Evidence and provenance
The pre-trade evidence standard: separate facts, calculations, interpretations, and risk rules
A five-layer method for making the evidence behind a pre-trade review inspectable—especially when one data lane is stale, incomplete, or unavailable.
Direct answer
What makes a pre-trade review evidence-ready?
A pre-trade review is evidence-ready only when it names source and freshness, separates fetched observations from deterministic calculations and system interpretations, preserves failed data lanes and warnings, and makes user-defined invalidation, risk, and no-trade rules inspectable. A page load, a successful provider response, or a confident pattern label is not evidence quality by itself.
The five evidence layers
Each layer answers a different question. Keeping the layers separate prevents a derived number from masquerading as a quote, an interpretation from masquerading as a fact, or an application output from masquerading as permission to trade.
A complete review can stop at layer five with no trade. That is not a failed review. It is the correct result when a required source, calculation, limitation, or user rule fails its gate.
What the Lumiere evidence record preserves
Lumiere's compact analysis-evidence record is designed to retain the state needed to inspect an output without treating a screenshot or a polished summary as provenance. The current schema preserves these field groups:
- Schema and time
- A version number and generation timestamp identify the shape and timing of the record.
- Provider state
- Provider id and label, trust or availability state, stale flag, provider generation time, last-tick time, age in milliseconds, and warnings.
- Request lanes
- Separate quote, intraday-chart, daily-chart, news, and local-connector statuses, plus bounded errors and latency.
- Inputs and checks
- Review thresholds, chart frame and session choices, and deterministic pass/fail checks with visible values and details.
- Derived metrics
- Percentage movement, range, volume measures, elapsed-session fraction, and the named basis for relative volume.
- Chart interpretation
- Timeframe, candle counts, active pattern, confidence, invalidation, support, resistance, and selected indicator values.
- News provenance
- Returned and material counts, catalyst reading, source note, and bounded source summaries with publisher, link, time, relevance, and takeaway.
- Warnings and scope
- Deduplicated warnings and the educational disclaimer remain part of the record instead of being discarded after rendering.
This is a bounded evidence object, not a complete market-data archive. It does not prove that a provider was correct, that every exchange or news source was available, that a pattern will persist, or that an eventual order would receive the displayed price.
Source labels and freshness are decision fields
“Connected” and “real time” are different claims. Interactive Brokers documents four market-data types: live, frozen, delayed, and delayed frozen. Its documentation says delayed data is generally 15 to 20 minutes behind, while availability also depends on permissions and the requested instrument. A review should therefore display the source type and time state rather than collapsing every successful response into “current.”
Primary sources, reviewed August 4, 2026: IBKR market-data types and IBKR TWS API documentation. Lumiere is not affiliated with or endorsed by Interactive Brokers.
Provider success also does not mean every requested lane succeeded. A quote can arrive while the daily-chart request fails. News can return zero items without proving there was no material event. A connector can be healthy while the selected source is stale. The evidence record must preserve those differences.
Worked synthetic record: partial data means no trade
This no-symbol example demonstrates record shape only. It is not current market data, a security analysis, a recommendation, or an observed production result.
The fresh last tick does not repair the failed daily lane. The 72% confidence label is an application interpretation, not an externally verified probability. The zero-headline result describes only what that request returned. Because the review requires the daily baseline, the decision record stops even though other lanes succeeded.
AI and system interpretations need an explicit boundary
The SEC, FINRA, and NASAA warn investors that AI-generated information can be inaccurate, incomplete, misleading, or outdated and recommend confirming underlying sources and using multiple sources when making investment decisions. In a pre-trade workflow, that means a model summary or pattern classification should point back to the observable record and remain open to contradiction.
Primary source, reviewed August 4, 2026: SEC, FINRA, and NASAA investor alert on artificial intelligence and investment fraud.
Acceptable interpretation
“The system labels this structure as a possible pattern with 72% internal confidence; inspect the displayed candles, invalidation, failed lanes, and limitations.”
Unsupported leap
“The pattern is 72% likely to win, so the trade is approved.” Confidence is not a verified win rate, suitability judgment, or order instruction.
Protected saved product records and anonymous analytics are different data systems
When a signed-in user saves a protected journal or paper-plan record, that saved product record may include the user-selected symbol, structured analysis, source links, compact evidence, trade-plan fields, outcome fields, and notes. Those fields support the user's account-scoped workflow and are covered by Lumiere's export and deletion practices.
Anonymous operational analytics are narrower. Vercel pageviews and Lumiere's fixed-schema analytics events do not accept Lumiere email, user id, stock symbol, journal text, broker data, private financial-profile data, authentication payloads, or arbitrary properties. Detailed product-usage events require separate Product Analytics consent. Read the Privacy Policy for the current complete statement.
A six-question evidence check before the order ticket
- Name the source.Record the provider and data type instead of writing only “market data.”
- Check time.Compare generation time, last-tick time, stale state, and the timeframe required by the decision.
- Keep lanes separate.Preserve fulfilled, missing, rejected, and zero-result states for quotes, charts, daily context, and news.
- Reproduce the arithmetic.Show inputs and formulas for deterministic metrics and risk ceilings.
- Label interpretations.Keep pattern, catalyst, relevance, and confidence outputs distinct from fetched facts.
- Apply the user's gate.Use prewritten invalidation, risk, execution, and no-trade rules; do not let a preferred position size overwrite them.
Apply the standard