Equity research report

Equity Research Report

Build a report through an explicit research workflow. Treat an attractive narrative as insufficient unless the evidence and completion gates pass.

Workflow

1. Resolve the instrument, exchange, currency, locale, and report as_of timestamp. Stop if the identity is ambiguous.
2. Publish a concise execution plan containing data collection, specialist analyses, synthesis, critique, and validation. Do not expose private chain-of-thought.
3. Collect one normalized market-data bundle first. Reuse it across specialists instead of fetching the same quote independently.
4. Collect point-in-time evidence for market data, financial statements or filings, valuation inputs, risk history, and current catalysts. Record source, timestamp, and limitations for every item.
5. Run independent technical, fundamental, valuation, risk, and catalyst analyses. Use deterministic calculations for indicators, multiples, returns, volatility, drawdown, and scenarios.
6. Retry only missing or failed evidence, at most twice per gap. Fall back to deterministic analysis when an LLM fails but verified inputs remain available. Mark fallback output degraded and cap its confidence.
7. Run a critic pass. Check unit scale, stale data, contradictory prices, unsupported targets, missing filings, circular citations, undeclared primary-source substitution, unreconciled cross-source contradictions, and conclusions that exceed the evidence.
8. Build the report and its machine-readable manifest. Include the decision trace: plan, tools used, evidence IDs, checks, retries, degraded steps, and unresolved limitations.
9. Run scripts/validate_equity_report.py MANIFEST.json. Claim complete only when it exits successfully. Otherwise label the report partial or blocked and retain the failed checks.

Read references/completion-gates.md before validating or changing completion criteria.

Primary-Source Honesty

These rules exist because of an observed failure: an "verify ROE via the SEC 10-Q" task had every EDGAR fetch fail, silently substituted aggregator sites, and then presented a confident conclusion — twice, with mutually inconsistent equity bases ($106B vs $83B → ROE 115% vs 147%) both labeled as evidence.

1. Declare substitution. If the task names a primary source (SEC filing, exchange notice, official IR release) and it cannot be retrieved, state that at the TOP of the conclusion. Numbers from secondary/aggregator sources must be labeled "not verified against the primary source", and the report cannot claim the verification the task asked for.
2. Reconcile or flag contradictions. When sources disagree materially, never list both as evidence without comment. Identify the basis difference (e.g. TTM-average vs quarter-end equity), reconcile to one number, or report the range with the discrepancy as an explicit open issue.
3. Do not recommend what just failed. A tool that failed repeatedly in this session may only appear in next-step suggestions with that failure noted (e.g. "peer_comparison failed 4x on network errors — retry later").

Tool Routing

• Prefer normalized market-data, filing, factor, risk, news, and web-search tools over model memory.
• In Aria Code, use the market-data service once, then pass its bundle to /team or specialist agents. Use the full team for comprehensive reports.
• Use web search for current filings, investor-relations releases, regulatory events, and other facts that can change. Cite the primary source.
• Never substitute an LLM estimate for unavailable market, filing, or risk evidence.

Required Outputs

Create:

1. A human-readable report with executive summary, company and industry, fundamentals, valuation, technicals, risk, catalysts, scenarios, limitations, and sources.
2. A JSON manifest accepted by the validator. Preserve source timestamps, units, agent status, and evidence provenance.

Keep investment language conditional. Separate observed facts, calculated metrics, model estimates, and analyst judgment. Do not present the output as personalized investment advice.