Brian's method, and better than v2's whole-record median: the pre-trigger
window is definitionally quiet (the buffer captured before the trigger fired),
whereas a median is merely robust to the event. Requiring the floor to hold
across pre-trigger / middle / end rejects transients that a median cannot.
per channel: pre/mid/end medians, spread = max - min
offset when |pre| >= floor AND spread <= 0.02 in/s
real fault >= 3 consecutive flagged events on that channel
The empirical noise floor justifies the threshold and validates the decoder:
across 19,244 non-flagged channel-events the pre-trigger floor is 62.7% exactly
0.000, 94.5% within +/-1 quantisation unit, median +0.0000, mean -0.0008. There
is no systematic zero-point bias — an independent confirmation of the
32000-count geo scale.
The result is threshold-insensitive across a 2x range (0.020 to 0.040 in/s),
which is what separates a real signal from a tuned one:
FINAL: 5 of 45 units (11%) — BE9558, BE11529, BE12599, BE13117, BE18438
BE11007 and BE10895 drop out; the spread test identifies them as transients
rather than pedestals. v1's 11% headline was right by luck — it included
BE11007 and named the wrong channel on most units.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HgTe8CamXAHcAmaQ6QNcog
Brian challenged the v1 finding that offsets "come and go", against field
experience that a unit which develops one stays broken until the geophone is
replaced. He was right; v1 had two flaws, both of which manufactured false
recoveries:
1. It scored only the axis with the largest peak, so a real event on one axis
hid a persistent pedestal on another. BE12599 on 2026-08-21 read "clean"
because Long had a 1.065 in/s event, while Tran sat at +0.4732 in/s and was
never examined.
2. It used the mean, which a real transient perturbs. The median is the resting
baseline and a blast does not move it. Same event, Long channel:
mean +0.0783 vs median -0.0050.
offset_scan2.py flags a CHANNEL when |median| >= 0.025 in/s (5 A/D counts,
Instantel's own criterion) and treats >=3 consecutive flagged events as the
real signal. No m/p ratio guard is needed — that existed only to compensate for
the mean.
Corrected results:
units with any flagged event 6 -> 19 of 45
units with a sustained pedestal 8 of 45 (18%)
runs >=3 consecutive 29; 1-2 event runs (noise) 69
Also corrected: the affected channel is most often Vert, not Tran (v1 named
whichever axis had the largest peak, so it was frequently wrong). BE10895 and
BE18003 were invisible to v1. BE12599's fault began 2026-08-14, not 08-17.
The decode itself was never in question and is confirmed against Blastware's
own ASCII export: on K558LJN3.BK0W, BW shows Tran parked at +0.265..+0.375 in/s
for the entire record while Vert and Long sit at ~0.005 — Instantel's "parallel
lines above or below the zero line".
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HgTe8CamXAHcAmaQ6QNcog
The DL2 export keeps a byte-identical `Sent/` copy of its root, so walking it
counts every binary twice: 127,035 histogram paths are 63,535 distinct files,
and 13,077 waveform paths are 6,577. offset_scan.py now keeps the first
occurrence of each basename.
Corrects the previous commit's changelog claim of 8 recovered files — it is 4:
K440HJCN.3C0H and K557IF1U.8K0H (stride 252), T191HVNP.0S0H (92), T193L0XM.CI0H
(612). Still zero regressions. The per-unit breakdown reading exactly 2-2-2-2
should have given the doubling away.
The 14,338-exact verification result is unaffected: ASCII exports are not
mirrored (14,340 paths, 14,340 distinct names), and the harness enumerates
those rather than the binaries.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HgTe8CamXAHcAmaQ6QNcog
detect_multi_interval_stride() confirmed a candidate stride on a third block
header whenever the body was long enough to contain one. But a body can exceed
two strides and still hold only two real blocks: a partial final block leaves
trailing padding. BE18193 T193L0XM.CI0H — 51 intervals at 2 s, i.e. one full
30-interval block plus a 21-interval remainder in a 2787-byte body — had every
decisive check pass at stride 612 (header at 0, header at 612, block counter
256 -> 257) and was then rejected for the absent third header at 1224. It
decoded to nothing.
A missing third header now means end-of-stream rather than disqualification.
The block-counter check is untouched — that is the test that prevents the
false positives which once handed 9,082 standard-block files to the
multi-interval walker.
Found by running the full DL2 archive against its preserved Blastware ASCII
exports (14,340 paired files, 11x the previous ground-truth corpus).
Measured over 127,035 archive histogram binaries:
recovered 8 files (strides 92, 252, 612; BE18193, BE18191, BE9557, BE9440)
regressed 0 files
Full-corpus verification: 14,337 -> 14,338 exact of 14,338 decodable pairs
(the 2 excluded are series-4 IDF, a different codec).
Also adds scratch/verify_against_ascii.py (per-sample decoder verification
against BW exports, with a saturation carve-out — BW clamps clipped events to
the range max while the decoder reports true counts) and scratch/offset_scan.py.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HgTe8CamXAHcAmaQ6QNcog
274 series-3 waveform events across 6 episodes on 2 units (BE9558,
BE18438) whose dominant geo axis sits pinned at a DC offset above that
unit's own geo trigger level — the "offset" hardware fault that makes a
unit retrigger continuously and flood the ACH queue.
Detection rule: dominant-axis |mean|/peak > 0.7 AND |mean| >= 0.9 x the
unit's geo trigger level. Bare |mean|/peak is useless on quiet events —
a trace at the 0.010 in/s noise floor clears any ratio threshold.
Not a decode artifact: these reproduce exactly in Blastware's own ASCII
export. Kept as the starting point for the archive-wide analysis.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HgTe8CamXAHcAmaQ6QNcog
The segment-channel scoring analyzer (from scratch/next_experiment_skeleton.py)
ran and immediately confirmed the rotation hypothesis:
SP0 seg 0: best fit Vert 508/508 ✓
SP0 seg 1: best fit Long 508/508 ✓
SP0 seg 3: best fit Tran 508/508 ✓ (Tran continuation)
SP0 seg 5: best fit Long 508/508 ✓
SP0 seg 9: best fit Long 508/508 ✓
V70 seg 0: best fit Vert 508/508 ✓
V70 seg 1: best fit Long 508/508 ✓
Channels rotate Tran → Vert → Long → MicL per 40 02 segment header.
Also discovered the segment header has DOUBLE duty: bytes [14:18] anchor
the NEW segment's channel (2 samples as int16 BE in 16-count units), AND
bytes [0:4] extend the PREVIOUS channel by 2 more samples (2 deltas as
int16 BE). This is the same "2 anchors + delta stream" structure as the
body preamble for Tran.
decode_waveform_v2 now returns full per-channel sample dicts.
Byte-exact verified ranges:
V70: Tran 512, Vert 512, Long 512 (all first segments)
JQ0: Tran 512, Vert 258
SP0: Long 1536 (all 3 L segments)
Still open: the 30 NN block format (high-amplitude packed deltas) —
appears mid-segment when single-byte deltas can't carry the magnitude.
6 new tests bring the count to 46. All passing.
Three things to make pickup smoother:
1. analysis/README.md (NEW): catalogues the ~25 scratch scripts.
Categorizes them as "still useful" / "superseded — keep for
archaeology" / "pure exploration". Tells a fresh engineer which
files to read first and which to ignore.
2. scratch/next_experiment_skeleton.py (NEW): stub + spec for the
segment-channel scoring analyzer. Includes the fixture loader,
block walker, and decode-segment-as-channel helper — just enough
scaffolding that the next pass starts from "fill in
score_segment_against_all_channels()" rather than from scratch.
Already runs and confirms 13 segments per 3-sec event with sample
starts going to 6590 (way past the 3328 actual samples) — strong
evidence that not all segments carry Tran.
3. Removed decode-re/ duplicate. It was a mirror of tests/fixtures/.
Analysis scripts that hardcoded decode-re/ paths updated to point
at tests/fixtures/. CLAUDE.md note updated: future event uploads
go directly into a dated subdirectory under tests/fixtures/.
All 40 tests still pass. Skeleton runs.
serversdownandClaude Opus 5