# Designing a Financial Research Workbench Before Choosing a Provider

**Public-candidate active-development protocol — review required**  
**Architecture and evaluation design; not investment advice, provider selection, or performance evidence**

## The decision before the data

Market-data APIs make it easy to retrieve prices and fundamentals before deciding what those values mean, whether they were available at the simulated decision time, whether they may be retained, or whether they can legally be shown to another person.

This candidate protocol reverses that order. It defines the evidence and authority requirements before choosing a provider or building a trading workflow.

## Provider eligibility

A provider must be evaluated against current official documentation and contractual terms for:

- stable security identity and ticker history;
- price adjustments, dividends, delistings, and corrections;
- GAAP/non-GAAP definitions and filing/availability dates;
- historical universe membership and survivorship risk;
- attributable event sources and corrections;
- latency class;
- local retention and model-processing rights;
- redistribution rights;
- rate limits, pagination, outages, and backfills;
- cost and exchange or user constraints; and
- credential, logging, telemetry, export, and support boundaries.

No named provider is selected by the protocol. Candidate providers remain discovered until their current official terms and behavior are reviewed.

## Immutable source and derived lineage

Every permitted retrieval should produce an immutable source record containing provider and dataset identity, retrieval and effective timestamps, security identifier, checksum or licensed equivalent, schema and ingestion versions, use classification, latency class, correction lineage, and success or failure details.

Derived values must reference their source records and transformation versions. Missing point-in-time support cannot be repaired by model inference.

## Point-in-time discipline

A historical replay must not use information before its recorded availability. Restatements remain distinct from originally available values. Universe membership must be reconstructed at the simulated date or labeled survivorship-biased. Corporate actions, identifier changes, bankruptcies, and delistings remain visible.

The protocol begins with a small candidate slice rather than claiming a complete market universe. Even the proposed market-cap threshold remains a hypothesis to be evaluated at each observation date.

## Historical replay

The candidate sequence is:

1. freeze the simulated decision timestamp;
2. load only records available by that time;
3. generate research objects independently where independence is part of the test;
4. preserve prompts, runtime identities, sources, outputs, and failures;
5. synthesize disagreements without erasing them;
6. require a human or pre-authorized experimental rule for any paper order;
7. apply documented cost, spread, slippage, and execution assumptions;
8. compare against named benchmarks and a simple baseline;
9. wait for the defined evaluation horizon; and
10. append corrections without rewriting the original decision through hindsight.

## Evaluation has three layers

### System integrity

Provenance completeness, stale/conflicting data detection, point-in-time violations, identity and corporate-action errors, jurisdiction violations, reproducibility, and correction closure.

### Research quality

Falsifiability, source coverage, contradiction handling, confidence calibration, useful disagreement, stability under irrelevant context, and source-faithful explanation.

### Paper outcomes

Returns relative to named benchmarks, drawdown, volatility, turnover, concentration, hit rate, holding period, transaction-cost sensitivity, regime behavior, and failure attribution.

No single metric grants trading or production authority.

## Human boundaries

- Alex separately approves provider purchases and licensing commitments.
- Provider eligibility does not authorize live trading.
- Model output does not authorize an order.
- Initial evaluation remains paper-only.
- Client-facing or redistributed data requires explicit rights.
- Results remain separated from interpretation and correction history.

## Current state

This is an active-development protocol. No provider has been selected, no complete historical trial has been established by the source document, and no investment outcome is claimed. Its present value is architectural: it makes provenance, licensing, temporal validity, evaluation, and human authority prerequisites rather than cleanup tasks.