# SAGEN-Bench public release v1

Data, results, instruments, and a standalone verifier for **SAGEN-Bench**,
the pre-registered evaluation of the SAGEN situational-awareness
architecture (paper: https://jakelawrence.xyz/research/sagen).

The study's source repository is private; this bundle is the public,
self-verifying record. Everything deterministic re-derives from what is in
this directory alone.

## Verify (Node >= 20, no dependencies)

    node verify.mjs

Re-derives: the corpus hash, the corpus itself (from the frozen seed), the
S3 knockout lattice (64/64 predicted-in-advance cells), the S1 oracle-mode
coverage table, and the power analysis - and compares each against the
committed results byte-for-byte.

## Contents

- corpus.v1.json + manifest.v1.json - the frozen confirmatory corpus
  (72 machine-generated scenarios; sha256 in the manifest)
- results/ - committed results for the deterministic arms
  (S3 lattice, S1 oracle coverage, power analysis)
- code/ - the pure instrument modules (engine port, event grammar,
  generator, knockouts, probes, prediction matrix, lattice, coverage
  detectors, power model)
- protocol/ - the binding pre-registration document
- SHA256SUMS - hashes of every file in this bundle

## Headline results (deterministic arms)

- S3 knockout lattice: 64/64 ablation-by-capability cells predicted in
  advance and observed as predicted (pilot AND frozen generated corpus).
- S1 oracle-mode mechanical coverage (n = 72): raw buffer 0.160, full
  transcript 0.195, rolling summary 0.268, SAGEN at its 300-token design
  budget 0.937, SAGEN at 2000 tokens 0.966.
- Disclosed negatives: the 300-token budget truncates the injection
  envelope on deep scenarios (machine-parseable-output 0.78); entity
  re-typing is unexercised by generator v1; four rubric dimensions are
  unexercisable by the conversation adapter on scripted input.

License: CC0 1.0 Universal. No human subjects: every arm observes software.