An engineering project by Eric Caskey
How Conviction Read works
A model output, its publication time, and a later outcome describe different events. This site keeps those distinctions visible.
I built the market research engine whose outputs appear here. Conviction Read is its public reading surface: a place to inspect the board, follow dated calls, and examine how the record is published.
Research and education, not investment advice. The engine ranks; it has not demonstrated edge net of costs; every historical magnitude is a survivorship-affected upper bound.
Scheduled publication keeps reading separate from analysis
- Model outputs
The engine records scores and dated verdicts.
- Scheduled publication
Exporters prepare the public data artifacts.
- Reader pages
The site displays those artifacts and their source dates.
The engine runs outside this web application and records its scores and verdicts. Scheduled exporters turn those records into public JSON artifacts. A statically exported Next.js site, served through S3 and CloudFront, reads those artifacts in the browser.
Opening a board, research page, or record does not start a new research report. The tradeoff is that you are reading a published snapshot. Its date and any missing data matter.
The Grade form is a separate API lookup of an already-computed score. It asks for an email and has a daily request limit. It does not run a fresh analysis. The published board and full record can be read without an email.
An export timestamp does not make an old score current
Reader pages distinguish the source date from the time an artifact was exported. Exporting yesterday's verdict again does not reset its source age. A missing date remains unknown; a dated saved snapshot is identified as a fallback.
Research commentary also keeps its original date and score context. A current quant score does not turn an older research note into a new analysis.
Outcomes have their own coverage
Outcome follow-up happens after a call has been recorded. The original ledger's outcome attachments and the independent follow-up are separate sources, with different coverage and methods. The record labels both instead of combining their totals.
A repeated call on one company is another observation, not another independent trade. Pending outcomes remain pending. A stock's price change is not a portfolio return, and the published follow-up does not establish returns after trading costs.
What you can verify
Read the dated record or download the published JSON. Original verdict fields are preserved by the publication contract. Outcomes can be attached later, and latest-factor fields can change. The entire live file is therefore not an immutable object.
The archive manifest lists snapshot files and their SHA-256 hashes. A matching hash confirms that the bytes agree with the served manifest. It does not independently prove historical continuity: both could be replaced. An independently retained copy gives you something to compare across time.
Changes made while operating the site
Score explanations now distinguish the 0-to-100 model score from the 0-to-1 conviction measure. Freshness checks use source dates instead of treating a new export as new evidence. Outcome summaries label their own denominators so readers can see which calls were measured.
These changes improve what a reader can infer from the interface. They do not change the historical calls or establish that the model predicts returns reliably.
Working software and investment evidence answer different questions
The pipeline and published artifacts show how the system operates. The measured outcomes determine what can be said about its predictions. A running pipeline and a long list of calls do not, by themselves, demonstrate a tradable edge.
Historical measurements have data limitations. Coverage differs by source and horizon, observations can repeat the same companies, and many calls have no mature outcome. Keep those limits alongside any number taken from the site.
Questions about the implementation
Contact Eric Caskey about the architecture or publication method. For a short project description and downloadable snapshots, see Project facts.