Fully Autonomous
Venture Fund

Introduction

A venture fund is, at its core, a firm of minds: people who have each developed a view of a domain, who hold positions, and whose combined judgment is worth more than that of any individual. We asked whether that firm could be implemented in software. Not assisted by software, but actually implemented as software. Fair is the answer we arrived at.

The system is structured as an orchestrating agent that coordinates several specialised sub-agents. Each has a domain, a memory, and a developing view of the world. They hold positions, surface disagreements, and write to a shared memory that compounds over time. The result behaves less like a tool and more like a team: one that doesn't sleep, doesn't forget, and doesn't let quality slip when deal volume spikes.

We chose crypto as the testing environment because the conditions are unusually favourable: access to deals is permissionless, positions are liquid, and feedback loops are fast enough that the system can learn from its own decisions in weeks rather than years. The alpha hypothesis runs deeper than that. At the earliest stages of any market, a significant portion of signal lives in social exhaust, what builders, researchers, and operators surface publicly before it reaches any database. Human analysts can't process this at scale. Agents can.

Agent Roster

One decider, many producers. Sub-agents build the context package. Only Fair issues execution intents.

The Decider

Fair, the MD

THE DECIDER
The single accountable decision-maker. Receives the full decision packet, applies its world model (beliefs, active theses, principles) and emits structured intents. The only agent authorized to propose execution. Thinks, but holds no wallet access.
Intent ThesisUpdate DataRequest
Research Layer: build the context package
🔍

Sourcing Analyst

Producer
Scans 5,000+ monitored accounts every tick to surface new projects before they reach broader awareness. Evaluates novelty and early signal strength, not investment quality. Optimized for high recall at the top of the funnel.
CandidateSet
📑

Principal

Producer
Produces structured theses from raw candidates. Each thesis includes explicit invalidation conditions: specific, falsifiable claims about what must be true for the investment case to hold. Evaluated on whether claims hold post-exit.
ThesisDraft
🌐

Macro

Producer
Synthesizes market reports and on-chain flows into a regime assessment. Outputs a risk-on/off signal, narrative momentum summary, and sector rotation view for context-setting in the decision packet.
RegimeAssessment
📡

Portfolio Monitor

Producer
Tracks every project on the watchlist and in the portfolio. Detects momentum shifts, thesis invalidation events, key-person changes, and material developments. Generates deltas for Fair's review each cycle.
WatchlistDeltas
📊

Market Strategist

Producer
Ranks narratives by phase (emerging, hot, crowded, fading) across the market. Detects rotation early by monitoring cross-community signal patterns and the pace of adoption among high-signal accounts.
NarrativeCards
Deterministic Layer: no LLM in the loop
🛡️

Policy Gate

Deterministic
Validates every intent against hard rules before any capital moves: position limits, portfolio concentration, drawdown breakers, liquidity requirements. Returns APPROVED, RESIZED, VETOED, or EMERGENCY with a reason code. No language model participates in this decision. It is fully deterministic. Survival constraints override everything.
PolicyVerdict APPROVED / RESIZED / VETOED
🔄

Executor + Reconciler

Deterministic
Translates policy-approved intents into market actions. Confirms fills, updates portfolio state in the ledger, and emits reconciliation results. Paper mode and live mode share an identical interface, switched by configuration, not code change.
ExecutionReport ReconciliationResult

The Knowledge Model

Every object in the system has a defined type with explicit mutation rules. This removes ambiguity about what "state" means.

Fair operates on a formal knowledge typology: nine distinct object types, each with defined mutability and provenance requirements. Raw observations enter as Receipts and are progressively transformed through normalization, scoring, interpretation, and structured decision-making into Executions and Outcomes. This ladder is not abstract. It is the schema that governs every object in the system's ledger.

In practical terms, this means that at any point in time, we can reconstruct exactly what the system knew, what it had access to, and why it made the decisions it made. Every step from raw observation to capital deployment is traceable, auditable, and reproducible.

Object Typology: Knowledge Ladder
Receipt
append-only
Fact
append-only
Derived Signal
recomputable
Belief
versioned
Thesis
versioned
Conviction
derived
Intent
immutable
Execution
immutable
Outcome
append-only
Type Definition Mutability Example
Receipt Raw observation from a source. No interpretation whatsoever. append-only Tweet, price tick, on-chain event, group chat message
Fact Normalized, validated receipt. A statement of what happened, without inference. append-only "Project X launched at 50K market cap on Solana"
Derived Signal Deterministic transform of one or more facts. Tagged with function version and source receipts for full reproducibility. recomputable Momentum score, volume z-score, social velocity index
Belief Interpretation about the state of the world. Narrative phase, founder quality, macro regime. versioned "AI agent infrastructure narrative is in the 'hot' phase"
Thesis Actionable claim about a specific project with explicit invalidation conditions. Always falsifiable. versioned "Project X will reach 5x because Y. Invalid if Z occurs."
Conviction Scalar derived from belief strength and thesis quality. Maps directly to position sizing. derived 0.8 conviction → 3% portfolio allocation
Intent Structured decision emitted by Fair. Specifies action, instrument, sizing, take-profit, stop-loss. immutable BUY 2% of NAV, TP at 3x, SL at -50%
Execution Result of a policy-approved intent. Confirmed fill, partial, or rejection with reason. immutable Filled 1.2 SOL at $54K market cap, tx: 0x...
Outcome Realized P&L and post-investment state. Closes the feedback loop for all three evaluation targets. append-only +90% realized, thesis held through exit, no thesis drift

Two-Layer State: Ledger and Snapshot

State lives in the harness, not in any agent. The ledger is truth; the snapshot is a performance-optimized projection of it.

Ledger (append-only)

The immutable record of everything that happened. Nothing is edited. Corrections are new events. The complete audit trail.
  • Every receipt and fact ingested
  • Every artifact emitted by producers
  • Every decision packet assembled as_of timestamp
  • Every intent from the decider
  • Every policy verdict with reason code
  • Every execution report and fill
  • Every reconciliation result
  • Every outcome and P&L attribution

Snapshot (current view)

A cached projection rebuilt from the ledger. What agents read for performance. Never the source of truth, always derivable from the ledger.
  • Portfolio: positions, P&L, exposure, risk budget
  • Beliefs: narrative assessments, regime view (versioned)
  • Theses: active theses with invalidation conditions
  • Watchlist: tracked projects and metadata
  • CandidateSets: latest output from Sourcing Analyst
  • NarrativeCards: ranked narratives by phase
  • RegimeAssessment: latest macro brief
  • Derived Signals: with function_version provenance

How Fair Works

The executive loop runs on an hourly cadence and interrupts immediately on material events.

System Architecture
flowchart TD
  subgraph SOURCES["DATA SOURCES"]
    direction LR
    S1["X / Twitter"]
    S2["Telegram"]
    S3["Farcaster"]
    S4["On-chain"]
    S5["Market Data"]
    S6["Group Chats"]
  end

  subgraph INGEST["INGEST ADAPTERS"]
    direction LR
    IA1["Social"]
    IA2["Chain"]
    IA3["Market"]
  end

  subgraph PROD["ARTIFACT PRODUCERS"]
    direction LR
    P1["Sourcing Analyst"]
    P2["Principal"]
    P3["Macro"]
    P4["Monitor"]
    P5["Narrative"]
  end

  subgraph ROUTING["PERCEPTION"]
    AR["Attention Router"]
    DP["Decision Packet"]
    AR --> DP
  end

  FAIR["Fair, the MD"]
  POLICY["Policy Gate"]

  subgraph EXEC["EXECUTION"]
    direction LR
    EX["Executor"]
    RC["Reconciler"]
    EX --> RC
  end

  LEDGER["Ledger"]
  SNAP["Snapshot"]

  SOURCES --> INGEST
  INGEST -->|"receipts"| PROD
  PROD -->|"artifacts"| AR
  SNAP -->|"beliefs + portfolio"| DP
  DP -->|"Decision Packet"| FAIR
  FAIR -->|"Intent"| POLICY
  POLICY -->|"Approved"| EXEC
  RC -->|"events"| LEDGER
  FAIR -->|"events"| LEDGER
  PROD -->|"events"| LEDGER
  LEDGER -->|"materialise"| SNAP
  SNAP -.->|"read"| PROD

  style FAIR fill:#fffbeb,stroke:#ea580c,stroke-width:2.5px,color:#0c1024
  style POLICY fill:#fef2f2,stroke:#dc2626,stroke-width:2px,color:#0c1024
  style LEDGER fill:#eff6ff,stroke:#2563eb,stroke-width:2px,color:#0c1024
  style SNAP fill:#eff6ff,stroke:#2563eb,color:#0c1024
      
1

Perceive

Ingest adapters pull receipts from all sources. Producers transform facts into typed artifacts: CandidateSet, ThesisDraft, RegimeAssessment, WatchlistDeltas, NarrativeCards. Each artifact is logged to the ledger before it is consumed.

2

Assemble Packet

The Attention Router surfaces what matters this cycle. The Decision Packet assembles the full context: current beliefs, portfolio state, fresh artifacts, pinned to an as_of timestamp. The packet is itself logged: Fair's inputs are always reconstructible.

3

Decide

Fair (the MD) receives the packet and applies its world model. Emits structured intents: buy, sell, add to watchlist, update thesis, request more data. Every intent is logged immutably before it proceeds to the gate, regardless of whether it is approved.

4

Gate

The Policy Gate validates each intent against hard rules: position limits, drawdown breakers, liquidity thresholds, concentration limits. Returns APPROVED, RESIZED, VETOED, or EMERGENCY with a reason code. Deterministic. No language model is involved at this stage.

5

Execute + Reconcile

Approved intents are executed, paper or live, same interface. The Reconciler confirms fills and updates portfolio state in the ledger. An ExecutionReport and ReconciliationResult are emitted and logged for every action taken.

6

Adapt

Outcomes feed three independent learning targets: routing quality, thesis quality, and execution quality. Each operates at a different feedback speed. Beliefs update. Producer accuracy is re-evaluated. The policy gate is reviewed on a slower cadence. The system improves with each completed cycle.

Scheduling: Cadence + Interrupt Triggers
flowchart LR
  subgraph CADENCE["REGULAR CADENCE"]
    T1["Hourly tick"] --> T2["Assemble packet"] --> T3["Run full loop"]
  end

  subgraph INTERRUPTS["INTERRUPT TRIGGERS"]
    direction TB
    I1["Watchlist spike"]
    I2["Stop / TP hit"]
    I3["Drawdown breach"]
    I4["Thesis invalidated"]
    I5["Liquidity collapse"]
  end

  INTERRUPTS -->|"immediate"| T2

  style T1 fill:#eef2ff,stroke:#4f46e5,color:#0c1024
  style T2 fill:#eef2ff,stroke:#4f46e5,color:#0c1024
  style T3 fill:#eef2ff,stroke:#4f46e5,color:#0c1024
  style I1 fill:#fef2f2,stroke:#dc2626,color:#0c1024
  style I2 fill:#fef2f2,stroke:#dc2626,color:#0c1024
  style I3 fill:#fef2f2,stroke:#dc2626,color:#0c1024
  style I4 fill:#fef2f2,stroke:#dc2626,color:#0c1024
  style I5 fill:#fef2f2,stroke:#dc2626,color:#0c1024
        

The Safeguards

Designed assuming the system can be exploited or make mistakes. The security model does not rely on any single component being correct.

The agents that think have no access to wallets. The agent that has wallet access doesn't think. No single component can independently discover, research, decide on, and execute an investment. An adversary would need to compromise multiple architecturally isolated systems to move funds, and even then, the append-only ledger would record every step.

Deterministic Policy Gate

Every execution intent passes through a rule-based gate before any capital moves. No language model participates in this decision. Position limits, drawdown breakers, liquidity requirements, and concentration limits are checked against hard numerical thresholds. The gate returns a structured verdict with a machine-readable reason code. Survival constraints override every other consideration. There is no mechanism for Fair to bypass this check.

Append-Only Audit Trail

The ledger is immutable. Every receipt, artifact, decision packet, intent, policy verdict, execution, and outcome is written as a new event. Nothing is overwritten, nothing is deleted. Corrections are new events that reference what they correct. This means every decision Fair has ever made is reconstructible from first principles, with the exact inputs it operated on at the exact moment it decided. The system can be audited, backtested, and interrogated at any point in its history.

Architectural Domain Separation

The intelligence layer, decision layer, and execution layer are architecturally isolated. The Sourcing Analyst and Principal have no awareness of portfolio state. The Policy Gate has no access to research. The Executor has no access to theses, social data, or any intelligence output. It only receives structured, policy-approved intents. These boundaries are enforced by the system architecture, not by social convention. There is no single point of failure that bridges all three domains.

Attention Routing Is Separate from Decision Inputs

What the Attention Router surfaces for a given cycle is recorded independently from what the Decision Packet contains. This separation prevents attention heuristics from silently changing Fair's decision inputs. The router can prioritize what to highlight, but it cannot modify the underlying facts, signals, or beliefs that Fair reasons about. Both layers are logged independently.

Interrupt-Driven Risk Response

The normal hourly cadence can be interrupted immediately by critical events: drawdown breach, liquidity collapse, confirmed rug pull or exploit, or thesis invalidation. These interrupts bypass the regular scheduling and trigger an emergency policy review. The escalation tier (CRITICAL, HIGH, MEDIUM, INFO) determines response speed and routing. CRITICAL events produce an immediate interrupt; others are handled within the next regular cycle.

Provenance at Every Layer

All derived signals cite the receipts they were computed from, along with the function version used. All beliefs cite the facts they were inferred from. All theses carry explicit invalidation conditions. This chain of provenance means that any claim the system makes can be traced back to the original observation that grounded it, and any inference that does not survive scrutiny can be identified and corrected without contaminating the rest of the knowledge base.

Non-Negotiable Invariants

  • One decider. Only Fair emits execution intents. Producers generate artifacts, not actions.
  • LLM proposes, policy disposes. Execution is deterministic and auditable. Survival constraints override everything.
  • Ledger is truth, snapshot is convenience. State changes are events. The snapshot is always rebuildable.
  • Attention routing is separate from decision inputs. The router surfaces; it does not alter what is seen.
  • Provenance everywhere. All signals cite receipts, all beliefs cite facts, all theses have explicit invalidation conditions.

The Learning Loop

Three independent evaluation targets. Each part of the system improves on its own signal, at its own speed.

Fair gets better with every investment, not as a metaphor but as a measurable property of the system. Every Outcome closes a feedback loop that flows back through three independent evaluation targets: how well the system identified what was worth paying attention to, how well its theses about individual projects held, and how cleanly the execution layer translated approved intents into positions. These three targets update at different speeds because they carry different signal-to-noise characteristics. Execution quality is immediate; thesis quality requires holding periods to measure; routing quality accumulates across hundreds of sourcing decisions before meaningful patterns emerge.

Routing Quality

Measures: Sourcing Analyst + Attention Router
  • Did surfaced candidates lead to profitable theses?
  • Precision: % of candidates that became positions
  • Recall: % of winning projects that were surfaced early
  • Time-to-awareness: how fast from publication to ingestion
  • Signal-to-noise ratio of attention router output
  • False positive rate among surfaced candidates

Thesis Quality

Measures: Researcher + Decider
  • Did the thesis claim hold at exit?
  • Was invalidation respected when its conditions were triggered?
  • Conviction calibration: does high conviction predict higher returns?
  • Quality of invalidation conditions (specific vs. vague)
  • Thesis-to-outcome alignment over time
  • Time from thesis to first profitable exit

Execution Quality

Measures: Execution Layer
  • Slippage: expected vs. actual execution price
  • Execution reliability: % of approved intents successfully completed
  • Stop/TP adherence: did exits trigger at the right levels?
  • Timing: execution speed after policy approval
  • Reconciliation accuracy: book matches reality
  • Fee optimization over time
Feedback Speeds
flowchart TD
  OUT["Outcomes + P&L"] --> R["Routing Quality"]
  OUT --> TH["Thesis Quality"]
  OUT --> EQ["Execution Quality"]

  R -->|"fast"| CTX["Perception tuning"]
  TH -->|"medium"| WM["World model update"]
  EQ -->|"slow"| POL["Policy review"]

  style OUT fill:#fdf2f8,stroke:#db2777,color:#0c1024
  style R fill:#ecfeff,stroke:#0891b2,color:#0c1024
  style TH fill:#f5f3ff,stroke:#7c3aed,color:#0c1024
  style EQ fill:#ecfdf5,stroke:#059669,color:#0c1024
  style CTX fill:#f7f8fb,stroke:#c4c9d6,color:#5a6378
  style WM fill:#f7f8fb,stroke:#c4c9d6,color:#5a6378
  style POL fill:#f7f8fb,stroke:#c4c9d6,color:#5a6378
      

Design Principles

Non-negotiable properties of the system. Not guidelines. Structural constraints.

Specialized agents with narrow mandates outperform general-purpose models across every function.
A single general-purpose model reviewing everything produces mediocre results across the board. The Sourcing Analyst is optimized purely for early discovery. The Principal is optimized purely for thesis construction. The Policy Gate is optimized purely for risk enforcement. Specialization creates depth; the executive loop creates breadth.
Language models propose. Deterministic policy gates dispose. No exceptions.
Language models are capable reasoners. They are not reliable rule-enforcers under adversarial conditions. Every execution intent generated by Fair, however confident and well-reasoned, must pass through a deterministic gate before it reaches capital. The creative capacity lives in the producers and the decider. The rule-enforcement capacity lives in a layer that cannot be reasoned with or convinced.
Every project analyzed makes the system better at analyzing the next one.
Research accumulates in a structured ledger. Pattern recognition improves across thousands of data points. The monitored account graph becomes more precise over time as accounts are promoted and demoted by signal quality. Unlike a traditional fund, where institutional knowledge degrades when analysts leave, Fair's knowledge is permanent, versioned, and growing.
Every claim traces back to a raw observation. Every decision to the exact inputs that informed it.
The system can reconstruct, at any point in time, why it believed what it believed and why it did what it did. This is not a compliance feature. It is the property that makes genuine self-improvement possible. A system that cannot explain its decisions cannot reliably improve them.
What the system pays attention to is separate from what it decides on.
Conflating attention and decision inputs creates a subtle but critical failure mode: the routing layer begins to silently shape outcomes rather than transparently surface inputs. In Fair, the Attention Router and the Decision Packet are separate components, logged independently, evaluated independently, and improved independently.