The Memory Layer
This is the architectural gap that compounds into an uncrossable moat.
Deal Outcome Database
Twenty-two pre-close prediction fields (EV, IRR, MOIC, WACC, exit multiple, identified risks, risk-adjusted EV) snapshot at DD completion. Eighteen post-close actuals captured at exit. Ten learning-signal columns track prediction accuracy by deal, sector, and risk category. Update outcomes by chat — "we passed because price went too high" — and Haiku extracts the structured fields. Over time: "we classified 20 similar situations as high risk, and 14 had material problems — a 70% hit rate."
Cross-Deal Pattern Matching
When analyzing a new deal, the system retrieves the most structurally similar completed deals using multi-dimensional similarity scoring (industry, size, deal type, risk profile, financial characteristics). A manufacturing take-private with customer concentration and covenant risk triggers retrieval of prior deals with the same profile — and their outcomes.
Analyst Correction Tracking
Every time a human analyst overrides, corrects, or adjusts a system output, the correction is categorized by risk category, deal type, and analyst. These corrections feed back into the system's prompts, tool configurations, and extraction parameters. The system learns from structured feedback loops, not model retraining.
Calibration Dashboard
Prediction accuracy tracked over time across deal types, risk categories, and confidence levels. A firm can see whether the system's "high confidence" predictions are actually more accurate than "medium confidence" ones — and by how much. Measurable improvement, not claimed improvement.
Flywheel Context Builder
Before any specialist agent begins analysis, the flywheel assembles institutional context: relevant prior deals, firm-specific risk tolerances, historical corrections for this deal type, calibration data for similar situations. The agent starts with institutional knowledge, not a blank slate.
Playbook Rule Engine
Firm-specific investment criteria encoded as executable logic. Sector exclusions, minimum EBITDA thresholds, maximum leverage limits, geographic restrictions — all run as a Pass 0 hard-reject gate before any AI analysis begins. This is not "the model thinks the deal doesn't fit." This is "the deal fails the firm's own encoded criteria."
From correction to next month’s grader.
The capabilities above are exposed by a five-stage engineering pipeline. Capture layers are running today; training layers fire as trajectories cross training-size — first monthly cycle June 2026.
Legacy ADK retired
Single canonical agent runtime — no parallel orchestrators drifting against each other.
Analyst correction capture
Inline correction triggers on every citation, both in the web UI and the Cowork plugin. Structured corrections persist to `analyst_corrections` with `metadata_` JSONB stamped by surface.
Nightly trajectory exporter
Each grader-stamped run becomes a JSONL training trajectory on GCS (`gs://ololand-training-trajectories-{env}`). Risk-extractor and verifier dataset builders consume the same persistent run log — the audit trail is the training corpus.
Per-tenant LoRA training
Monthly Vertex SFT job over a firm-scoped trajectory window. Your corrections train your firm’s adapter; nothing pools without an explicit opt-in. First retrain fires June 2026 once trajectories cross training-size.
Anonymized-pooled cross-tenant
Cohort-tuned base adapter for firms that opt in to the shared pool — k-anonymity audit on every record. Opt-in only, gated to the Dev/Pro tier as the price-of-entry exchange.
The model is the analyst. OloLand is the underwriting control system — and every analyst correction becomes a training trajectory.
Defensibility That Scales With Time, Not Funding
Time Dependency
The outcome database requires deals to close (or fail) and outcomes to be observed. A 12-month head start means 12 months of outcomes a competitor cannot replicate by hiring engineers.
Experience Dependency
Cross-deal patterns require N deals of sufficient diversity. The 50th manufacturing take-private teaches the system things the 5th could not.
N-of-1 Personalization
Every firm's playbook, correction patterns, and risk tolerances are unique. The system calibrates to each firm individually. A competitor would need to re-learn each firm from scratch.
Switching Cost Escalation
After 100 deals, the institutional memory is deeply embedded. Switching to a competitor means abandoning the calibration curve and starting fresh.
Super-Linear Utility
The 200th deal is more valuable than the 20th — not linearly, but super-linearly — because cross-deal patterns emerge at scale that are invisible at low volumes.
No Competitor Has Institutional Memory as an Architectural Primitive
Hebbia
Strength: 1M+ daily queries, $500K average contracts
Gap: Does not track what happened after the deal closed
Datasite
Strength: Owns the virtual data room — the source documents
Gap: Does not track deal outcomes or calibrate predictions
Harvey
Strength: Legal DD for PwC across thousands of matters
Gap: No cross-deal calibration layer
ChatGPT / Claude
Strength: Frontier reasoning capability
Gap: Stateless by design — every conversation is independent
The Venture Community Has Converged on This Thesis
Foundation Capital
"Context Graphs" — the next trillion-dollar platforms are systems of record for decisions, not objects. Decision traces become training data; overrides become case studies.
Greylock
"New New Moats" — systems of intelligence that become their own system of record represent the highest-defensibility category.
Alpha-Matica
"Data Moat 2.0" — the new defensible moat is a living, proprietary feedback loop that continuously captures, refines, and redeploys human judgment at scale.
NFX
50%+ of VCs identify "quality or rarity of proprietary data" as the factor most likely to create durable moats.
OloLand is building exactly this: a living feedback loop where every deal, every correction, and every outcome makes the next analysis more accurate.