EXPERIMENT DESIGN / 03
Find the next real lift.
Four testable ideas grounded in the field record. Each has a mechanism, a control, and a rule for deciding whether capture improved.
The hypothesis queue
Plausible mechanisms. Effects unproven.INDEPENDENT RESULTS CHECK
Let the rule decide.
Preview an aggregate result against the frozen rule. This workspace cannot publish utilities, start experiments, or add to the capture census.
What counts as evidence?
- Register before observations begin; use six matched pairs of 06:00–12:00 UTC blocks.
- Enter deduplicated global first-seen counter boundaries. Keep identifiers and request content out.
- Verify coverage, listing visibility, and hold telemetry; declare contamination.
All six paired effects must be positive, with mean lift ≥20/h. Mean lift ≤0 does not support capture benefit. Missing evidence or intermediate effects remain inconclusive.
Imported aggregates are operator supplied. This preview cannot independently verify provenance or establish randomization. Downloading a draft does not register it.
Before rollout: validate the classifier, exclusion rules, randomization, and washout. Passing this preview is not a statistical significance test or proof of causality.
No new hypothesis has been evaluated. Existing T0–T6 findings remain in the experiment record.
THE RECORD WE BUILD ON
T0 / T1
FalsifiedFast stacking and sibling volume did not establish a repeatable capture engine.
T2
Listing recipe supportedA useful pack showed a few hundred nickname-days in prior windows. Its rate is not a growth guarantee.
T3 / T4–T6
Attribution limitedConfounded pack windows, handshake observation, listing holds, and the daily identity clock call for a cleaner test.