ALTA
Autonomous LLM Trading Asterism
A virtual trading platform operated by specialized LLM agents.
Trade the opportunity. The stock, ETF, or option is only its carrier.
ALTA is an evidence-first research system for public markets. Specialized LLM
agents search independently, turn weak signals into falsifiable opportunities,
argue both sides, compare expressions, pass a separate risk audit, and measure
what happened in a replayable Shadow ledger.
It is built for a question that conventional agent demos usually avoid:
Can a team of autonomous researchers preserve curiosity without giving up
evidence, accountability, recovery, or portfolio discipline?
[!IMPORTANT]
ALTA is experimental, research-only software. It is not investment advice,
a recommendation, an order-management system, or evidence of profitable
Alpha. There is no live-trading mode. The only broker boundary is explicit,
operator-authorized Tiger Paper execution; live credentials and real
capital must never be connected.
中文说明 · Architecture ·
Getting started ·
Research scope · Security ·
Website
Latest: clearer decisions, fresher evidence, safer recovery
- Find the signal: distinguish live, current and expired clues; keep old
research inspectable without promoting it as a fresh opportunity.
- Follow the work: an operating brief, focused opportunity cards, Agent
handoffs and a searchable decision history in English or Chinese.
- Keep your place: retain valid data through disconnects and malformed
responses; prevent delayed polls from undoing acknowledged operator changes.
The five-minute mental model
ALTA treats an idea as a lifecycle, not a chat response:
flowchart LR
S[Sense\nindependent Scouts] --> C[Candidate\nevidence + lineage]
C --> F[Foundry\ndedupe + complete]
F --> D[Debate\nbull · bear · catalyst]
D --> R[Rank\nportfolio-aware]
R --> X[Express\nstock · ETF · option]
X --> A[Audit\nrisk + authority]
A --> E[Execute\nShadow or Tiger Paper]
E --> M[Monitor\nobserve · exit · attribute]
M --> L[Learn\ncalibration + Trader Mind]
L -. new search priors .-> S
The LLMs own open-ended research and judgment. Deterministic code owns the
parts that must remain exact: schemas, lineage, permissions, capital limits,
idempotency, leases, state transitions, recovery, and accounting.
That separation is the design. Agents get room to think; the system never has
to trust their prose as authority.
One opportunity, many minds
| Desk |
Responsibility |
What it must produce |
| Scouts |
Search filings, market structure, price/volume, macro, ownership, options, cross-asset and open-web clues |
A cited Candidate or an honest abstention |
| Foundry |
Normalize, deduplicate and freeze claim lineage |
One stable, falsifiable Opportunity |
| Underwriters |
Build independent bull, bear, catalyst and implementation cases |
Structured disagreement, not consensus theatre |
| Research Director |
Allocate attention and rank competing work |
Decision value adjusted for urgency and portfolio overlap |
| Expression desk |
Compare stock, ETF and option carriers |
Executable payoff candidates with costs and liquidity |
| Risk & audit |
Challenge evidence, sizing, concentration and authority |
Approve, resize, defer, Shadow-only or reject |
| Execution & monitoring |
Revalidate prices, submit authorized Paper intents and watch exits |
Durable intent, fill history, telemetry and attribution |
| Learning loop |
Score forecasts, costs and realized outcomes |
Downside-only calibration and evolving Trader Minds |
No single agent can create evidence, approve its own risk, and move broker
capital. Debate diversity comes from independently routed model roles, while
the final mutation boundary remains deterministic and fail-closed.
What is actually implemented
| Capability |
Current state |
| Opportunity lifecycle |
Durable Candidates, Opportunities, assessments, debates, ranks, expressions, audits, positions and exits |
| Autonomous research |
Multi-lane Scouts with bounded tool use, retries, source health, citation checks and independent abstention |
| Long-horizon continuity |
Due-time scheduling, exact-question follow-ups, thesis deadlines and restart-safe work recovery |
| Portfolio intelligence |
Stress, underlying, catalyst, Alpha-source and systematic-exposure concentration controls |
| Expression selection |
Carrier comparison, price freshness, liquidity, costs, payoff shape and execution reserve |
| Evaluation |
Point-in-time forward outcomes, benchmark-relative Alpha, selection adjustment and forecast calibration |
| Execution |
Shadow by default; gated Tiger Paper-only intents with authorization generation, lease and reconciliation |
| Operations |
One-command console, authenticated API/SSE, live Agent activity, decision ledger, health and control surfaces |
| Resilience |
PostgreSQL source of truth, disposable Redis support state, cycle recovery, orphan cleanup and bounded degradation |
ALTA deliberately does not claim that these contracts produce Alpha. Its
job is to make that claim testable without confusing persuasive model output,
backtests, or Shadow marks with real evidence.
Operator console
The bilingual console follows the entire research-to-position path. It shows
what each Agent is doing, which tools it used, what changed, why a decision was
made, and which authority boundary is active. Secrets are write-only: the UI
can report configured/healthy/expired, never reveal a stored value.
Select any screenshot to open the full-size image.
| System health |
Agent activity |
 |
 |
| Health and research progress. |
Roles, tool calls and handoffs. |
| Decision history |
Opportunity detail |
 |
 |
| Replay the decision trail. |
Review the thesis and evidence. |
Captured from the current production build on September 9, 2026, with the
English interface and explicitly labeled, read-only synthetic preview data.
No real credentials, account details or market-provider payloads are shown.
Agent-authored artifacts retain their original language.
After starting the console, append ?preview=1 to its URL to explore this view
without starting research or placing orders.
The console is an operational lens, not a performance advertisement. Shadow
returns, confidence ranges and model opinions are labeled according to their
actual evidence level.
Start with one command
Prerequisites: macOS or Linux, Node.js 22+, Corepack, Python 3.12+, uv, Docker
Desktop or Docker Engine, and about 20 GB of free disk space.
git clone https://github.com/kyky2347/ALTA.git
cd ALTA
corepack pnpm install --frozen-lockfile
./alta dashboard
./alta dashboard builds the web console when needed and opens it on a stable
loopback endpoint. Re-running the command reuses an installed console instead
of creating a confusing second port. From the console, Start prepares the
locked Python environment, wakes an installed OrbStack or Docker Desktop on
macOS when necessary, starts PostgreSQL and Redis, and brings up the research
backend. ALTA never installs a container engine or gains broker authority on
its own.
For a deterministic first run with no paid providers:
./alta fixture seed
./alta run --cycle-id first-look
Useful operator commands:
./alta doctor # dependencies and safe configuration
./alta status # concise runtime state
./alta service status # autonomous supervisor state
./alta service start # continuous Shadow research
./alta service stop # graceful stop and child cleanup
./alta test # first-party Node/harness suite
See Getting started for provider setup,
environment isolation and clean shutdown.
Modes and authority
Replay fixtures → autonomous Shadow → explicitly armed Tiger Paper
deterministic no broker mutation Paper account only
- Replay reproduces orchestration and state contracts with synthetic data.
- Shadow uses research and market inputs but simulates positions locally.
- Tiger Paper is off by default and requires a valid Paper account,
operator authorization, a current preflight, account binding and fresh risk
approval. Revocation is fail-closed and may enter close-only drain.
- Live trading does not exist. Unknown, stale, duplicated or mismatched
broker state becomes
manual_review; it is never guessed through.
Agent-requested size is treated as a proposal. Final whole-share quantity is
clipped by deterministic risk, liquidity, broker buying power, concentration,
price freshness and current authorization. Every mutation begins as a durable
intent and is reconciled after interruption.
Built for interruption
Long-running research fails in less glamorous ways than a model benchmark:
power disappears, providers time out, processes overlap, and a thesis can wait
weeks for a catalyst. ALTA therefore makes recovery part of the domain model.
flowchart TD
DB[(PostgreSQL\nsource of truth)] --> Q[Claim or resume cycle]
Q --> W[Bounded Agent work]
W --> K[Durable checkpoint]
K --> H[Heartbeat + lease renewal]
H -->|healthy| W
H -->|restart / stale owner| Q
W --> P{Capital mutation?}
P -->|no| K
P -->|yes| I[Persist intent first]
I --> B[Tiger Paper boundary]
B --> Z[Reconcile broker + ledger]
Z --> K
Key properties:
- one claimed autonomous cycle at a time, with stale-owner takeover;
- exact checkpoints instead of replaying an entire Agent conversation;
- monotonic capital authorization checked inside the mutation lease;
- backoff and circuit breakers for optional data sources;
- graceful termination of Agent, App Server and supervisor process trees;
- idempotent migrations and restart reconciliation;
- authenticated control plane bound to
127.0.0.1 by default.
The console retains its last valid snapshot through malformed responses and
connection failures. Late polls cannot overwrite acknowledged operator changes;
cached opportunity freshness is rechecked on every read. See the
local reliability review for verification
scope and remaining limitations.
These are strong local-runtime contracts, not a claim of exchange-grade or
bank-grade availability. Operational assumptions and remaining limitations are
kept explicit in the runbook.
ALTA's research loop is designed to reduce common sources of imaginary edge:
- Every claim keeps source and observation lineage.
- Discovery, debate, expression and audit are separate decisions.
- Point-in-time evaluation uses completed market sessions and frozen forecasts.
- Trial volume is recorded so selection-adjusted confidence can be shown.
- Comparable closes feed forecast and execution-cost calibration.
- Calibration can only reduce new risk; it cannot manufacture expected return.
- Shadow evidence and Paper fills remain clearly separated from live results.
The intended edge comes from breadth of search, independent disagreement,
cross-domain clue synthesis, better expression, disciplined abstention and a
durable learning record—not from asking one model for a ticker.
Verified, not promised
The release baseline remains 0.26.0; newer changes are recorded under
Unreleased. The September 9 local review verified:
| Automated checks |
Browser and integration checks |
| 644 passed: 201 Node · 386 runtime · 38 Paper boundary · 19 dashboard |
84 view/language/viewport combinations in Chromium |
| TypeScript, lint, formatting and production build passed |
7 authenticated read routes through PostgreSQL, Python and the Node gateway |
| Recovery and freshness regression coverage |
Disconnect/reconnect, malformed responses, delayed polls and bounded secure-session retry |
The real-stack smoke was read-only, with autonomous research and Tiger disabled.
Temporary services and test data stores were stopped afterward. These results
do not establish current broker connectivity, multi-day availability, physical
power-loss recovery, universal browser compatibility or profitable Alpha.
See the reliability review for limits and
the publication security audit for scan scope.
Repository map
.
├── alta project-local operator command
├── alta-dashboard/ bilingual React operator console
├── alta-src/ Node gateway, launch control and bounded tools
├── alta-runtime/
│ ├── python/ Opportunity OS, orchestration, API and SSE
│ ├── capital-python/ isolated Tiger Paper-only executor
│ └── compose.yaml loopback PostgreSQL and Redis
├── vendor/openai-codex/ pinned, attributed Codex Rust substrate
└── docs/ architecture, operations, audits and roadmap
PostgreSQL is the system of record. Redis is disposable support state.
ALTA-owned application behavior stays outside vendor/openai-codex/ whenever
practical.
Read next
License and provenance
ALTA-authored source is licensed under the Apache License 2.0.
Vendored and package-managed dependencies retain their own licenses, copyright
notices and terms.
ALTA includes a modified Apache-2.0 snapshot of
OpenAI Codex at a pinned commit as its local
App Server and Agent-harness substrate. Modified files, retained notices,
package dependencies, API-only integrations and conceptual references are
listed in ATTRIBUTION.md and
vendor/openai-codex/CHANGES.md.
ALTA is independently maintained. It is not an OpenAI product, is not endorsed
by any named data or brokerage provider, and must be used only within the
research scope.
ALTA 是一个仅用于研究的多智能体 LLM 交易平台：它把微弱的市场信号转化为有证据支撑的机会，由相互独立的智能体角色进行辩论，再经过一道独立的风险审计，并且只执行模拟交易或经明确授权的模拟盘交易。