Check the actual data first
Download the replay ZIP and unpack it. The two HDF5 files are the GWOSC O1 preliminary version 2 records. Each has 131,072 unitless strain samples at 4096 hertz, beginning at GPS 1126259446. The manifest records file lengths and SHA-256 values. Read the DATA, CBC CAT1 and no compact-binary hardware injection bits by name. Livingston’s separate continuous-wave flag is not set. It should stay visible in the record. GWOSC data release · Data-quality definitions
Fit the two sites with one template family
Use Python 3.12.12 with the pinned numerical packages listed in the README. From the unpacked directory run the fitting script. The noise average uses 25 off-event windows. Apply the same passband to both the measured and fitted strain and subtract them in white units.
python -B prototype.py . new-fit-report.json
The conditional maximum in the saved grid gives scores about 19.36 at Hanford and 14.20 at Livingston. The root-mean-square whitened trace over the short display window falls from 0.665 to 0.342 at Hanford and from 0.544 to 0.350 at Livingston. Use the same window for both sides of each comparison. Over the full record, the whitened squared energy removed equals the squared single-site score under these conventions. The separate time-domain dot calculation agrees with the FFT output.
Recover the tutorial values by their own convention
The original exercise uses a different noise average and the nominal midpoint of its template file as time zero. Recompute it separately.
python -B tutorial_reference.py . new-reference-report.json
The independent replay gives 18.56 and 13.21. Hanford’s timestamp is GPS 1126259462.439453 in that convention. The time of peak template amplitude comes about 16 milliseconds earlier than the file midpoint. Comparing differences between the detectors preserves the roughly seven-millisecond delay. Pinned GWOSC notebook, matched-filter section
Try a deliberately bad fit
At Hanford choose the mistimed template. Notice how the residual grows even though the measured data never change. Repeat at the low end of the mass grid and move the template back to its own maximum. Changing arrival time cannot remove the difference in the spacing of the cycles. For a quantitative check, save both analyses and compare the residual samples at identical array indices.