vestige/benchmarks
Sam Valladares c6b642a31a pre-registration: commit the ablation interpretation before running it
The sync arm differs from rag on THREE axes, not one: event anchoring, a
hand-authored causal edge, and a shared write bus (vestige_log). The published
caveats admitted the edge. They did not admit the bus, and run-rag.sh states in
its own header that it has 'NO shared bus and NO coordination'. A reviewer is
entitled to read the current headline as 'coordinated agents beat uncoordinated
agents'. That reading is now named in the repo rather than left to be found.

Three new arms isolate one axis each: rag-bus (coordination only), sync-noedge
(no causal edge), anchor-dense (event anchoring on a plain cosine stack).

The decision table states, in advance, which outcomes demote Vestige and by how
much - including that if anchor-dense matches sync, event-anchoring belongs to
nobody and gets promoted to the headline while the product becomes an
implementation detail. Stopping rule fixed at 5 trials per arm before any spend.

Committed before the experiment. git log on this file versus the results commit
is the check.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-22 21:10:17 +08:00
..
silent-rotation pre-registration: commit the ablation interpretation before running it 2026-07-22 21:10:17 +08:00