# Frozen pilot protocol, before inspection of any detection or recovery results

Question: In a 100-target public TESS Sector 1 sample, does an explicit flare/artifact competitor and stricter gap completeness improve recovery at a matched empirical un-injected target-trigger rate? This addresses a pilot subgoal, not the root requirement for labeled astrophysical false alarms.

Sampling: intersect the MIT Sector 1 two-minute target list with the MAST Sector 1 light-curve download manifest. Retain catalog Tmag in [8,12]; order by SHA256('sj-tess-controls-v1:' + decimal TIC ID). In this order retain the first 100 files with header RADIUS in [0.7,1.5] solar radii, TEFF in [4500,6500] K, at least 10000 finite QUALITY=0 PDCSAP flux/time pairs and at least 20 days between first and last. No rejection for variability or positive events. Order selected TICs by SHA256('sj-tess-split-v1:' + decimal TIC ID); first 40 development, last 60 evaluation. Missing or inaccessible files must be retried or recorded as a deviation, not silently skipped.

Input and preprocessing: inject into the archived two-minute PDCSAP flux before any new masking, binning, normalization, or trend removal. This is after the mission's original pipeline, which is NOT rerun. Retain QUALITY=0 finite flux/time. Bin on a 10-minute uniform grid, averaging valid cadences. Normalize by median flux; interpolate empty bins only to estimate a three-day quadratic Savitzky-Golay trend; subtract that trend. Unobserved residual bins contribute zero to convolutions. Estimate white-noise scale robustly from successive valid-bin residual differences (MAD / 0.67448975 / sqrt(2)). This scale is a ranking device; correlated-noise probabilities are not inferred from it.

Baseline: positive matched-filter improvement score max(Z,0)^2 for a bank of small-Einstein-radius, linear-limb-darkened central-chord shapes of durations [3,4.5,6,9,12,18] hours, limb coefficient 0.6. All kernels live in a three-day window, are orthogonal to a constant and linear time, and are normalized in Euclidean norm. Require at least 80% observed bins in that window and no endpoint padding.

Controls: compare against a bank of Davenport (2014) empirical classical-flare profiles with FWHM [0.25,0.5,1,2,4,8] hours, positive single-bin impulses, and positive/negative exponential discontinuity profiles with decay scales [0.5,2,6,12] hours. Use the maximum competitor improvement within six hours of each candidate center. Controlled score = lens score - lambda * competitor score and requires 95% observed bins in the three-day window. Choose lambda from [0.25,0.5,1] by maximum development recovery at a 10% empirical background-target-trigger fraction (ties choose smaller lambda). Also report each candidate lambda's development results and gap-only ablation (lambda=0 with stricter completeness). Do not optimize using evaluation targets.

Injections: one at a time, nine per target, all combinations of white-dwarf mass [0.3,0.6,1.0] solar masses and period [100,300,1000] days. Use host mass RADIUS^1.25 solar masses, circular orbit, impact parameter independently uniform [0,0.8], linear limb coefficient 0.6, and Nauenberg white-dwarf radius. Einstein radius squared = 4GM_lens*a/c^2. Net fractional surface amplification (2*RE^2-Rwd^2)/Rstar^2 times local limb intensity divided by disk-averaged intensity. Apply inside the projected chord; neglect finite-Einstein-radius boundary smoothing. Integrate each exposure with five uniform sub-exposure samples. Draw center uniformly over valid cadences at least 1.5 days from sector endpoints, including those near internal gaps. Per-target RNG seed is the first 16 hexadecimal SHA256 digits of 'sj-inject-v1:' + decimal TIC ID, modulo 2^32. No flares or artifacts are added to the primary sample.

Metrics: each un-injected target contributes its full-search maximum score. Each injection contributes maximum score within max(20 minutes, physical duration/4) of true center; evaluate the un-injected score in the same neighborhood. Primary recovery is newly detected localized pulses: injected score > threshold and local un-injected score <= threshold. Also report raw localized detection without the latter restriction. Threshold is the linearly interpolated 90th percentile of target-level background maxima, with strict >. Primary evaluation is a descriptive equal-background-rate ROC comparison: separately estimate these thresholds on the 60 evaluation null curves, without tuning any detector parameter there. This should give 6/60 triggers absent ties, which must be checked. It is not a deployment threshold or an unbiased estimate of survey false-alarm probability. Secondary analysis applies each development-fitted threshold unchanged to evaluation and reports both achieved trigger fractions and recovery. Report 5% and 20% evaluation operating points as exploratory sensitivities.

Uncertainty: paired target-block bootstrap of 60 evaluation stars with replacement, 2000 draws, seed 20261023; keep every selected star's nine injections together and recompute both background thresholds within each draw. Report percentile 95% intervals for the recovery difference, conditional on the selected development lambda. Duplicated maxima in bootstrap may yield fewer than 10% strict-threshold triggers; this discrete empirical-quantile limitation must be stated. Report recovery by mass and period descriptively, no separate significance claims. No p-values or population occurrence rates will be claimed.

Report all performed analyses, including unfavorable results and all implementation corrections as deviations. Save code, parameter values, external file URLs/digests/sizes, sampling audit, per-injection results, and runnable container environment. This sample neither labels individual background events nor claims an astronomical detection. Known public flare/instrument-event validation remains work for the swarm.