← Work · Case 05
An evidence-gated research digest
Collection runs without a model. Writing runs with one. Evidence rules decide what gets in.
StatusProduction
TypeScheduled routine
RoleConcept, design and build, with AI agents
01 · Problem
Problem
Industry news is mostly noise, and AI summaries make it worse: invented quotes, reconstructed articles, confident claims with no source behind them.
02 · Approach
Approach
A scheduled collector gathers RSS, Bluesky and Reddit without any model, de-duplicates against a ledger and saves a run file. A second step writes the digest under hard rules: no source URL, no line; quotes checked against the source; a headline is never expanded into an article.
03 · Architecture
Architecture
04 · Guarantees
Guarantees
- 01No URL, no line.
- 02Quotes are checked against the source text.
- 03Content is never reconstructed from a headline.
- 04The output is a file. Nothing is sent anywhere.
05 · Stack & decisions
Stack & decisions
| Python | A model-free collector that runs on a schedule. |
|---|---|
| Composio | Authenticated access to the sources, with the payload read safely even when it is stored in a file. |
| JSONL ledger | Append-only memory of what has already been seen. |
| Claude | The writing step, under the evidence rules. |
06 · Proof
Proof
- collection runs without a model. writing runs with one.
- feeds scroll away: a day not collected is lost for good,
- so collection never skips. writing can catch up later.
- 1no URL, no line
- 2no number that is not in the source
- 3already in the ledger: not repeated, not even reworded
- 4nothing new: one line, "Nothing new."
- 5a note for us only when it is concrete
- 6source down: "source unreachable". never pretend to have read it.
- always banned: impact scores, confidence %, invented priorities
Excerpt from the real system, translated and sanitized: no names, identifiers or data.