps.”
Methodology check against firm policy via the portfolio MCP. The reviewer flags — it does not run the deterministic math, does not compute the variance, does not block IC approval, does not cite the source page.
“This mark implies 14.2× EBITDA, 2.1 turns above the peer median. Monte Carlo 5th-percentile EV is $X. Beneish M-Score is above the −1.78 manipulation-risk threshold. Two high-priority assumptions have no supporting evidence — sponsor margin expansion (+340 bps over 3 years, no comparable precedent) and customer retention (95% assumed, churn ledger shows 87%). IC approval is blocked until both assumptions are resolved.”
Figures tie to source documents, pages, and formulas. The grade gate enforces citation coverage at render time, not at prompt time.
That’s not a better plugin. That’s a different product category — a verifiable intelligence layer, with deterministic engines and a persistent deal record underneath.
Measured on Vals AI Finance Agent v1.1
Stanford + a Global Systemically Important Bank wrote the questions. We ran OloLand’s verifier stack on top of two frontier models from two different labs — Anthropic Claude Opus 4.7 and Google Gemini 3.1 Pro — with the same harness, same tools, same questions, same judge. Three independent seeds per model.
compute_cagr alone (3-seed leave-one-out)OloLand’s biggest contribution is not the accuracy ceiling — it’s the floor. Predictable, defensible outcomes on every run, every model.
See how the loop runs — and where OloLand bolts in
Interactive 15-step walkthrough of the agentic loop: keystroke, first POST, tool_use, subagent fork, cache hit, IC gate. Verifiable at every step.
Anthropic captures the session. OloLand captures the institution.
Same Cowork surface. Same Claude harness. Different depth — and a persistent deal record that lives across the portfolio, not the conversation. Install both. They compose.