Mira

English | 中文

Agent-native investment research system for evidence-tracked, refreshable investment theses.

Mira is an AI investment research system for agents and research users. It is designed to turn multi-source material into traceable, reviewable and refreshable investment judgment, rather than a one-off stock report or chat answer.

Mira focuses on the investment thesis: what evidence supports it, when it is valid, what would disconfirm it, and how it should be monitored.

Collaboration Model

Mira is built for a three-part division of labor:

• Humans own the purpose, question framing, constraints, risk budget, final judgment and any real-world action.
• AI models expand cognitive bandwidth by reading, summarizing, comparing, extracting claims, generating hypotheses and drafting analysis.
• Mira constrains model output into evidence-tracked, refreshable and falsifiable research work.

Model output is not evidence by itself. Mira's job is to turn useful model work into source-backed claims, explicit uncertainty, refresh conditions and research actions that remain separate from trading or portfolio instructions.

Disclaimer

This repository is for research workflow design, documentation, and historical examples only. It does not constitute investment advice and does not provide legal, tax, accounting, or financial advice. The same disclaimer applies to any AI/agent output produced with this repository, including answers, memos, research packages, and derived analysis based on public sources, user-provided materials, local files, market data, or combined datasets. Treat all such outputs as research support, not advice, recommendations, or verified facts. Public case packages and generated outputs may be incomplete, inaccurate, or stale after their stated cutoff or refresh boundary.

Start Here

If you just want to start using Mira, read START_HERE.md. It is the single user-facing entry card for layered prompts, examples, the help matrix and usage boundaries.

If you open the repository in an agent and only say hi Mira, 你好 Mira, Mira mode, or ask how to use Mira, the agent should return a concise START_HERE.md summary. If the first prompt is already a concrete research task, it should route the task directly instead of interrupting with onboarding.

If you are using a non-Codex / non-Claude product, or you are unsure whether the product will load this repository's rules, use docs/chatgpt-conversation-instructions.md. It is the compact, self-contained entry gate that keeps the agent in Mira Mode instead of answering as a generic assistant, and you can paste it directly when the product cannot read local repository files.

If you want Codex, Claude Code or another code agent to use this repository directly, the main documents are:

Common commands:

scripts/mira_update.sh is the safe update entrypoint: it fetches, refuses dirty/ahead/diverged worktrees, fast-forwards only, and validates the repository after a successful update. Freshness checks do not update the repository.

Quickstart

Minimal workflow:

1. If the user explicitly wants to update Mira itself, run scripts/mira_update.sh.
2. Otherwise, for standard/deep_dive research run scripts/check_updates.sh (local-first by default, 24h remote TTL; quick_map skips it; add --always-fetch to force a remote check now). A blocked fetch degrades to cached local refs — never elevate sandbox permissions for a freshness check.
3. Read OPERATING_CONTRACT.md for the lazy-loading map.
4. Run loops/analysis-routing.md before formal analysis.
5. Load only the routed loop, skill and templates needed for the task.
6. Produce artifacts with evidence logs, time boundaries, refresh conditions and downgraded conclusions where evidence is weak.
7. Validate formal cases with templates/delivery-checklist.md and the relevant script.

For more prompt types, use the Help matrix in START_HERE.md.

What Mira Does

Mira supports:

• first-pass coverage or thesis rebuild for a stock, industry, ETF, macro variable or market theme
• monitoring updates that separate incremental evidence from thesis-changing evidence
• daily and weekly market briefings that separate market snapshot, driver attribution, calendar risk and research escalation
• earnings, SEC filing, macro, industry-concept and ETF specialty analysis
• structured interpretation of sell-side, institutional or user-provided research reports, including claim extraction, valuation assumptions, expectation deltas and thesis impact
• evidence logs that classify facts, claims, assumptions, opinions, market pricing and derived calculations
• user-provided materials, public API outputs and authorized third-party data through a controlled ingestion layer
• thesis system objects such as expectation maps, event deltas, decision logs and postmortems
• position and portfolio reviews when the user provides holdings, weights, mandate or risk constraints
• methodology research for adopting, trialing or retiring reusable research methods

Mira is not a trade bot, market-data daemon or autonomous portfolio manager.

Core Concepts

Task Routing

Formal work starts with loops/analysis-routing.md. Common destinations:

Single-equity research should also run:

• thesis horizon routing
• framework routing
• overlay routing, when a focused evidence path materially improves the answer

Documentation Map

Examples

Prefer these canonical examples before older cases:

• cases/aapl-2026-04/: single-equity research package with memo, evidence log, expectation map, thesis ledger, decision log and actionability bridge.
• cases/nvts-2026-05/: earnings/event package with peer verification and actionability bridge.
• cases/abf-2026-05/: industry concept package.
• cases/etf-discovery-2026-05-09/: ETF discovery example.
• cases/etf-listing-analysis-2026-05-09/: ETF listing analysis example.

All public cases are historical examples and should not be treated as live recommendations.

Validation

Validate the repository or a formal case:

python3 scripts/run_quality_gate.py
python3 scripts/validate_repo.py
python3 scripts/validate_repo.py cases/<case-id>

SEC supplement or filing package validation:

python3 scripts/validate_sec_filing_package.py path/to/sec-supplement-source-note.csv
python3 scripts/validate_sec_filing_package.py path/to/sec-filing-package-dir

More validation and package details: docs/research-artifacts.md.

Project Boundaries

Mira is a research protocol, not an adviser, trade executor or automated data platform.

• No autonomous trading or order generation.
• No position-size or portfolio-construction conclusion without user-provided holdings, weights, mandate and risk budget.
• No unsourced market view should be presented as a conclusion.
• Weak evidence must downgrade the conclusion.
• Real research outputs need source trails, time boundaries and refresh conditions.
• Public examples are historical and may be stale.

Open-source policies:

• License: Apache-2.0
• Contribution guide: CONTRIBUTING.md
• Security policy: SECURITY.md
• Data/source policy: DATA_POLICY.md