W3 R94/R94-A — original and independent exact floor sensitivity evidence¶
Study boundary: binary bounded originally missing stratum, exact observable masses, known randomized audit invitation, hypothesized EXTERNAL lower response-probability guarantees. Neither an empirical study nor an estimator qualifying real MNAR gaze or whole-grid 90% bands.
Original exact proof (research DRAFT; not merged)¶
| Field | Immutable source |
|---|---|
| Original research PR | #459 |
| Source head SHA | 142d63a551d0e61d776f3b604b5d3cd865b0c2e7 |
| Original GitHub Actions science | #39863137968, 4/4 jobs PASS |
| Original immutable archive | Artifact #12461918445, name: w3-r94-exact-floors-54-record-original-archive |
| Original archived members | r94_exact_evidence.json; SHA256SUMS |
| Dataset size | 54 mathematical cases = 3 nonreturn rates × 3 known invitation probabilities × 6 hypothetical floor assumptions |
| Identified set | For \(p=P(Z=1)\), \(L=\max\{q_1,1-q_0/\ell_0\}\), \(U=\min\{1-q_0,q_1/\ell_1\}\); ratio terms omitted at zero floor |
| Status | Research-only; production_release_ready=false; no package public API, tag, or PyPI change |
Original source contracts explicitly reject missing/tampered factorial evidence and demonstrate exact endpoint models at all feasible intervals. The original archive includes all 9 incompatible cells rather than silently excluding them.
Independent exact rederivation (research DRAFT; not merged)¶
| Field | Immutable audit evidence |
|---|---|
| Separate auditor PR | #460, stacked on #459 |
| Audit source SHA | ff6d8ec2a77452e79a1c6f041e3609996a9be908 |
| Dedicated R94-A workflow | #39863696029, 4/4 jobs PASS |
| Independent archive | Artifact #12462204809, name: w3-r94-A-independent-original-54-case-audit |
| Dependency gate | Three separate Python version contract jobs plus mandatory download and full replay of the actual original source archive, validating original SHA256SUMS |
| Independent mathematics | Allocate unreturned binary \(Z=1\) mass \(t\) and \(Z=0\) mass \(\delta-t\); derive bounds on \(t\) without importing original R94 code |
| Complete replay | 54/54 original archived cell records, 45 feasible endpoint-witness sets and 9 incompatible external assumptions |
| Status | Independent mathematical/source verification, not empirical calibration |
The auditor independently reconstructs every original denominator, exact rational lower/upper, endpoint completion law, boolean guardrail and nonreturn count. It refuses incomplete archives, duplicates, noncanonical fractions, invalid hashes, unexpected fields and missing case combinations.
Figure-to-source ledger¶
| Documentation figure | Source evidence | Denominator |
|---|---|---|
| Exact 99% return assumption ladder | δ=1/100, six prespecified floors and exact sharp bounds, R94 original verified by R94-A | Six deterministic assumption regimes; no simulation |
| Width by return fraction | δ=1/1000,1/100,1/20 and exact common/moderate/asymmetric/strong widths | Three fixed δ values × four illustrated floor assumptions; not independent replications |
The SVGs display exact-source-derived sharp population sets/widths with rounded labels for readability. They are not 90% intervals, Monte Carlo uncertainty bars or estimates of real eye-tracker quality.
Scientific limitations and release restrictions¶
The crucial external lower completion floors are hypothetically imposed: they are not identified by random invitation alone and must not be passed off as measured return propensities. A strong floor equal to the observed overall return probability forces uniform completion in this symmetric binary special case, not in arbitrary field data.
No claim is made for arbitrary functional response mechanisms, real raw-missing gaze recovery, point estimation without those assumptions, finite-sample confidence regions, 90% function-wide W3 slope bands, empirical statistical calibration, public API qualification or v1.2 release.
Keep research PRs #459 and #460 draft/unmerged; issues #289 and #296 OPEN. Stable 1.1.0 and no GitHub/PyPI release unchanged. Publication of these documentation-only pages is gated on full exact-head CI and a separately successful protected-main Pages build/deploy.