W3 R91/R91-A — original randomized completed-recovery evidence ledger¶

Research provenance, independent-audit gate and interpretation limits. This evidence concerns a bounded synthetic finite-population mean and an idealized genuinely complete random recovery process, not an installed package estimator.

Frozen original prospective source study¶

The archive retains cases.json, summary.json, source-commit.txt and SHA256SUMS. Eight separate four-block source shard artifacts retain the entire frozen source experiment. All original source and conditional retrieval records, including unfavorable errors, are preserved. The original source itself is not an independent scientific audit.

Separate numerical verification requirement¶

The separate R91-A draft PR #451 reconstructs the frozen original experiment without importing its source generator or HT implementation. The independent science workflow #39767059171 has now passed 4/4 exact-head jobs: Python 3.11/3.12/3.13 numerical and tamper fixtures and the mandatory actual original 6,144-record SHA256 archive download/replay. Its independently reconstructed artifact #12421700110 is preserved with ZIP digest sha256:bf9cf35cfe080953f58627080ff4d1bec5e8b6306621734b95a84448d8dd754a. It must independently regenerate the original complete X/Y source, selective first-stage and secondary observation masks, all 16 nested Bernoulli completed-retrieval masks, every observed-only HT point and design variance, simulator-only oracle variance and truth, naive estimate, sharp bound width and original fixed-cell summary. It must verify the exact original checkout and every original SHA256 member, and output a distinct independent replay evidence artifact.

Independent original-source numerical verification is complete: original 12/12 source jobs and separate 4/4 independent jobs passed, including complete immutable source audit. These checks verify numerical implementation and provenance, not validity of nominal confidence intervals, functional-slope inference, or field-data assumptions.

Original source-level descriptive errors (independently reconstructed)¶

The source simulator yields the following selected results. Each fixed cell has 32 independent complete-source blocks with 16 dependent conditional recovery repetitions per block. The values below are not source-independent confidence intervals or calibrated uncertainty statements.

The HT conditional unbiasedness formula and variance identities are verified analytically by exact two-record Bernoulli enumeration, independent of this simulation. Sixteen finite recovery draws are insufficient to claim exact zero finite-source error. In these original simulated conditions, HT has larger empirical RMSE than the naive measured-only estimator, despite smaller signed errors in expectation. This is a legitimate bias-variance trade-off, not an error to delete or a general method qualification.

Boundaries and publication gate¶

The only target is the realized finite mean of clip((Y+1)/2,0,1) among originally missing synthetic records. An ideal external validator that truly recovers every Bernoulli-selected record with a known positive inclusion probability is assumed. Random audit assignment with selective noncompletion is explicitly insufficient. The HT estimate need not lie in [0,1], and no confidence interval or functional-band method is provided.

The R91 article, methods companion and two actual data SVG figures are documentation-only publication candidates. The R91-A independent original-archive verification has passed; the separate exact-head 25/25 docs PR gate and protected-main Pages deployment must still pass before any claim of website publication. Both original and independent scientific PRs remain draft/unmerged.

No claim of field MNAR selection identification, general raw gaze recovery, function-on-scalar slope inference, 90% simultaneous-band coverage or production v1.2 release is authorized. Scientific blockers #296 and #289 stay OPEN. Stable public 1.1.0 and API unchanged, v1.2 GitHub Release/PyPI HOLD.