23 October 2026 — W3 evidence update¶
Research notes · Independent mathematical audits · Not a package release
What changed¶
The programme progressed from empirical stress-testing of functional confidence bands to a precise identifiability question. In an originally missing binary-outcome stratum, what can be learned about unreturned outcomes using observed audit returns, randomized invitations and external information?
- R93 · Same observed law, different hidden populations
Even near-complete response does not force point identification when completion depends on the unreturned binary outcome.
Exact observational equivalence
- R94 · Guaranteed completion floors
An externally defensible lower bound on each outcome's completion probability can conditionally narrow sharp latent-mean intervals. Incompatible floors must be rejected.
- R95 · Know the direction—or admit you do not
A known direction of selective completion can halve a symmetric identified set; monotonic selection without a known sign supplies no reduction.
Interpretation: these are conditional identified sets, not confidence intervals. R95 comprises 54 deterministic rational cases, not 54 independent datasets. Its original PR #462 and independent PR #463 remain draft/unmerged. R95 original and audit ledger.
What did not change¶
- Published stable package: 1.1.0, Python 3.11–3.13.
- Unpublished development source: 1.2.0rc2.dev0, not a separately published qualified literal RC2.
- Research blockers: #289 (simultaneous bands) and #296 (observation-dependent gaze).
- GitHub Release / PyPI for 1.2: on hold.
The site and research evidence are intended to expose unqualified or negative results clearly—not to turn them into advertised stable functionality.