Historical record · v0.1.0

BABEL v1

The benchmark and replication protocol remain public; they do not validate execution-failure prediction.

archivedEmpirically demonstrated

Does the frozen benchmark support reproducible compositional semantic-failure tests?

Internal and annotation-derived labels did not establish prediction of real execution failure.

Why it was pursued

The benchmark froze compositional semantic-failure tasks for public reproduction and challenge.

Decision date

2026-07-14; status current to 2026-07-17.

No longer claimed

  • The structural diagnostic predicts execution failure
  • Benchmark labels alone validate real-world safety consequences

What survives

  • Frozen benchmark artifact
  • Evaluation and replication protocol
  • Scoped reported benchmark results

Successor records

priority

BABEL v2

The claim is unverified; advancement requires the frozen partner evaluation and held-out gate.

Record