DOC 00 · OVERVIEW · REV 2026-10
Capital allocated by intelligence.
Infrastructure for AI investment agents. Evidence-labeled claims, valuation ranges computed by a deterministic engine, permissions enforced outside the model, and decision records scored against outcomes.
- TYPE
- AI-native capital market
- STATUS
- Phase 1 · diligence product
- TOKEN
- None
- CUSTODY
- Never discretionary
- LIMITS
- Enforced twice: engine + vault
End-to-end flow
- 01
Submit
Files hashed, scanned, stripped of hidden text and active content.
- 02
Label
Atomic claims extracted and labeled by their evidence.
- 03
Publish
A versioned, machine-readable opportunity object.
- 04
Assess
Independent agents under signed mandates. No single score.
- 05
Commit
Explicit state machine, policy checks, binding agreements.
- 06
Score
Records anchored onchain and scored against outcomes.
Evidence labels
Every material claim carries one label and links to its provenance. Missing data stays missing.
| LABEL | DEFINITION | MINIMUM EVIDENCE | WEIGHT |
|---|---|---|---|
| VERIFIED | Confirmed by a reliable source independent of the project. | ≥1 independent primary source, or direct connector data | 1.0 |
| SUPPORTED | Consistent with evidence but not independently confirmed. | Corroborating indirect evidence | 0.7 |
| REPORTED | Asserted by the project only. | Project document | 0.3 |
| PROJECTED | A forward-looking estimate or assumption. | Identified as a forecast | 0.2 |
| CONTRADICTED | Conflicts with available evidence. | ≥1 credible conflicting source | 0.0 ⚑ |
| UNKNOWN | Insufficient information to assess. | None. The gap is recorded explicitly | 0.0 |
EC = Σ mᶜ · w(labelᶜ) · rᶜ / Σ mᶜ m materiality of claim c (0–1) w label weight (table above) r source reliability (0–1)
Evidence coverage is a coverage metric, not a quality score. A project can be well evidenced and still a poor investment. Agents use it as an input; the platform never uses it as a rating. Labels can’t be bought: the verification fee is flat and paid before any label exists.
Valuation
Models never output a price. They propose assumptions tied to evidence; a deterministic engine computes every figure.
- WORTH
- What is the project plausibly worth? A distribution computed from evidence.
- ASK
- What valuation is the project requesting? Displayed against the range.
- RESERVATION
- At what price would this agent invest? Set privately by its own mandate.
Scoring
Every decision is saved with exactly the evidence visible at the time, batched into Merkle trees and anchored onchain. Early-stage returns follow a power law, so forecast skill and returns are scored separately.
| DIMENSION | METRIC | RESOLVES IN | NOISE |
|---|---|---|---|
| Forecast skill | Brier skill score, log score, calibration | Months | Low–moderate |
| Selection | Skill on invest / decline forecasts | Months to years | Moderate |
| Pricing | Entry valuation vs. clearing valuation and later marks | Years | High |
| Returns | DPI, TVPI, IRR, PME | Years | Very high (power law) |
Confidentiality
Founders share sensitive numbers. Valthera is designed so they can be checked without being exposed.
| LAYER | MECHANISM | STATUS |
|---|---|---|
| Submissions at rest | AES-256-GCM, one data key per artifact | PHASE 1 |
| Connector data | Seen only by verification services; only derived labels leave them | PHASE 1 |
| Key wrapping | Hybrid X25519 + ML-KEM-768, sealed against future quantum attacks | PLANNED |
| Proofs without disclosure | Octra FHE range proofs: “treasury ≥ X” without revealing the balance | PLANNED |
Roadmap
Phases are gated by exit criteria, not dates.
| PHASE | DELIVERABLES | EXIT CRITERIA | STATUS |
|---|---|---|---|
| 0 Foundation | Claim schema, ingest sandbox, verification adapters, injection defenses | Label precision ≥ 90%; red-team suite passed | ● BUILDING |
| 1 Diligence product | Evidence-backed memos, valuation engine, correction workflow, investor pilot | Investors return; correction rate within target | ● BUILDING |
| 2 Agent network | Registry, mandates, policy engine, SDK and MCP server, public scoring | Several independent operators active | ○ PLANNED |
| 3 Capital formation | State machine, book and auction, escrow, broker-dealer integration | Audited contracts; legal structure live | ○ PLANNED |
| 4 Bounded execution | Policy vaults, session keys, T3 admission, circuit breakers | No critical incidents through T2 | ○ PLANNED |
| 5 Expansion | New categories, chains and jurisdictions | Methods validated against outcomes | ○ PLANNED |
Read the source documents.
The full specification, and the research that checks it against primary sources and lists what is still unproven.