Refrozen repository snapshot containing the treatment actually verified live.
GOOSE EVIDENCE LAB / 2026-09-15
Evidence
before advice.
Goose does not turn every SEO or AI-search observation into a rule. This lab keeps provider guidance, independent research, winning pages, owned-site experiments and counter-evidence separate.
Current truth: the current Ptichi cohort revision has exact production parity and is ready for bounded observation. No ranking, citation or acquisition result is claimed yet.
- External records
- 14
- Treatment
- 12
- Controls
- 6
- Frozen queries
- 12
- Measured wins
- 0
LIVE EXPERIMENT / PTI-DA-001 R2
Production parity is proven. Keep the cohort frozen.
Exact source SHA recorded by the reviewed Ptichi production evidence.
The exact refs match. Bounded observation may proceed without changing cohort membership or query text.
The original 2026-09-09 freeze remains historical HOLD evidence because its exact implementation was never verified live. Before the first later verified publication, treatment content changed while the 12 treatment / 6 control assignment and 12-query panel stayed fixed. R2 binds that unchanged design to the exact production snapshot 13b49e9a…. No post-deployment Search/AI outcome was used to select or rewrite the cohort, and later undeployed source changes are outside this observation state.
12 treatment entities
Existing useful Ptichi pages only. No indexable pages were manufactured to reach a generic sample size.
6 unchanged controls
Comparable published entities stay outside the treatment during the observation window unless a correctness fix is required and recorded as contamination.
14 / 28 / 56 days
Review windows remain fixed. Provider-native evidence is interpreted separately from deployment truth, and immature windows remain HOLD for outcome decisions.
View the frozen 12-query panel
q01-explain-clearlyhow to explain something clearly at workq02-explain-technicalhow to explain technical ideas to non technical peopleq03-project-statushow to give a concise project status updateq04-recommendationhow to present a recommendation at workq05-difficult-questionhow to answer a difficult question at workq06-project-delayhow to explain a project delay to stakeholdersq07-project-riskhow to communicate project risk to stakeholdersq08-uncertaintyhow to communicate uncertainty at workq09-before-meetinghow to rehearse what to say before a work meetingq10-stop-ramblinghow to stop rambling at workq11-speak-conciselyhow to speak more concisely in meetingsq12-recorded-voicewhy does my recorded voice sound different
LEARNING SYSTEM
Observe. Challenge. Retire.
Adding recommendations is easy. Goose is designed to also weaken or remove them when evidence changes.
External evidence
Separate provider requirements from independent observations and expert experiments.
Winner Observatory
Track entrants, drops, top-10 persistence and citation persistence for frozen query cohorts.
Controlled cohorts
Freeze treatment, controls and queries before outcomes are visible, then block measurement until production parity.
Recommendation review
Move advice through fresh, review-due, challenged, contradicted and retire-candidate states without silent registry mutation.
EVIDENCE IS NOT ONE SCORE
Keep the stages apart.
A passing implementation is not a deployment, a deployment is not exposure, and a citation is not a useful visit.
Implementation
Did the repository change exist and pass checks?
Deployment
Is that exact implementation live?
Exposure
Did Search/AI systems expose the page?
Citation
Was the page cited, and did the citation persist?
Visit
Did identifiable traffic arrive?
Useful action
Did the visit lead to a relevant product action?
EXTERNAL EVIDENCE / REVIEWED 2026-09-09
14 records. Different jobs.
Official platform documentation is strongest for provider requirements and controls. Observational studies and expert experiments can generate hypotheses, but they do not expose hidden ranking weights.
Best suited to provider requirements, eligibility, controls, reporting semantics and explicitly documented recommendations.
Useful for prevalence, association, drift and cohort behaviour; does not establish causality.
Useful for a narrowly specified feature behaviour when setup and limitations are inspectable; not automatically generalizable.
The full dated research record remains the source of truth: knowledge/research/2026-09-09-external-evidence-winner-loop.json.
RECOMMENDATION LIFECYCLE
Advice can decay.
fresh→review-due→challenged→contradicted→retire-candidate
Source age, provider changes, winner counterexamples and owned-site negative evidence can all create review attention. No event silently rewrites the recommendation registry.
INSPECT THE METHOD