Critiques

Third-Party Analyses, Reviewed

Technical reviews of third-party audio-forensic analyses, tools, and the claims built on them. Each entry lays out the method, where it breaks down, and what the evidence actually supports.

truth@evades:~$ forensic-analysis

COMMON METHODOLOGY ERRORS

  • ✗ Confusing echoes/reflections with separate gunshots
  • ✗ Treating visual waveform features as evidence of editing
  • ✗ Ignoring codec compression artifacts in spectral analysis
  • ✗ Applying indoor acoustic models to outdoor environments
  • ✗ Cherry-picking frequency bands to support a narrative
  • ✗ Misidentifying AGC behavior as volume manipulation
  • ✗ Using frame-based video sync for TDOA multilateration
  • ✗ Hardcoding conclusions before running the analysis
ACOUSTIC TOOL CRITIQUE METHODOLOGY ERROR

"calcrack" — A Fit-Checker Presented as a Solver

Jordan Henshaw's calcrack (@jordanhenshawhq) is a Blender add-on that scores how well a user-supplied shooting geometry fits observed crack-thump delays. The physics it implements is correctly derived — but by its own README it "does not itself solve for a solution." The user positions the rifle and target; the tool only evaluates the fit.

✗ Forward model, not an inverse solver — the user supplies candidate positions; the tool only scores the fit
✗ No uncertainty quantification — output is a single delay error in seconds; no confidence intervals, no Monte Carlo
✗ Assumes a single constant bullet velocity; the "Consider Air Drag" mode is an unimplemented stub
✗ Tunable inputs, no objective fit floor — the trajectory and velocity are user-supplied, so a low error reflects the inputs, not the data
READ FULL ANALYSIS →
FORENSIC METHOD EXPLAINER

Crack-to-Bang: Ranging, Not Localization

Crack-to-bang (crack-thump) is a real forensic technique — but it measures a distance, not a location. A single crack-to-bang gap draws a circle around the microphone: it gives range, not direction. Worked through the publicly released Maher/Kester correspondence on the Kirk shooting, plus what holds up under Daubert/Frye.

▸ One gap → one distance → a circle around the mic; locating a shooter needs an independent bearing or multiple synced sensors
▸ Assumes bullet velocity and on-axis geometry — both inject unquantified error in reflection-heavy scenes
▸ Even a leading expert reached only "consistent with ~140 m" from one rough recording — the method's honest ceiling
READ FULL ANALYSIS →
ACOUSTIC TRIANGULATION DEBUNKED

Bray's "Acoustic Fingerprint" TDOA Analysis

Claims to locate a shooter via 5-signature TDOA multilateration across 6 cameras. Cross-correlation reveals 125–321ms sync errors (43–110m spatial error) — 4–5x the camera baseline. Shooter position hardcoded before analysis runs.

✗ Synchronization errors of 125–321 milliseconds — equivalent to 43–110 meters of spatial error
✗ Shooter position hardcoded as a constant in the script before the analysis runs
✗ Cameras producing inconsistent results manually excluded without statistical justification
READ FULL ANALYSIS →
MULTI-DOMAIN ANALYSIS DEBUNKED

The "Epicenter Theory" — Shaped Charge, Decoy Shooter, Ballistic Gel

Claims a micro shaped charge in a microphone battery killed the victim, not a rifle. Fatal internal contradiction: "radial expansion" evidence is incompatible with directional shaped charge jets — the author himself acknowledges this. Flawed optical flow code, circular Strouhal calculation, synthetic "replication" scripts.

✗ Foundational optical flow code misuses physics terminology — detects wavefronts, not epicenters
✗ Published "replication" scripts generate synthetic data from hardcoded values
✗ Bray's own article states cell phone audio cannot reliably distinguish the sources he claims to identify
READ FULL ANALYSIS →
SCRIPT ANALYSIS METHODOLOGY ERROR

Synthetic Signals and Hardcoded Conclusions

Jon Bray's acoustic analysis scripts (followtheepicenter.com) are presented as forensic audio analysis but contain fundamental architectural issues that disconnect the outputs from the evidence recordings.

✗ Generates synthetic test signals rather than analyzing actual recorded audio
✗ Hardcodes shooter positions before "running" the analysis
✗ LLM-generated methodology with fundamental physics errors
✗ Outputs look like forensic results but are mathematically disconnected from evidence

When the conclusion is in the code before the data is processed, it's not analysis — it's illustration.