isolated agent laboratory · evidence before promotion
Test the hypothesis.
Measure the delta.
Keep production clean.
Hermanity Lab turns isolated experiments on Bunsen into readable, falsifiable evidence. Every candidate stays off Herman until its numbers, tests, and promotion boundary are explicit.
BOUNDARY ACTIVE 0 Herman runtime mutations
SPECIMEN B-01 · BUNSEN
hostisolated lab VMmemoryHoncho disabledphase01 · measured
// phase 1 notebook
What the lab found
performancecandidate
Parallel npm audits cut doctor wall time in half
11.99 s median
52.6% faster
10/10 benchmark runs; focused 3/3; full suite red/stalled at 83%
7b1461cde013…
reliabilitypass-with-observation
Twenty-five isolated lifecycle cycles leave no survivors
No process growth · 0 survivor cycles
gateway stop max 1.24 s
25/25 lifecycle safety; workload compliance 22/25
779ec6a5a7a2…
toolingmeasured
The web tool works; exact-format obedience is the weaker link
10/10 factual · 9/10 exact format
Median 35.49 s
10 isolated fetches; corrected secret scanner PASS
934787131af7…
identity safetycandidate
Content fingerprints close a rapid-config identity gap
SHA-256 content key detects same-size, same-mtime edits
424 passed
10/10 synthetic identity routes plus full Honcho plugin suite
a6acf18bd24c…
model routingpass
Fallback works and the usage receipt names the served model
Fallback to minimax-m3 in 25.95 s
Canonical config hash unchanged
Primary, direct fallback, and forced fallback all passed
096e57bdea1a…
source hygienehold-and-promote
Seven local patches were sorted before promotion
7 apply to fork and upstream; 0 already landed
65 targeted tests passed
TTS/footer candidate; cron toggle held for repair
26ab28ebedb6…promotion boundary
Candidate ≠ installed.
Doctor and Honcho fixes are reviewable candidates. Source patches are sorted. None has permission to change Herman. Jon’s explicit approval remains the final gate.
See the six-part method