An autonomous quantitative research corporation.
Ten departments. Seven market desks. Seventy-six role charters. Agents that observe, hypothesize, experiment, argue in meetings, decide, build strategies, manage risk, trade on paper, remember everything — and expand the organization themselves as the evidence justifies it.
Status: M7 complete — the company has a window.
Mission Control renders the facility from the registries and lights it from the
record. Every number on screen names its source, and Figure(42) without one
is a TypeError. · 2026-09-05
Research software. No live trading adapter exists. Nothing here is proven profitable. Read DISCLAIMER.md.
pip install -e ".[dev]"
aurelis db init # schema, invariants, the org chart
aurelis agent hire # staff the launch roster
aurelis research review # the demonstration
aurelis memory import # inherited trials, gap and all
aurelis station serve # Mission Control on http://127.0.0.1:8787/HYP-0001 CONFIRMED -> REFUTED
claimed max_drawdown < 0.20, measured 0.12364208
universe 3 names (still trading)
objection OBJ-0001 SURVIVORSHIP, critical
test the same rule, universe restored to point-in-time
re-run universe 6 names
restored LUNC/USDT, FTT/USDT, HOTAIR/USDT
max_drawdown 0.12364208 -> 0.64507263
verdict UPHELD
chain chain verified: 68 events, seq 1..68
A researcher registers a drawdown claim over the instruments still trading,
runs it, and it is confirmed. A Critic names SURVIVORSHIP — it does not
write the test; the taxonomy generates it from the specification under review.
The Chair dispatches it. The point-in-time re-run restores three delisted names
and drawdown goes from 12% to 65%. The objection is upheld and the claim is
refuted by a measurement.
Nobody intervenes at any point.
martex-quant found this same defect on real crypto history, where it took a Sharpe of 1.47 to 0.86. Those figures belong to that corpus; the ones above are what this engine measured on fixture instruments where the bias is present by construction.
aurelis memory import
imported 21 ledger entries from martex-quant (29 hypothesis documents)
claimed by the source 125
documented by its entries 120
unallocated 5
carried because Documented per-hypothesis deltas do not sum to
the ledger's stated total. The gap is reported,
not absorbed.
reconciles yes
The import reproduces the source's own arithmetic instead of tidying it.
125 claimed, 120 accounted for by its committed documents, and a five-trial gap
that the source itself says would be fabrication to distribute. Deflated
Sharpes arrive as published — 0.99 against 65 trials, never re-deflated
against Aurelis's own count, because that would restate a figure somebody else
computed.
That is the snapshot bundled in the installed wheel, so the import is
reproducible from the lockfile alone. --bundle <repo> reads a live
repository instead — 174 claimed, 169 documented, the same gap of five. The
reconciliation row stores the SHA-256 of whichever ledger was read, so a corpus
that changed under a re-import is detectable rather than silently merged.
Ask whether an idea is new, and the answer comes from the record:
aurelis memory prior-art "Do funding rate extremes predict forward returns?" --family info.derivatives.funding
MQ-H08 (martex-quant, killed) - close match on extremes, funding
That answer is now in every Brainstorm's evidence pack before anyone speaks.
Confidence is derived, never stored, which is what makes it degrade on its own:
aurelis memory confidence FND-0001
FND-0001 none
verdict confirmed
independent support 1
capped by OBJ-0001 was upheld by measurement: survivorship
The finding still says confirmed. Nobody edited it, and nobody had to
remember to. An objection was upheld against the claim, so the company is no
longer entitled to believe it — and the reason is on the record rather than in
somebody's head. Support is counted the same way: three results that correlate
above 0.7 collapse to one, and the discount says what it discounted instead
of quietly returning a smaller number.
There is no confidence column to go stale, which is the whole point (ADR-0008).
aurelis station serveA facility drawn from the registries and lit by the record. Ten department rooms, a bay per desk, the Graveyard as a full room — and the Registry and the Vault with no corridor, because you cannot walk into a process boundary. Staff figures are the headcount: a room with three people has three, and a room with none is drawn unlit at full size rather than left out.
The rule that makes it trustworthy is a type:
Figure(42) # TypeError — no source
Figure(42, Source.table("agents", "...")) # fine, and hovering shows the queryFigure has no constructor that omits its source, so a number cannot reach a
page without naming the row, artifact or registry entry it came from. "Nothing
on this page was typed" is checkable by reading the type rather than by
auditing every call site. Where nothing was measured, the page says NO DATA
and why — never 0, because a zero is a measurement and the two justify
different conclusions.
Open /hypothesis/HYP-0001 after the demonstration and the whole story is one
page: the claim, the preregistration hash and its lock time, the criteria
committed before the run, the code version and data fingerprint, every metric
with computed_by = ENGINE, the survivorship objection, and the measurement
that killed it. Why does the company believe this? is answered by scrolling.
aurelis station build writes the same record to a single file that fetches
nothing — no stylesheet, script, font or image — stamped with the ledger head
and the chain verification, so a finding can be cited years after the database
has moved on.
aurelis mission run opens a mission with a Kickoff meeting, plans a
project into three dependency-sequenced tasks, runs them, and closes with a
Retrospective that scores the kickoff's forecasts against what happened.
INTEL briefs the desk, QUANT checks that briefing against a window it
measured itself, LEAD-R decides. There is no orchestrator — a task whose
dependency has not succeeded is simply invisible to the queue.
aurelis research run takes a single hypothesis from claim to verdict:
propose, screen for prior art, lock a preregistration, design, run, and
derive the verdict from criteria fixed before anything executed. On 240 bars it
returns UNDERPOWERED, because 240 bars genuinely cannot detect a Sharpe of
0.05 — and saying so is the point.
Look around:
aurelis org show · org desks · org charters |
the company as designed |
aurelis agent list · agent show INTEL |
what one agent holds, sees, writes, may invoke |
aurelis mission show MSN-0001 |
every task, its status, what it waits on |
aurelis meeting show MTG-0001 |
the transcript, who changed their mind, who dissented |
aurelis meeting calibration |
how good the company's forecasts have been |
aurelis research show HYP-0001 |
every metric, its interval, and who computed it |
aurelis research graveyard |
everything killed, and why |
aurelis research defects |
every market defect and how it is settled |
aurelis tick · aurelis doctor |
advance the working day; check the workspace |
| Doc | Covers |
|---|---|
docs/00-audit.md |
What already exists on this machine and what it is worth |
docs/01-architecture.md |
The system: layers, repository, agent runtime, communication, desks, engines, self-improvement, platform, cost, safety, testing |
docs/02-organization.md |
The company: 10 departments, 7 desks, all 76 role charters, launch roster, role fission, permissions, skills, careers |
docs/03-meetings.md |
How agents work together: 11 meeting types, the 7-phase protocol, brainstorming, debate, cost control |
docs/04-domain-model.md |
Every entity, its lifecycle, and who may write it |
docs/05-lifecycles.md |
Research → strategy → portfolio → risk → trading |
docs/06-mission-control.md |
The station: the facility, drill-down, every view |
docs/07-roadmap.md |
M0–M13, each with an acceptance test |
docs/adr/ |
The nine decisions that are hard to reverse |
A corporation of AI agents that researches markets and builds systematic strategies. The software's job is to make the company function: departments, desks, teams, colleagues, meetings, tools, memory, budgets, careers, and a building to work in.
EXECUTIVE ─── missions, priorities, org development, the Chair
│
├── MARKET INTELLIGENCE ────┐
├── QUANTITATIVE RESEARCH ──┤
├── STRATEGY LABORATORY ────┤ × 7 DESKS
├── PORTFOLIO & RISK ───────┤ crypto · equities · options · futures
├── TRADING OPERATIONS ─────┤ commodities · FX · memecoins
├── AUDIT & GOVERNANCE ─────┤
├── KNOWLEDGE & MEMORY ─────┘
├── INFRASTRUCTURE
└── INSTITUTIONAL GOVERNANCE ─── serves the other nine, replaces none
Agents work individually. Teams work together. Meetings decide.
1. New repository, new architecture. Aurelis is its own system.
- martex-quant is a tool in the toolbox — one research engine (crypto),
a validated data lake, and a statistics library, reached only through
engines/martex/. It helps researchers with part of their work. It generates no hypotheses, decides nothing, and no agent ever sees it except as tool calls. - nullius contributes platform patterns (hash-chained ledger, preregistration triggers, evidence typing, budget accounting) and staffs one service department — Institutional Governance — whose eleven officers serve the other nine. They have no authority over research direction and replace nobody.
- Everything else is Aurelis's own. (ADR-0001)
2. Wide scope from the start. Seven market desks: crypto, equities,
options, futures, commodities, FX, memecoins. A desk is an orthogonal
dimension crossing every department; an agent is (role, desk). Opening a desk
is registering a config and staffing it — no architectural change.
(ADR-0004)
3. Subscription-first, minimum cost. Everything runs on the Claude Pro subscription through a provider abstraction until a budget is set; switching to a metered API is a config change. Deterministic work costs nothing, models are tiered by seniority, meetings are budgeted, idle is free, and the entire system is testable at zero cost through a mock provider. (ADR-0007)
4. Meetings are first-class — and they work. Real multi-round discussion with a full kept transcript. Kickoff at the start of every mission and project, Retrospective at the end, both enforced by the state machine. Plus Brainstorm, Research Review, Debate, Strategy Committee, Risk Committee, Incident Review, Standup, Board and All-Hands as needed.
The mechanism that makes debate end in evidence rather than in exhaustion: every objection must carry a discriminating test — an executable spec that would settle it — and the Chair runs it, in the meeting, with everyone watching. (ADR-0002)
The full org chart from CLAUDE.md §4 exists from the start — as charters.
Seventeen launch agents hold them as generalists, and each agent's record says
exactly which future specialists it is standing in for.
| Launch | Holds | Becomes |
|---|---|---|
| AG-04 INTEL | all 9 Market Intelligence charters | Fundamental, News, Sentiment, Technical, Macro, Regime, AltData, Source Reliability, Head |
| AG-06 QUANT | 7 research charters | Statistical, Backtest, Simulation, ML, Factor, Data Scientist, Quant |
| AG-13 TRADE | all 7 Trading charters | Setup, Planner, Approval, Execution, Monitor, Post-Trade, Head |
| AG-14 AUDIT | all 6 Audit charters | Research, Data, Backtest, Execution, Behaviour auditors, Chief |
| … | … | … |
Then the company splits its own roles as evidence justifies:
17 agents ──▶ ~28 ──▶ ~45 ──▶ ~80 ──▶ 100+
1 desk 2 4 7 7
Role fission is the mechanism. Every agent carries measured load and quality; when a threshold trips, the Org Development Lead proposes a split with evidence and a predicted effect, a Board meeting decides, the new specialist is hired and onboarded, and the effect is measured afterwards — recorded even when the split made nothing better.
That is CLAUDE.md §16's "4 researchers, 2 fundamental analysts" reached by
measurement. Nothing in the runtime changes across those stages — agents are
rows, roles are charters, desks are configs. Growth is data.
(ADR-0003)
Self-improvement is core and stays core. Three measurements, none circular:
- Forecast calibration. Every agent records a probability before every experiment and every meeting outcome, and it is scored afterwards. One cheap call each, per-agent quality signal, no LLM grading another LLM's prose.
- Training scenarios. A synthetic engine generates research problems where the answer is known — planted effects of stated size, planted leaks, planted survivorship bias, and a large fraction with nothing in them. New agents are scored before they start work; org changes are measured by running the same suite with and without; playbook revisions that lower the catch rate fail CI. Runs offline and cheaply. (ADR-0005)
- Forward paper performance. The backtest-vs-live gap, per strategy and per desk — the only measurement where reality gets a vote.
Agents interpret. Software computes. No metric, verdict or confidence is ever produced by a model. A turn or finding containing a numeral that is not in the evidence pack or a tool result is rejected by a validator.
Everything traces to evidence. Findings carry evidence refs; evidence carries artifact hashes; artifacts carry the spec, seed, data fingerprint and code version that produced them. In Mission Control, every number on screen opens its source.
Supporting them: preregistration locked and hashed before any run (database trigger, not a prompt), an append-only hash-chained ledger, risk that cannot be bypassed (foreign key, not instruction), sealed out-of-sample data behind a process boundary with a counted query budget, and no live trading adapter in the repository at all (ADR-0006).
The primary interface. A cutaway industrial facility — pixel-art-inspired, dark sci-fi, dense but readable — where every room is a department, every bay on the floor is a desk, and every figure on screen names the artifact it was read from.
Click a department to open it. Click an agent to see what it is doing, what it can see, what it can write and what it costs. Click a meeting to read the argument, the evidence cited, who changed their mind, and who dissented. Click a strategy to see why the company believes in it. The Graveyard is a full room, not a hidden tab.
Two rooms have no corridor into them — the Registry and the Vault — because you genuinely cannot walk into a process boundary.
By M5: given a strategy specification whose universe was chosen with
hindsight, the Strategy Critic raises a SURVIVORSHIP objection, attaches a
discriminating test, the Chair dispatches it inside a Research Review meeting,
the point-in-time run comes back with the Sharpe collapsed, the author concedes
on the record, and the hypothesis is refuted — with no human in the loop.
That is a real discovery from the existing research corpus, reproduced automatically by the company, five milestones in.
| M0 ✅ | Foundations | ledger, budgets, artifacts, queue, provider abstraction |
| M1 ✅ | Agent runtime | 76 charters, 17 agents, permissions, views, tools, the loop |
| M2 ✅ | Missions | missions → projects → tasks, dependencies, the working day |
| M3 ✅ | Meetings | seven-phase protocol, forecasts, objections with tests, dissent |
| M4 ✅ | Research lifecycle | engines, preregistration, experiments, verdicts |
| M5 ✅ | Critique & audit | market defects, point-in-time, the review that kills |
| M6 ✅ | Memory & knowledge | graph, lessons, corpus import, vault export |
| M7 ✅ | Mission Control | the live facility, every figure sourced |
| M8 | Strategy, portfolio, risk | versions, gates, veto |
| M9 | Paper trading | approval chain, the backtest-live gap |
| M10 | Training scenarios | onboarding and playbook regression |
| M11 | Org development | the company grows itself |
| M12 | Multi-desk | equities → options → futures → commodities → FX → memecoins |
| M13 | Scale | 100+ agents, seven desks, hardening |
Full acceptance criteria in docs/07-roadmap.md.
Multiple specialized agents genuinely collaborating. Research that is reproducible. Failed research preserved and used. Strategies versioned and evidence-backed. Risk that is independent. A company that learns from its results and improves its own structure on measured evidence. And a human who can operate and understand the whole thing through Mission Control without opening a terminal.
Profitability is what the company is built to pursue — through evidence, across seven markets, with an organization that keeps getting better at looking. It is not assumed, and the system is built so that it can say when it has not found one.