Legenda

Limits and checks

This page shows Legenda's current limits. It lists what passed, which conclusion was not supported by its source, and what still needs human review. The point is simple: every public claim should have a receipt.

Current check

Legenda evidence check: 43 case records checked, 42 passed, 1 unsupported inference flagged.

Legenda shows the flagged inference because this is exactly what the report is supposed to catch: a conclusion that reaches beyond its source.

43case records checked
42passed checks
1unsupported inference flagged
0unexpected issues

What this proves

  • The checked corpus still passes the current evidence rules.
  • Source passages, witness scope and supported conclusions are checked every time.
  • Unsupported conclusions are flagged before they can become report findings.

What this does not prove

  • It does not prove raw text intake is fully automatic.
  • It does not prove every language, document type or domain is supported.
  • It does not decide legal truth, motive, emotion, or what every person must have known.

The unsupported conclusion

Example: the source material says a set of keys was left behind on a table. A draft conclusion tried to treat those keys as proof that the person was permanently gone. Legenda flags that as unsupported: the keys may support absence, interruption, or residue, but not by themselves death or permanent departure.

Self-check

A generated test case was turned into text and read back with 81.82% agreement on the core facts.

This is an internal Legenda stress test, not a customer dossier. It is published because the misses show where Legenda still needs care.

81.82%core fact recovery
22source facts tested
18found in all three checks
4core misses

Named misses

  • Two red-cup traces were too weakly recovered.
  • One hesitation gesture was missed.
  • One witness link was missed.

Failure types

The current stress test found checking gaps: one pass weakened a subtle hesitation trace, another flattened object evidence into generic gestures. The checked case records themselves still pass.

Failure classes

Failures are useful only when Legenda can say where they happened.

Source wording gap
The source text did not make a visible fact clear enough to recover. Fix the source wording or demo text.
Checking gap
The source carried the fact, but the check missed or weakened it. The red-cup and hesitation misses are in this bucket.
Record-shape gap
The material contains something real, but the case record has no stable place for it yet. That becomes product work, not copy polish.
Derivation gap
The case record carried the fact, but Legenda failed to derive the supported conclusion. This is the strictest bug class.

What it can and cannot do

Not sentiment analysis. Not an instant upload box.

Can it handle other languages? English and Dutch passages have been tested. Other languages stay named as untested until they have checked artifacts.
Can I upload a technical report or incident file? Not as a fake instant self-service flow. The current offer is concierge: bounded narrative, investigation or interaction material goes through a human-in-the-loop pipeline and returns as a source-backed report.
Is this sentiment analysis? No. Legenda records visible facts, source support, witnesses and exact passages. Public output must point back to the passage it used.
Why not just ask ChatGPT? A summary can sound right. Legenda's claim is narrower: exact source passages, repeatable checks, and independent comparison. If a conclusion cannot point to source support and a quote, it does not graduate into certainty.