Skip to content

02-organization.md

Latest commit

 

History

History
515 lines (400 loc) · 25.9 KB

02-organization.md

File metadata and controls

515 lines (400 loc) · 25.9 KB
Raw

02 — The Organization

Date: 2026-09-04 Status: proposal, v2.

The complete company: ten departments, seven market desks, sixty-seven role charters, and the mechanism by which the company grows itself from a launch roster of sixteen agents to forty, then to hundreds.

1. Principles

  1. A role is a charter, not an agent. The charter is the full remit. An agent holds some coverage of a charter. Several agents may hold the same charter on different desks.
  2. Every charter area is owned by someone at all times. A launch agent holding six charters is explicitly standing in for six future specialists. Coverage is never dropped — only reassigned.
  3. Growth is earned and recorded. Splitting a role, hiring an agent, opening a desk: each is an OrgChange with a trigger, evidence, and a meeting decision behind it.
  4. Specialisation is the point. A researcher may not trade. An analyst may not decide portfolio exposure. Risk may veto anyone. An auditor may challenge anything. Separation of duties is enforced by write scopes, not by instruction.
  5. Seniority is real. Junior agents produce; senior agents review; leads assign and represent the team in meetings; directors set direction. Model tier follows seniority, which is also how cost follows importance.

2. The ten departments

# Department Owns Head
1 Executive / Mission Control Company direction, missions, priorities, org development, the Chair Company Manager
2 Market Intelligence Observation, evidence gathering, briefings, market state Head of Market Intelligence
3 Quantitative Research Hypotheses, experiments, statistics, modelling, research engineering Research Director
4 Strategy Laboratory Strategy discovery, synthesis, debate, adversarial testing, validation Head of Strategy
5 Portfolio & Risk Construction, allocation, exposure, correlation, risk authority Chief Risk Officer
6 Trading Operations Setup analysis, trade planning, approval, execution, monitoring Head of Trading
7 Audit & Governance Independent challenge of research, data, backtests, execution, agents Chief Auditor
8 Knowledge & Memory Institutional memory, archive, registries, ledger, knowledge graph Chief Knowledge Officer
9 Infrastructure Data systems, compute, scheduling, runtime, observability, integrations Head of Infrastructure
10 Institutional Governance Preregistration, custody, evidence typing, forecast scoring, budget ledger — in service to the other nine Governance Director

Department 10 is where nullius's eleven officers live. They are a service department, like legal or compliance in a real firm: every department uses them, none reports to them, and they replace nobody.

3. The seven market desks

Desks cross departments. An agent is (role, desk).

Desk Instruments Engines Key data
CRYPTO spot, perpetuals, funding, basis martex-quant ccxt lake, funding, on-chain
EQUITIES single names, ETFs, indices equities engine prices, fundamentals, filings, factor models
OPTIONS listed options, vol surfaces options engine chains, IV, greeks, term structure
FUTURES index, rates, term structure futures engine continuous contracts, roll calendars
COMMODITIES energy, metals, agriculture futures engine curves, inventories, seasonality
FX majors, crosses, carry fx engine bars, rate differentials, central bank calendar
MEMECOINS micro-cap tokens, launches martex-quant meme/ DexScreener, GeckoTerminal, wallet cohorts

Desks open in stages. CRYPTO first (the data lake and engine already exist), then EQUITIES, then OPTIONS, then the rest. Opening a desk is registering a DeskConfig and staffing it — no architectural change.

Desk × role is the multiplier that takes the company to hundreds of agents. Seven desks × a forty-role roster is a 280-agent ceiling, reached only where evidence justifies each hire.

4. The full role roster — 67 charters

Every role listed in CLAUDE.md §4, plus the roles the architecture requires. Tier is the model tier. Launch says which launch agent stands in for it (see §5).

4.1 Executive / Mission Control

# Role Charter Tier Launch
1 Company Manager (CEO) Company direction, mission portfolio, final escalation, chairs the Board High AG-01
2 Research Director (CIO) Research agenda, what gets studied, allocation of research effort High AG-02
3 Operations Director Throughput, blockers, scheduling, department health Mid AG-03
4 Mission Orchestrator Decomposes missions into projects and tasks, tracks progress Mid AG-03
5 Chief of Staff / Chair Runs meetings: agenda, rounds, budget, minutes, action items Mid AG-03
6 Org Development Lead Watches org metrics, proposes fission/fusion/hiring, runs org experiments Mid AG-02

4.2 Market Intelligence

# Role Charter Tier Launch
7 Head of Market Intelligence Briefing quality, source coverage, desk intelligence priorities Mid AG-04
8 Fundamental Analyst Financial statements, valuation, earnings, sector economics Mid AG-04
9 News Analyst Event detection, materiality, timeline reconstruction Low AG-04
10 Sentiment / Social Analyst Positioning, crowding, social signal with base rates attached Low AG-04
11 Technical Analyst Price structure, trend, volatility, microstructure observation Mid AG-04
12 Macro Analyst Rates, growth, inflation, policy, cross-asset linkage Mid AG-04
13 Regime Analyst Regime identification, transitions, conditional behaviour Mid AG-04
14 Alternative Data Analyst On-chain, flows, positioning, novel sources and their reliability Mid AG-04
15 Source Reliability Officer Tracks every source's freshness, revisions, historical accuracy Low AG-04

4.3 Quantitative Research

# Role Charter Tier Launch
16 Lead Researcher Owns a research programme, assigns work, defends findings High AG-05
17 Quant Researcher Designs and runs experiments, forms hypotheses Mid AG-06
18 Statistical Researcher Test selection, power, multiple testing, confidence intervals Mid AG-06
19 Backtest Researcher Backtest design, cost realism, walk-forward protocol Mid AG-06
20 Simulation / Monte Carlo Researcher Path simulation, bootstrap, prop-rule and ruin analysis Mid AG-06
21 ML Researcher Learned models, feature design, purged CV, calibration Mid AG-06
22 Factor Researcher Factor construction, exposure, spanning tests Mid AG-06
23 Research Engineer Builds experiment specs, engine adapters, data pipelines, tooling Mid AG-07
24 Data Scientist Exploratory analysis, visualisation, anomaly surfacing Mid AG-06

4.4 Strategy Laboratory

# Role Charter Tier Launch
25 Head of Strategy Strategy portfolio, promotion agenda, chairs Strategy Committee High AG-08
26 Strategy Architect Turns findings into strategy designs: entries, exits, sizing High AG-08
27 Strategy Discovery Agent Systematic search over the hypothesis space for candidates Mid AG-08
28 Strategy Synthesizer Combines signals and strategies into coherent specifications Mid AG-08
29 Strategy Critic Attacks the design: assumptions, fragility, hidden costs Mid AG-09
30 Adversarial Researcher Actively tries to break results; hunts for the reason it is wrong Mid AG-09
31 Replication Researcher Independent re-test with a deliberate declared variation Mid AG-10
32 Robustness Researcher Parameter, period, universe, cost and regime stress grids Mid AG-10
33 Validation Researcher Runs promotion gates; owns the out-of-sample request Mid AG-10

4.5 Portfolio & Risk

# Role Charter Tier Launch
34 Chief Risk Officer Risk authority, limits, veto, kill switches, chairs Risk Committee High AG-11
35 Risk Manager Per-strategy and per-desk risk assessment on every proposal Mid AG-11
36 Portfolio Manager Portfolio construction from approved strategy signals Mid AG-12
37 Exposure Analyst Gross, net, concentration, sector, desk and factor exposure Mid AG-11
38 Correlation Analyst Inter-strategy correlation, drawdown interaction, diversification Mid AG-11
39 Capital Allocation Analyst Allocation across strategies and desks; capacity analysis Mid AG-12
40 Stress Testing Analyst Scenario and historical stress, tail behaviour, liquidity shock Mid AG-11

4.6 Trading Operations

# Role Charter Tier Launch
41 Head of Trading Operational integrity of everything that reaches a broker Mid AG-13
42 Market Setup Analyst Whether current conditions suit an approved strategy Mid AG-13
43 Trade Planner Turns target exposure into an executable plan Mid AG-13
44 Trade Approval Agent Final gate: checks approval chain, limits, and halts Low AG-13
45 Execution Agent Submits orders through the broker adapter; nothing else Low AG-13
46 Position Monitor Live positions, drawdown, divergence from expectation Low AG-13
47 Post-Trade Analyst Slippage, cost attribution, backtest-vs-live gap Mid AG-13

4.7 Audit & Governance

# Role Charter Tier Launch
48 Chief Auditor Independent challenge authority across the whole company High AG-14
49 Research Auditor Preregistration compliance, HARKing, cherry-picking, spec drift Mid AG-14
50 Data Auditor Timestamps, revisions, survivorship, point-in-time correctness Mid AG-14
51 Backtest Auditor Look-ahead, cost realism, liquidity assumptions, leakage Mid AG-14
52 Execution Auditor Approval chains, limit breaches, unapproved orders Low AG-14
53 Agent Behavior Auditor Role-boundary violations, fabricated figures, unsupported claims Mid AG-14

4.8 Knowledge & Memory

# Role Charter Tier Launch
54 Chief Knowledge Officer What the company knows, and how confident it is entitled to be Mid AG-15
55 Institutional Memory Keeper Lessons learned, standing rules, what not to repeat Mid AG-15
56 Research Archivist Every experiment, result and document, retrievable Low AG-15
57 Strategy Registrar Strategy registry, versions, states, promotion history Low AG-15
58 Hypothesis Ledger Keeper Every hypothesis and trial ever registered, including the dead Low AG-15
59 Knowledge Graph Curator Nodes, edges, dependency and contradiction structure Mid AG-15

4.9 Infrastructure

# Role Charter Tier Launch
60 Head of Infrastructure Platform health, capacity, reliability Mid AG-16
61 Data Systems Engineer Ingestion, validation, storage, freshness, backfills Mid AG-16
62 Compute Manager Job scheduling, resource limits, queue health Low AG-16
63 Agent Runtime Manager Agent lifecycle, stuck workers, restarts, model routing Low AG-16
64 Observability Engineer Metrics, alerting, dashboards feeding the station Low AG-16
65 Integrations Engineer External APIs, brokers, feeds, credentials handling Mid AG-16

4.10 Institutional Governance (the nullius officers)

# Role Charter Tier Launch
66 Governance Director Owns the integrity machinery; reports findings to the Board Mid AG-17
67 Registrar Locks and hashes preregistrations before any run None AG-17
68 Custodian Holds sealed out-of-sample data; releases counted queries None AG-17
69 Evidence Officer Enforces evidence typing; rejects speculation promoted to fact None AG-17
70 Forecast Scorer Collects pre-run forecasts, scores them, publishes calibration None AG-17
71 Ledger Officer Verifies the hash chain, reconciles totals, reports tampering None AG-17
72 Budget Officer Hierarchical budgets, dispatch refusal, spend reporting None AG-17
73 Provenance Officer Every figure traces to an artifact; flags orphaned numbers None AG-17
74 Skeptic-in-Residence Standing adversarial review, independent of Strategy Lab Mid AG-17
75 Replication Officer Schedules and adjudicates independent replications Low AG-17
76 Peer Reviewer Structured review before any claim becomes institutional Mid AG-17

Seventy-six charters. Most of Institutional Governance is deterministic software wearing a badge — which is deliberate, because "who computed this?" must be a checkable field, and because these officers cost nothing to run.

5. The launch roster — 17 agents

Every charter above is covered from day one. Launch agents are generalists standing in for future specialists, and each one's record says exactly which charters it currently holds.

Agent Name Department Holds charters Tier
AG-01 CEO Executive 1 High
AG-02 CIO Executive 2, 6 High
AG-03 OPS Executive 3, 4, 5 Mid
AG-04 INTEL Market Intelligence 7–15 (all nine) Mid
AG-05 LEAD-R Quant Research 16 High
AG-06 QUANT Quant Research 17–22, 24 Mid
AG-07 ENG-R Quant Research 23 Mid
AG-08 STRAT Strategy Lab 25–28 High
AG-09 CRITIC Strategy Lab 29, 30 Mid
AG-10 VALID Strategy Lab 31–33 Mid
AG-11 RISK Portfolio & Risk 34, 35, 37, 38, 40 High
AG-12 PM Portfolio & Risk 36, 39 Mid
AG-13 TRADE Trading Operations 41–47 (all seven) Mid
AG-14 AUDIT Audit & Governance 48–53 (all six) Mid
AG-15 KNOW Knowledge & Memory 54–59 (all six) Mid
AG-16 INFRA Infrastructure 60–65 (all six) Low
AG-17 GOV Institutional Governance 66–76 (all eleven) Mixed

Seventeen agents, one desk (CRYPTO), seventy-six charters covered, zero gaps.

The company is recognisably itself on day one: an Executive that sets missions, Intelligence that observes, Research that experiments, a Strategy Lab that builds and attacks, Risk that can veto, Trading that only paper-trades, Audit that challenges everyone, Knowledge that remembers, Infrastructure that keeps it running, and Governance that makes cheating impossible.

Then it grows.

6. Role fission — how the company builds itself

This is the growth mechanism, and it is the answer to "start with sixteen, end with forty, then hundreds."

6.1 The idea

Every agent's record carries:

charter    the full remit of its role(s)
coverage   the charter areas this agent currently holds
load       measured: queue depth, latency, tasks/day, backlog age
quality    measured: calibration, replication survival, objection hit rate,
           audit findings against it, cost per accepted output

When load or quality metrics cross a declared threshold in a coverage area, the Org Development Lead proposes a fission: split the coverage set, hire a specialist for the split-off part, and hand over.

        AG-04 INTEL                          AG-04 INTEL
   holds charters 7–15          ──▶     holds 7, 11, 12, 13
   backlog: 34 tasks                          +
   news latency: 19h                    AG-18 NEWS  (charters 9, 10, 15)
                                        AG-19 FUND  (charters 8, 14)

The reverse — fusion — merges two agents whose outputs overlap above a measured threshold, or whose combined load no longer justifies two.

6.2 Triggers

Fission and fusion are proposed on measurement, never on a hunch:

Trigger Threshold example Proposal
Backlog depth > 20 open tasks in one coverage area for 7 days Fission
Response latency Briefing older than the decision that needs it Fission
Quality degradation Calibration falling across a coverage area Fission or retrain
Coverage starvation A charter area with zero outputs in 30 days Fission or explicit closure
New desk opened Desk has no staff for a department Hire (role, desk)
Output overlap Two agents' artifacts > 80% duplicative Fusion
Underuse An agent with < 5 tasks in 30 days Fusion or suspension
Scenario failure An agent fails the training suite for its charter Retrain, reassign, or suspend

6.3 The procedure

1. Org Development Lead detects a trigger and assembles the evidence
2. Writes an OrgChange proposal: kind, affected charters, justification,
   expected effect, and how the effect will be measured
3. Board or Executive meeting debates and decides
4. On approval:
     - new Agent row is created with the split coverage
     - handover: open tasks, channel membership, memory scope transfer
     - the new agent runs the training scenario suite for its charters
     - its starting record is its scenario performance
5. Effect is measured after a declared window
6. Result is recorded — including when the split made nothing better

Step 6 matters. Every org change is an experiment with a predicted outcome, and the company keeps the record of which ones worked.

6.4 The growth path

Stage Agents Desks Shape
Launch 17 CRYPTO Generalists; every charter covered
Stage 2 ~28 + EQUITIES Intelligence splits into 3–4; Research splits statistical/backtest/ML; Audit splits research/data/backtest
Stage 3 ~45 + OPTIONS, FUTURES Most charters held by a dedicated agent; desk-specific analysts and researchers
Stage 4 ~80 + FX, COMMODITIES, MEMECOINS Full roster per active desk; teams per desk; multiple researchers per programme
Stage 5 100+ all seven Multiple agents per charter per desk where load justifies: 4 researchers, 2 fundamental analysts, 3 critics

Stage 5 is the CLAUDE.md §16 target — "4 researchers, 2 fundamental analysts" — reached by measurement rather than by decree.

Nothing in the runtime changes across those stages. Agents are rows; roles are charters; desks are configs. Growth is data.

7. Teams

A Team is a durable group with a lead, a charter, and usually a desk. Missions assign teams, not only individuals.

Team:
    team_id · name · department · desk
    lead_agent · members[]
    charter · standing_channels[]
    active_projects[] · metrics

Examples at Stage 3: TEAM-CRYPTO-MOMENTUM, TEAM-OPTIONS-VOL, TEAM-EQUITY-FACTORS, TEAM-INTEGRITY.

Teams hold their own standups and their own retrospectives. A team's output quality is a measured property, and consistently unproductive teams are dissolved by the same OrgChange mechanism that created them.

8. Permissions — three scopes

Every agent's permissions resolve to three scopes, all enforced in the runtime and, where it matters, in the database.

Read scope (what it can see)

A view registry, as in CLAUDE.md §3. Widening what a role can see requires registering a view — so "what did the Strategy Critic know when it approved this?" is answered by reading a registry, not by tracing calls.

Examples: a Researcher sees the hypothesis, the data spec, and prior findings — not the sealed out-of-sample window. A Trader sees approved instructions and market state — not research in progress. An Auditor sees everything, which is what makes it an auditor.

Write scope (what it can create or change)

Entity Who may write
Hypothesis Researchers, Strategy Lab
Registration (locked) Registrar only
Run / metrics Engines only — no agent
Finding / interpretation Analysts, Researchers
Objection Critics, Adversarial, Auditors, Skeptic-in-Residence
Strategy version Strategy Architect proposes; Head of Strategy promotes
Risk assessment Risk roles only — no other agent, ever
Trade approval Trade Approval Agent, requires a risk assessment
Order Execution Agent only, requires an approval
Org change Org Development Lead proposes; Board decides
Event ledger Append-only — no agent

The three that carry the most weight: agents never write numbers, only Risk writes risk, and execution requires an unbroken approval chain.

Tool scope (what it can invoke)

Bound per agent from its role and desk. A Crypto Technical Analyst has engine.backtest.crypto and data.ohlcv; it does not have broker.submit. Tool calls are logged with inputs, outputs, cost and duration.

9. Skills

Three layers, all versioned, so "skill" is never an unversioned prompt fragment.

Capability — deterministic, tested software

stats.deflated_sharpe@2 · integrity.leak_scan@1 · research.purged_cv@1 · portfolio.factor_attribution@1 · options.implied_vol_surface@1 · futures.roll_yield@1 · stats.calibration@1

Versions are immutable. Results name the version that produced them, so a capability change cannot silently restate history.

Competence — a role's right to see and duty to produce

(role, view_id, output_schema, validators[], capabilities[])

Granting a competence widens a view and extends an output contract. It is a registry change with a test, not a prompt edit.

Playbook — a reusable procedure, LLM-facing

The thing that feels most like a "skill" to a human: a written, versioned, tested procedure an agent follows.

playbook: research.test_cross_sectional_signal@3
  when to use · required inputs · steps · required checks
  common failure modes · what disqualifies the result
  worked example · scenario suite that validates it

Playbooks are tested against the synthetic scenario suite. A playbook revision that lowers the catch rate on planted defects fails CI and does not ship. That is what makes them skills rather than folklore.

Skill categories mirror CLAUDE.md §5: general, market intelligence, quantitative research, strategy, risk, trading operations — with per-desk variants where the procedure genuinely differs (testing an options signal is not testing a crypto momentum signal).

10. Careers — agent lifecycle

Agents have histories, and the history is used.

HIRED ──▶ ONBOARDING ──▶ ACTIVE ──▶ SENIOR ──▶ LEAD ──▶ DIRECTOR
              │             │  │
              │             │  └──▶ RETRAINING ──▶ ACTIVE
              │             └─────▶ SUSPENDED ──▶ REASSIGNED | RETIRED
              └──▶ FAILED_ONBOARDING ──▶ RETIRED
  • Onboarding runs the training scenario suite for the agent's charters. An agent that cannot catch planted defects in its own specialty does not start work; the failure is recorded and the role's playbook is reviewed.
  • Promotion follows measured quality: calibration, findings that survive replication, objections upheld, low audit findings. Promotion raises model tier, which is how cost follows demonstrated value.
  • Suspension follows measured failure: repeated role-boundary violations, fabricated figures, persistently poor calibration.
  • Retirement preserves the record. A retired agent's outputs, evidence and history remain in the ledger permanently.

Agent performance is never evaluated on P&L. A researcher who correctly kills a bad strategy has produced valuable work, and the metrics say so.

11. What each department actually does in a day

Concrete, so the design is checkable against reality.

Market Intelligence — scheduled scans per desk; posts briefings to desk channels; flags anomalies as LEAD nodes (never findings); answers questions from Research; maintains source reliability records.

Quantitative Research — takes questions from the Research Director and from Kickoff meetings; forms hypotheses; preregisters through the Registrar; builds specs; runs engines; reports results with artifact hashes; defends findings in Research Review meetings.

Strategy Laboratory — turns surviving findings into strategy designs; Discovery searches systematically; Synthesizer combines; Critic and Adversarial attack; Replication and Robustness test; Validation runs promotion gates and requests the out-of-sample query.

Portfolio & Risk — assesses every proposal; sets and enforces limits; analyses correlation and exposure; allocates capital; runs stress tests; vetoes.

Trading Operations — evaluates setups; plans trades; checks the approval chain; submits to the paper broker; monitors positions; attributes slippage and reports the backtest-vs-live gap.

Audit & Governance — independently samples and challenges research, data, backtests, executions and agent behaviour; raises findings that any agent must answer; reports to the Board.

Knowledge & Memory — records lessons; curates the graph; keeps registries; answers "have we tried this before?" — which is one of the most valuable questions in the company and currently costs 174 trials' worth of context to answer by hand.

Infrastructure — data freshness, job health, agent liveness, alerting.

Institutional Governance — locks registrations, holds sealed data, scores forecasts, verifies the chain, enforces budgets, and files integrity findings.

Executive — sets missions, arbitrates escalations, chairs meetings, and develops the organization.