DumbMoney

An evidence-native, autonomous multi-agent fund operating system

Seven specialist systems. One portfolio brain. Two exclusive brokerage boundaries.

Research → falsification → assurance → policy → allocation → routing → reconciliation → learning.

DumbMoney is the public-source composition root for a local-first, multi-agent market research and portfolio orchestration system. It binds forecasting, technical analysis, experimental research, constitutional governance, cyber-assurance, portfolio allocation, and venue-specific brokerage cells into one replayable operating loop.

It is designed around a simple premise:

Intelligence may propose. Evidence may qualify. Policy may permit. Only the assigned venue owner may act.

DumbMoney is not an eighth peer agent. It is the organism around the agents: the scheduler, memory spine, portfolio allocator, router, reconciler, recovery system, and observability plane.

Product truth

Question	Current answer
What is it?	A public-source reference architecture and executable local harness for an autonomous, evidence-gated multi-agent fund.
What is integrated?	Seven exact, content-addressed ObtuseAI component snapshots plus the DumbMoney orchestration runtime.
What runs today?	Deterministic provider-free replay, read-only forward evidence collection, model-budget allocation, technical analysis, evidence courts, assurance gates, routing simulation, reconciliation, and dashboards.
What does not happen by cloning?	Live orders, broker authentication, capital activation, automatic strategy promotion, or profitability claims.
Who owns Robinhood?	Dopey, exclusively.
Who owns Kalshi?	Dummy, exclusively.
Can DumbMoney call a broker?	No. It produces venue-owned intents; it never owns broker API authority.
Can research promote itself?	No. Doofus can recommend a challenger but cannot self-promote or trade.
Is the system profitable?	Not established. Synthetic replay and operational health are not profitability evidence.
Runtime posture	Public source, private local state, ignored credentials, loopback status surfaces, explicit activation.

The organism

flowchart LR
    subgraph SENSE["SENSE"]
        DIMWIT["Dimwit<br/>chart + technical intelligence"]
        WATERBOY["Waterboy<br/>sports forecasting"]
        VENUEOBS["Dummy + Dopey<br/>venue observations"]
    end

    subgraph THINK["THINK"]
        DOOFUS["Doofus<br/>research + evidence court"]
        NIMROD["Nimrod<br/>provenance + cyber assurance"]
    end

    subgraph GOVERN["GOVERN"]
        BLUNDER["Blunder<br/>constitutional policy"]
        DUMBMONEY["DumbMoney<br/>portfolio brain"]
    end

    subgraph ACT["ROUTE + ACT"]
        DUMMY["Dummy<br/>Kalshi owner"]
        DOPEY["Dopey<br/>Robinhood owner"]
    end

    LEDGER["Content-addressed evidence<br/>reconciliation + memory"]

    DIMWIT --> DOOFUS
    WATERBOY --> DOOFUS
    VENUEOBS --> DOOFUS
    DOOFUS --> NIMROD
    NIMROD --> BLUNDER
    BLUNDER --> DUMBMONEY
    DUMBMONEY -->|"Kalshi intent only"| DUMMY
    DUMBMONEY -->|"Robinhood intent only"| DOPEY
    DUMMY --> LEDGER
    DOPEY --> LEDGER
    LEDGER --> DUMBMONEY
    LEDGER --> DOOFUS

This separation is structural, not just descriptive. The machine-readable topology in configs/architecture.v5.json binds every cell to its assignment, authority ceiling, input/output contract, and venue route. Runtime validation fails if those assignments drift.

Roles, skills, tools, and authority

System	Role	Core skills and tools	Produces	Authority ceiling
DumbMoney	Fund harness and portfolio brain	Scheduling, durable memory, budget auctions, allocation, routing, reconciliation, recovery, telemetry, API/dashboard	Portfolio envelopes, venue-owned broker intents, ledger events, health reports	No broker credentials, no broker API calls, no self-promotion
Blunder	Constitutional control plane	Permissions, policy evaluation, capital limits, quorum, stop/kill state, release gates	Policy decisions and bounded authority envelopes	Governs; does not place venue orders
Dimwit	Market perception and technical analysis	Chart perception contracts, OHLCV validation, SMA, EMA, MACD, RSI, ATR, Bollinger bands, support/resistance	Content-addressed technical observations	Research input only; no probability invention, credentials, or orders
Waterboy	Sports intelligence	Point-in-time sports data, projections, calibration, schedule and result normalization	Sports forecasts and evidence packets	Research input only; never routes directly to Kalshi
Doofus	Research laboratory and evidence court	Hypothesis generation, adversarial falsification, held-out evaluation, embargoes, leakage checks, calibration, champion/challenger experiments	Evidence verdicts, retest plans, challenger recommendations	Cannot trade or automatically promote its own work
Nimrod	Constitutional cyber-defense assurance	Provenance, integrity, replay, Witness records, cross-implementation checks, read-only observation	Security-assurance verdicts and exact blocked states	No target mutation, containment, trading, or self-authorization
Dummy	Exclusive Kalshi brokerage cell	Prediction-market research, market normalization, probability forecasts, Kalshi execution contracts, settlement and calibration	Kalshi receipts, fills, abstentions, outcomes	Only Kalshi; live action still requires explicit external authority
Dopey	Exclusive Robinhood brokerage cell	Equities, options, crypto research, broker/account reconciliation, defined-risk execution contracts	Robinhood receipts, fills, abstentions, account evidence	Only Robinhood; live action still requires explicit external authority

The component names are intentionally memorable. Their contracts are not casual: every crossing between cells is typed, content-addressed, replayable, and authority-bounded.

The autonomous loops

1. Market-intelligence loop

Dimwit receives chronological, timezone-aware OHLCV and produces deterministic descriptive indicators:

SMA20;
EMA12 and EMA26;
MACD;
RSI14;
ATR14;
Bollinger bounds;
rolling support and resistance;
alignment state and source digest.

It does not convert an indicator into a probability or expected return. Without a point-in-time, embargoed, leakage-cleared, cost-stressed mapping to outcomes, Doofus returns BLOCKED_RESEARCH_INPUT_ONLY.

flowchart LR
    OHLCV["OHLCV / chart evidence"] --> DIMWIT["Dimwit analysis"]
    DIMWIT --> INPUT["Research input only"]
    INPUT --> COURT["Doofus held-out court"]
    COURT -->|"insufficient evidence"| BLOCK["Blocked + experiment request"]
    COURT -->|"qualified"| CHALLENGER["Shadow challenger recommendation"]
    CHALLENGER -. "never auto-promotes" .-> PORTFOLIO["DumbMoney portfolio"]

2. Sports-to-Kalshi loop

Waterboy owns sports forecasting. Dummy owns Kalshi. They are deliberately not the same role.

Waterboy forecasts flow through Doofus falsification, Nimrod assurance, Blunder policy, and DumbMoney allocation before a Kalshi intent can reach Dummy. Waterboy never receives Kalshi credentials.

flowchart LR
    DATA["Sports data"] --> WATERBOY["Waterboy forecast"]
    WATERBOY --> DOOFUS["Doofus evidence court"]
    DOOFUS --> NIMROD["Nimrod assurance"]
    NIMROD --> BLUNDER["Blunder policy"]
    BLUNDER --> FUND["DumbMoney allocation"]
    FUND -->|"Kalshi route"| DUMMY["Dummy"]

3. Cross-venue thesis loop

A thesis may create different candidates for prediction markets and securities, but the system does not pretend those instruments are interchangeable.

Each venue receives:

its own market snapshot;
its own cost model;
its own liquidity and risk estimate;
its own evidence verdict;
its own allocation;
its own reconciliation record.

The shared thesis is learned across venues only after outcomes are frozen and settled. This supports cross-domain learning without cross-venue authority leakage.

4. Portfolio and routing loop

DumbMoney converts qualified candidates into a portfolio, not a list of independent bets. Allocation accounts for evidence strength, uncertainty, cost, risk, correlation tags, capital ceilings, and constitutional policy.

The resulting route is exact:

Destination	Required owner	Other cells
`dummy_kalshi`	Dummy	Rejected
`dopey_robinhood`	Dopey	Rejected
Direct broker call from DumbMoney	Nobody	Forbidden
Direct order from Dimwit, Waterboy, Doofus, Nimrod, or Blunder	Nobody	Forbidden

Provider-free replay emits SHADOW_SIMULATE_ONLY intents with zero broker calls and no execution authority.

5. Evidence and recursive-improvement loop

Every cycle separates generation, evaluation, promotion, and action:

stateDiagram-v2
    [*] --> Hypothesis
    Hypothesis --> FrozenCandidate
    FrozenCandidate --> AdversarialEvaluation
    AdversarialEvaluation --> Rejected
    AdversarialEvaluation --> ShadowChallenger
    ShadowChallenger --> SettledEvidence
    SettledEvidence --> Retest
    SettledEvidence --> PromotionRecommendation
    PromotionRecommendation --> HumanOrConstitutionalReview
    HumanOrConstitutionalReview --> ResearchOnly
    HumanOrConstitutionalReview --> BoundedDeployment
    BoundedDeployment --> SettledEvidence

Doofus can mutate hypotheses and recommend challengers. It cannot edit the rules that judge it, choose its own hidden test, grant itself execution authority, or erase losing evidence.

6. Security-assurance and recovery loop

Nimrod validates the local architecture digest, component health, source lineage, candidate authority, research verdicts, and immutable fixture digest before routing. A mismatch produces an exact blocked state.

DumbMoney adds:

restart-safe stage checkpoints;
a hash-linked SQLite event ledger;
durable stop and kill state;
content-addressed source snapshots;
idempotent imports;
watchdog heartbeats;
verified backups;
fail-closed service recovery.

Assurance is evidence supplied to governance. It does not authorize itself.

7. Model-economy loop

OpenRouter is an optional bounded reasoning layer, not an unlimited brain. DumbMoney allocates a hard $10/day budget:

Lane	Daily ceiling	Purpose
Research	$6.00	Hypothesis development and information gathering
Evaluation	$2.00	Critique, falsification, and uncertainty reduction
Operations	$2.00	Summaries, diagnostics, and bounded coordination

Requests compete on expected value of information and uncertainty reduction. Ambiguous provider outcomes are conservatively charged. Provider permission is opt-in, ZDR is requested, and model output never receives trading authority.

Why this architecture is different

Intelligence and authority are different data types

Most agent systems connect a model to tools and call the result autonomy. DumbMoney separates cognition from permission. A persuasive forecast remains untrusted until evidence, assurance, policy, allocation, and venue ownership all agree.

The fund is a harness, not a committee

DumbMoney is not another voter competing with its cells. It is the deterministic environment in which specialists operate: shared memory, capital accounting, routing, recovery, and observability live above the individual agents.

Research cannot grade itself

Doofus generates and challenges strategies, but promotion evidence must be point-in-time, held out, cost stressed, leakage checked, and preserved. Synthetic success is labeled synthetic and cannot silently become a live claim.

Security is inside the decision loop

Nimrod is not a perimeter add-on. Provenance and integrity become explicit inputs before capital routing, while Blunder remains the independent constitutional authority.

Brokerage boundaries are architectural

Dummy and Dopey do not share a generic execution client. Kalshi and Robinhood actions terminate in different cells, with different evidence, reconciliation, and failure semantics.

Recursive improvement compounds evidence, not confidence

The system grows by accumulating immutable observations, predictions, settlements, calibration, rejected hypotheses, and retest results. It is built to compound knowledge while preventing a successful demo from becoming its own authorization.

Capabilities

Plane	Implemented capabilities
Orchestration	Fifteen-stage restart-safe supervisor, continuous mode, deterministic replay, stage checkpoints
Evidence	Content-addressed observations, frozen candidates, outcome separation, settlement, calibration
Research	Adversarial verdicts, held-out TA court, champion/challenger recommendations, quarantine and retest
Portfolio	Proof-weighted allocation, risk/cost inputs, correlation tags, venue allocation
Technical analysis	Deterministic stock/crypto OHLCV indicators and research-only classifications
Dopey data plane	Verified 13,112-file catalog, normalized equity history, zero-authority options context, and research-memory provenance
Walk-forward research	Rolling train/holdout folds, outcome embargo, round-trip costs, drawdown, Brier calibration, and regime diagnostics
Shadow campaign	Content-addressed, deduplicated evidence units and ranked 20-symbol strategy results without promotion authority
Sports	Point-in-time sports observations and evidence-routed Kalshi candidates
Assurance	Seven-tree integrity checks, architecture invariants, Nimrod read-only verdicts
Governance	Blunder policy contracts, durable stop/kill, explicit authority ceilings
Routing	Exclusive Dummy/Kalshi and Dopey/Robinhood broker intents
Model gateway	Optional OpenRouter, hard daily/lane budgets, conservative accounting, ZDR request
Operations	Loopback GET-only API, dashboard, watchdog, Windows persistence plans, verified SQLite backups
Portability	Exact component materialization and isolated Python environment per cell
Data handoff	Secret-rejecting, content-addressed Dopey data intake without changing frozen source

A deterministic demonstration

The committed provider-free scenario is visibly synthetic and exercises the whole control flow without network, model, or broker calls:

uv sync --frozen
uv run dm architecture-status
uv run dm doctor
uv run dm replay

Current expected replay:

components healthy             7
Dimwit TA observations         1
TA candidates promoted         0
research inputs blocked        1
Nimrod assurance              PASS_READ_ONLY_ASSURANCE
portfolio candidates           4
Dummy / Kalshi intents         2
Dopey / Robinhood intents      2
provider calls                 0
broker calls                   0
live activation               false
profitability evidence         false

Replay digest:

28d0813b814d6d99966763a78555041caceef24483ebd3e4153d762333919037

Runtime surfaces

CLI

uv run dm bootstrap --plan-only
uv run dm bootstrap
uv run dm doctor --require-environments
uv run dm architecture-status
uv run dm technical-analysis --input C:\path\to\synthetic-ohlcv.json
uv run dm replay
uv run dm shadow --continuous --serve
uv run dm forward --continuous --serve --allow-network
uv run dm forward-status
uv run dm verify-ledger
uv run dm stop
uv run dm kill
uv run dm resume

OpenRouter is contacted only when --allow-provider is supplied and OPENROUTER_API_KEY exists in the process environment.

The Dopey laptop handoff accepts either the original verified v4 directory bundle or the simple content-addressed v5 ZIP, including SQLite research indexes. See Dopey Data Drop v5.

The imported data can drive the provider-free real-data research loop without changing the frozen Dopey source tree:

uv run dm dopey-data-catalog
uv run dm dopey-research
uv run dm dopey-research-status
uv run dm dopey-campaign-status

Dimwit produces technical observations, Doofus runs disjoint cost-stressed walk-forward holdouts, Nimrod verifies lineage and temporal controls, and DumbMoney may emit only SHADOW_SIMULATE_ONLY Robinhood intents to Dopey. Options and trade memory have zero directional authority. See Dopey Real-Data Research v6. The campaign ranks all 80 symbol/strategy pairs across the configured 20-symbol universe. Re-running unchanged market data records duplicate observations but does not inflate distinct evidence.

Dopey's final laptop-generated public runner configuration is a separate single-file handoff:

uv run dm dopey-runner-config-verify --source C:\path\to\dopey-robinhood-runner.v1.json
uv run dm dopey-runner-config-import --source C:\path\to\dopey-robinhood-runner.v1.json
uv run dm dopey-runner-config-status

The importer is secret-rejecting, content-addressed, and non-activating. See Dopey single-file runner configuration. The Windows service plan schedules this provider-free research once daily and runs one bounded cycle immediately after installation.

Read-only API

The status service is loopback-only and rejects POST:

GET /health/live
GET /health/ready
GET /v1/status
GET /v1/cycles
GET /v1/events
GET /v1/ledger/head
GET /v1/budget
GET /v2/forward/status
GET /v2/scorecards
GET /v2/predictions
GET /v2/settlements
GET /v2/snapshots
GET /v2/service
GET /v4/dopey-handoff
GET /v5/architecture
GET /v6/dopey-research
GET /v7/dopey-campaign

The forward dashboard runs at http://127.0.0.1:8791/dashboard when enabled. It presents evidence maturity, calibration, raw and independent sample counts, the ranked Dopey shadow campaign, model budget, runtime health, Dopey handoff state, and the seven-cell architecture.

Repository map

DumbMoney/
├── configs/                         # runtime, model, service, architecture
├── contracts/                       # versioned inter-cell contracts
├── docs/                            # architecture and evidence specifications
├── fixtures/                        # visibly synthetic deterministic scenarios
├── projects/
│   ├── blunder/                     # constitutional control plane
│   ├── dimwit/                      # chart + technical intelligence
│   ├── doofus/                      # research + evidence court
│   ├── waterboy/                    # sports intelligence
│   ├── nimrod/                      # cyber-defense assurance
│   ├── dopey/                       # Robinhood brokerage owner
│   └── dummy/                       # Kalshi brokerage owner
├── scripts/                         # bootstrap, service, handoff, validation
├── src/dumbmoney/                   # composition root and runtime
├── tests/                           # integration, replay, API, recovery tests
├── DUMBMONEY_WORKSPACE.json         # content-addressed workspace manifest
└── pyproject.toml

Each component keeps its own original README, landing assets, notices, and license. DumbMoney validates each integrated subtree against its recorded Git tree, preventing accidental edits to the embedded source snapshots.

Component provenance

Component	Upstream	Integrated tree
Blunder	`ObtuseAI/blunder`	`398736f9e72e73b86190bae45d2fb526a67fad39`
Dummy	`ObtuseAI/dummy`	`65794d672fa001f2b34489e307ce4883245a7830`
Dopey	`ObtuseAI/dopey`	`d0c581da3c5d8e429490e244ef01a4b2d82eaae7`
Doofus	`ObtuseAI/doofus`	`a48260166f317ebc224506af741a65c88a2ed9d0`
Waterboy	`ObtuseAI/waterboy`	`95bd0fefdd1462c1113a4c845b29f859d164446b`
Dimwit	`ObtuseAI/dimwit`	`77eddc59e2a70ce18647f8555b4dcb3f06722200`
Nimrod	`ObtuseAI/nimrod`	`0618ac4a9644f54905476e4e3855ea55e776c078`

Materialize any component as a standalone local Git repository without a network fetch:

python scripts/materialize_component.py dimwit C:\DumbMoneyRuntime\dimwit
python scripts/materialize_component.py nimrod C:\DumbMoneyRuntime\nimrod

Validation

python scripts/validate_workspace.py
python scripts/scan_secrets.py --root .
uv run python -m unittest discover -s tests -v
uv run dm architecture-status
uv run dm replay --assert-digest 28d0813b814d6d99966763a78555041caceef24483ebd3e4153d762333919037
uv run dm verify-ledger
uv build

CI runs the integrated workflow on Windows with Python 3.12 and 3.14.

Operating boundary

DumbMoney is public source with private local operation.

The repository intentionally excludes:

broker credentials and browser sessions;
account history and private market data;
model-provider API keys;
live runtime databases;
local state, logs, imports, and backups;
signed operator authority;
deployment-specific capital limits.

A clone is research-only until the operator separately supplies every required credential, dataset, policy, deployment receipt, and explicit authority. Nothing in this repository is financial advice, an offer to manage capital, a profitability claim, or authorization to place an order.

License and security

DumbMoney is source-available under LICENSE. Integrated component trees retain their own license terms under projects/<component>/LICENSE. Third-party materials remain governed by their respective terms.

See SECURITY.md for reporting and operational boundaries.

DumbMoney compounds evidence before it compounds capital.

Architecture · Contracts · Workspace manifest · Security

Name		Name	Last commit message	Last commit date
Latest commit History 18 Commits
.github/workflows		.github/workflows
configs		configs
contracts		contracts
docs		docs
fixtures		fixtures
projects		projects
scripts		scripts
src/dumbmoney		src/dumbmoney
tests		tests
.gitattributes		.gitattributes
.gitignore		.gitignore
CONTRIBUTING.md		CONTRIBUTING.md
DUMBMONEY_WORKSPACE.json		DUMBMONEY_WORKSPACE.json
LICENSE		LICENSE
NOTICE.md		NOTICE.md
README.md		README.md
SECURITY.md		SECURITY.md
pyproject.toml		pyproject.toml
uv.lock		uv.lock

Uh oh!

Folders and files

Latest commit

History

Repository files navigation

DumbMoney

An evidence-native, autonomous multi-agent fund operating system

Product truth

The organism

Roles, skills, tools, and authority

The autonomous loops

1. Market-intelligence loop

2. Sports-to-Kalshi loop

3. Cross-venue thesis loop

4. Portfolio and routing loop

5. Evidence and recursive-improvement loop

6. Security-assurance and recovery loop

7. Model-economy loop

Why this architecture is different

Intelligence and authority are different data types

The fund is a harness, not a committee

Research cannot grade itself

Security is inside the decision loop

Brokerage boundaries are architectural

Recursive improvement compounds evidence, not confidence

Capabilities

A deterministic demonstration

Runtime surfaces

CLI

Read-only API

Repository map

Component provenance

Validation

Operating boundary

License and security

About

Topics

Resources

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages