Skip to content

Latest commit

 

History

History
57 lines (42 loc) · 1.98 KB

File metadata and controls

57 lines (42 loc) · 1.98 KB

Activation Dogfood Usefulness - 2026-06-28

This report closes the bounded #2857 probe for activation usefulness. It asks a narrow question: when replay signal is available, does activation reduce manual search, wrong-route drag, or noisy surfacing on a small dogfood/replay slice without claiming live default quality?

The answer is yes for this bounded replay. It is not a Dream/default-foreground promotion.

Command

python benchmarks\aippocampus\benchmark_activation_dogfood_usefulness.py --json

Focused regression coverage:

python -m unittest tests.aippocampus.test_benchmark_activation_dogfood_usefulness -v

Result

Metric Cold no activation Warm replay signal Delta
Generated candidates 2 5 +3
Foreground exposed candidates 2 2 0
Verifier-seen candidates 2 5 +3
Useful source-open hits 1 2 +1
Source-open follow-through 1 2 +1
Manual-search fallback 2 1 -1
Wrong-route drag 1 0 -1
Noisy surfacing 1 0 -1

Warm replay roles: process_supervision=1, hard_negative=1, replay_sample=1, and positive_demo=1.

One case consumes a replayable live-agent trajectory packet with complete ordered source reopen. The other cases cover route-feedback replay, parked candidate lifecycle, and source-openable route preservation.

Decision

activation_probe_useful_on_bounded_replay=true.

The probe supports keeping activation signals in the bounded replay/dogfood pipeline because they reduce manual search and route drag while preserving source-open follow-through. It does not promote live default foreground activation, Dream default delivery, source truth from activation, or private history quality.

Boundary

This is a deterministic dogfood/replay usefulness probe. It does not claim causal real-user lift, live default foreground quality, private-history quality, default Dream delivery quality, or factual support from activation signals.