AlphaGuard is an automated system for producing auditable research artifacts, not an unstructured content generator. Every publishable output must therefore have two synchronized forms:
- a human-readable Markdown report built from an approved template;
- a machine-readable JSON sidecar validated by
src/output_standard.py.
The Markdown file communicates the research. The JSON sidecar proves what was claimed, which evidence supported it, which kernel evaluated it, and when the underlying data was current.
Use for event-driven earnings, filing, market-microstructure, supply-chain, or
model-behavior anomalies. The expected reading time is five to ten minutes.
Start from templates/telemetry_note.md.
Required sections:
- document control;
- executive summary;
- AlphaGuard performance scorecard;
- technical deconstruction;
- forensic evidence;
- strategic implication;
- methodology and audit trail;
- disclosures.
Use for multi-company, cross-market, or quarterly thematic research. Start
from templates/deep_dive_whitepaper.md.
Required sections:
- document control;
- investment question and scope;
- executive findings;
- evidence base and methodology;
- market or industry architecture;
- company-level forensic findings;
- valuation or scenario analysis;
- risks, disconfirming evidence, and invalidation conditions;
- audit trail;
- disclosures.
Section order, naming, document control, evidence labels, and disclosures are
fixed. Company names, data, equations, findings, and scenarios are dynamic.
Sections may be marked NOT APPLICABLE, but required sections may not be
deleted.
This fixed shell gives institutional readers a stable retrieval pattern while preserving analytical flexibility.
- Use restrained Markdown headings, tables, blockquotes, equations, and code.
- Do not use emoji, decorative icons, engagement bait, or promotional slogans.
- Use direct institutional English. Separate observation from interpretation.
- Use ISO-8601 timestamps with explicit timezone offsets.
- State currencies, units, fiscal periods, and market sessions explicitly.
- Label synthetic fixtures, estimates, scenarios, and backtests as such.
- Never present model output as verified market fact.
- Never publish a price, yield, failure rate, valuation, or performance claim without a source reference and as-of timestamp.
Every material assertion in the JSON sidecar must be assigned one type:
| Type | Meaning | Minimum evidence |
|---|---|---|
FACT |
Externally observable or reported value | At least one source |
INFERENCE |
AlphaGuard interpretation of evidence | Source and methodology |
SCENARIO |
Conditional forecast, valuation, or strategy case | Methodology and assumptions |
Facts and inferences must not be blended into one sentence. A statement such
as "revenue was USD 10 billion, proving the equity is undervalued" must become
two claims: one FACT and one INFERENCE.
Each source requires:
- a stable
source_id; - a
registry_document_idpresent inconfig/source_registry.json; - the exact registered title and HTTP(S) URL;
accessed_atmatching the registered snapshot timestamp;- the registered SHA-256
content_hash; - a precise page, section, table, exhibit, paragraph, cell-range, timestamp, line-range, or JSON-pointer locator.
The complete registration and locator contract is documented in
docs/SOURCE_REGISTRY.md.
Preferred evidence order:
- regulatory filings and exchange disclosures;
- company investor-relations material;
- official statistical or regulatory datasets;
- primary market-data feeds;
- reputable secondary research.
Social-media evidence may support behavioral analysis but must not be treated as a substitute for filings, transaction data, or official disclosures.
Every published finding must retain the shared BaseAuditor fields:
{
"kernel_id": "Module_15_SCGV",
"target": "AVGO",
"rigor_score": 2.1,
"data_quality_status": "REJECTED",
"structured_written_feedback": "The proposed reorder point omits lead-time variance."
}Module-specific evidence may be added. It may not replace the shared fields. Raw scorecards must not be rewritten to make a narrative more persuasive.
The canonical schema is
schemas/publication_artifact.schema.json. Runtime validation is implemented
in src/output_standard.py.
Schema version 1.1 requires registered document snapshots, SHA-256 hashes,
and exact citation locators. Version 1.0 artifacts must be migrated before
they can pass the current publication gate.
Validate an artifact before publication:
python3 src/output_standard.py examples/publication_artifact.example.jsonRequired top-level fields:
schema_version;artifact_id;document_type;title;as_of;generated_at;telemetry_timestamp_matrix;engine_version;classification;executive_summary;claims;sources;scorecards;disclosures.
An artifact is publishable only when all gates pass:
- repository validation succeeds;
- every scorecard satisfies
BaseAuditor; - every
FACTcites at least one known source; - every claim cites methodology;
- timestamps include timezone;
- required disclosures are true;
- synthetic data is explicitly identified;
- no placeholder values remain in the final report;
- Markdown and JSON use the same artifact ID and as-of time;
- a human reviewer confirms that the prose does not overstate the evidence.
- every cited document resolves to the source registry and an exact locator.
The timestamp matrix must satisfy
docs/TEMPORAL_STANDARD.md. All three local market
clocks must reconcile to one UTC instant, and that instant must equal the
publication's as_of timestamp. A weekday schedule projection must not be
described as an exchange-verified holiday calendar.
A subscription gateway may separate public and premium sections, but it must not hide required disclosures, manipulate scorecards, or imply guaranteed returns. Strategy parameters behind the gateway remain subject to the same source, timestamp, scenario, and risk-disclosure rules.
Every report must state that China Alpha Dispatch is an independent research entity, that the material is for informational purposes, and that derivatives or options involve significant risk where applicable. Nothing in the output contract converts research into individualized investment advice.