Skip to content

Research Demo

This demo shows a realistic research-analyst agent trace moving through Lerim's compiler so the next research run starts with the source-quality rules, the rejected leads, and the cited conclusions instead of re-deriving methodology from scratch. The trace and the extracted records below are real: the output was captured by importing the checked-in example through Lerim's research profile.

1. The completed run

A research agent assessed the competitive position of the EU neobank sector for Q2 2026. The valuable signal is not the report itself; it is the standing source-quality rule, the rejected uncorroborated lead, the coverage boundary, and the confirmed filing-backed conclusion.

The trace lives at docs/examples/traces/research-analyst-run.jsonl.

2. Import it

lerim trace import docs/examples/traces/research-analyst-run.jsonl \
  --source-name research-agent \
  --source-profile research \
  --scope-type domain \
  --scope research

3. What Lerim kept

Six records were extracted: three durable constraints (a source-quality hierarchy, a rejected-lead rule, and a coverage boundary), one confirmed fact, one analyst preference, and one archived episode for the research session.

Constraints (active — these govern future research):

  • Source-quality hierarchy for quantitative claims — ESMA filings and SEC equivalents are preferred over press releases and blog posts for confirming quantitative claims such as funding figures.
  • TechCrunch funding figures require regulatory corroboration — TechCrunch funding figures should not be used for EU neobank sizing unless corroborated by a regulatory filing. A single unnamed source with no filing corroboration was rejected.
  • Crypto-native institutions excluded from EU neobank coverage — Crypto-native and stablecoin-first institutions are excluded from the EU neobank coverage set unless explicitly requested; they distort funding-trend comparisons.

Fact (active, cited):

  • Q2 2026 EU neobank funding declined approximately 35% year-over-year — Driven by tightened MiFID II marketing restrictions. Confirmed by three ESMA-filed prospectuses. Evidence: line 4.

Preference (active):

  • Benchmark EU neobank valuations against traditional bank multiples — Not against US fintech multiples. US-fintech comparables overstate value because they ignore MiFID II capital constraints.

Episode (archived — history, not standing context):

  • The EU neobank funding research session itself.

Inspect them yourself:

lerim context records --source-profile research

4. What the next agent now knows

Before this trace was compiled, the next research run on EU neobanks would likely re-cite the uncorroborated TechCrunch funding figure, benchmark against US fintech multiples, and include crypto-native institutions in the coverage set — all of which this run rejected. After compilation, the source-quality hierarchy, the explicit coverage boundary, and the filing-backed conclusion are available before work begins.

5. The improvement loop

The source-quality rules and cited conclusions above are exactly the kind of approved, evidence-backed data for a smaller, private research-analyst model. The open core captures and cites; model specialization (distillation, RL, prompt and harness tuning) is the private Lerim layer, built on top of this open foundation.

Notes

  • Lerim is selective by design. The first extraction of an earlier version of this trace produced zero durable records because the methodology rules were stated only as in-the-moment narration. Adding one explicit "going forward, we should not use TechCrunch figures unless a filing corroborates them" line caused the rules to be captured as standing constraints. Routine sessions can and should produce zero records; that is the compiler rejecting noise.
  • Replace the checked-in example with your own cleaned research-agent source session for real evaluation.