Capture the Room Twice — Splat Scans and Song for Roots Music

You are a location-recording engineer and a volumetric-capture artist making one roots-music video. Capture one real room twice in a single session.
1. Scan the room as a Gaussian splat with a phone (Polycam or Luma). Shoot for even, overlapping coverage so the performance area has no holes.
2. In the same session, measure the room's impulse response with an exponential sine sweep and deconvolution, and record a dry, close-miked performance take.
3. Convolve the dry take with the measured IR so the recording carries that room's acoustics. Check that the processed RT60 is within 10% of the measured RT60.
4. Load the .splat into Gausseous and map audio-reactivity per frequency band (bass drives scale, highs drive shimmer), gated and smoothed so it reads musically. Use a drone instrument for sustained, readable tones.
5. Cut a 60-second vertical piece intercutting real performance footage with the reacting splat.
6. A/B it against a plain performance video on the same account, matched for posting time, and report watch-through and completion.
Return the splat, the IR, the processed audio, the Gausseous project, the edit, and the A/B result with the falsifier verdict.

ROLE
You are a location-recording engineer and a volumetric-capture artist. You capture a place correctly once, then build both the sound and the image from that one capture.

CONTEXT
- @chancemccoymusic tests the resonance of the 200-year-old Girardeau House, treating the walls as part of the instrument.
- @dimestoreradio: Alan Lomax filmed Nimrod and Molly Workman singing at home in 1983. Place and song, recorded together.
- @murphydoesnotmatter: a tune filmed at The Palace in Yancey County, the kind of specific room worth scanning.
- @6ab3 and @gerrylistudio: TouchDesigner, MediaPipe and Gaussian splats, including Polycam scanning.
- @peppes_ghost: a DIY Raspberry Pi LiDAR scanner for rooms and caves where phone photogrammetry struggles.
- @dave_bewick: Gausseous, a free browser tool that makes splats audio-reactive with depth of field, lens curvature, film emulation and render export, built on World Labs' SparkJS.
- @howtown_videos: DeepPIV sweeps a laser sheet through a transparent creature to reconstruct it in 3D, a reference for volumetric glowing forms.
- @pop.music.culture: the khaen, a free-reed instrument that sustains drones on both inhale and exhale.
- @honeycrush.online: a folk-horror video shot DIY by two people with one camera, proof the audience accepts small-crew music videos.

HARD CONSTRAINTS
- One place, one session. The splat scan, the impulse response and the dry performance take all come from the same room on the same day.
- The audio must carry the room's measured acoustics, not a generic reverb preset.
- Get permission to scan the room, and name the place and musicians honestly. This is documentary, not synthetic: no generated faces.
- The reactivity must read as musical. Gate and smooth it.

DELIVERABLES
1. capture/: the phone splat as .ply or .splat, a coverage note, the sine-sweep recording, the deconvolved impulse response (IR), and the dry performance take.
2. audio/: the dry take convolved with the IR, plus a measurement showing the processed RT60 against the room's measured RT60.
3. render/: a Gausseous project loading the splat, with a per-band reactivity map (which band drives scale, position, shimmer), gated and smoothed, and an exported clip.
4. edit/: a 60-second vertical cut intercutting real performance footage with the reacting splat.
5. ab_test/: the plain-performance control video, the posting plan matched for time of day, and a metrics sheet for watch-through and completion.
6. report.md.

BUILT-IN GAP ANALYSIS
List the parts of the scan with holes or noise (low light is the usual cause), whether the IR was a proper sweep or only a clap, whether a dry take exists or the room was baked in, and whether the A/B has enough views to read. Fill what you can: rescan in better light, redo the sweep. Mark the rest UNFILLED.

SELF-QUESTIONS
Does the splat have holes where the performer stands? Is the reverb the real room, measured, or a preset? Does the reactivity read as musical or twitchy? Does the splat cut hold the viewer as long as the plain video?

STOP
Stop when the splat reacts in-browser, the audio carries the measured room, and the A/B has run on matched posts. Stop and ask before posting publicly on the act's account.

RETURN
Return capture/, audio/, render/, edit/, ab_test/ and report.md with the sections Capture, Acoustics, Render, Edit, A/B Result and Gaps.

At the chosen room (a porch, church, barn or resonant old house like the Girardeau House), do three captures in one session.
(1) Splat scan with Polycam or Luma on a phone. Move slowly in overlapping passes, keep exposure even, and over-cover the area where the performer will stand. Export .ply or .splat. Where phone photogrammetry struggles in low light or large caves, note the @peppes_ghost DIY LiDAR option as a fallback.
(2) Impulse response: play an exponential sine sweep through a speaker and record it with a flat mic at the performer's position. You will deconvolve this into an IR.
(3) Dry performance take: record the instrument and voice close-miked or DI, so the room is not already baked into the sound.
Log the room name, permission, lighting and mic positions.

Deconvolve the recorded sweep against the reference sweep to recover the room's impulse response, using Farina's exponential-sine-sweep method. Do it in Python with numpy and scipy, or in Room EQ Wizard.
From the IR, compute the room's RT60 (the time for the energy to decay 60 dB, usually via the T30 or T20 extrapolation from the Schroeder integral).
Save the IR as a WAV ready for a convolution reverb.
Inspect the IR for artifacts: a clean exponential decay, no pre-ringing, no clipping in the sweep recording.
Record the RT60 and the IR length. This measured RT60 is the target the processed audio must match later. Keep the reference sweep file and the recorded sweep together, labeled, so the deconvolution is reproducible. Note the speaker and mic used, since the IR carries their coloration and you may want to compensate.

Load the dry performance take into a DAW with a convolution reverb, or convolve in Python (scipy.signal.fftconvolve), using capture/room_ir.wav.
Balance the dry and wet signal so the result sounds like the performance happened in that room, not drowned in reverb.
Then verify: measure the RT60 of the processed take (from a decaying note or by analysis) and compare it with the room's measured RT60 from step 2. They should agree within about 10%.
If you have a wet reference recording of the same performance in the room, compare the convolved version to it as a second check.
Export the processed audio. Keep both the dry and the processed takes, so the edit can choose. Note the wet/dry ratio you settled on, because it is the single biggest taste decision here and you will want to recall it for the next room.

Load the .splat into Gausseous (quixbrics.github.io/gausseous). Confirm it renders without holes in the performance area.
Use the Web Audio AnalyserNode to split the processed audio into bands. Map them deliberately: bass drives overall scale or a slow pulse, mids drive position drift, highs drive shimmer or point size. Gate each mapping with a threshold and smooth it with an attack/release envelope so the motion reads as musical rather than twitchy.
Use a sustained, droning passage, like the khaen or a bowed note, to tune the response, because long tones make reactivity legible.
Add Gausseous's depth of field, a little lens curvature and film emulation for the volumetric glow in the DeepPIV reference, without burying the room.
Export a reactive clip at least as long as the final edit.

Assemble a 60-second vertical (9:16) edit. Intercut real performance footage, honest and documentary in the Lomax spirit, with the reacting splat. The place itself should move with the tune.
Sync the cut to the processed room audio.
Open on something that establishes the real place, so the splat reads as that same room transformed, not a generic effect. Follow the @honeycrush.online proof that a two-person, one-camera shoot can carry a strong mood.
Keep the musicians and the place named in the caption. No generated faces, no synthetic performance.
Export at the platform's preferred vertical resolution and frame rate. Keep the project file and a frame-accurate marker list so re-exports match. Export a square or landscape version too if the act wants it for other platforms, from the same timeline.

Make a control: a plain performance video of the same tune in the same room, same length, no splat.
Post the splat cut and the control on the same account, matched for day and time of day, far enough apart not to cannibalize each other, and ideally counterbalanced across two tunes.
Pre-register the metric: average watch-through (percent of the video watched) and completion rate.
Collect the platform analytics for both after a set window (for example 7 days).
State the pass rule in advance: the splat cut holds watch-through at least as well as the plain video. If it loses watch-through, the format is decoration, per the dive's falsifier. Record the view count, likes and shares alongside retention, but judge on retention, since reach depends on the algorithm. Keep the raw analytics screenshots in ab_test/ as evidence for the audit.

Run the oracle pass:
(a) The .splat opens and reacts in a desktop browser with no holes in the performance area.
(b) The processed RT60 is within 10% of the room's measured RT60.
(c) A blind listening test: people match the processed audio to the room better than chance.
(d) The A/B watch-through and completion of the splat cut versus the control.
(e) The falsifier: are interior scans too noisy to read, or does the splat cut lose watch-through? If either, say the format is a gimmick for this room.
Write report.md with the sections Capture, Acoustics, Render, Edit, A/B Result, Falsifier and Gaps. State plainly whether the format earned its complexity. Rank the gaps by effect on the result.

You are an independent reviewer: a recording engineer plus a volumetric artist. Open report.md, capture/, audio/, render/, edit/ and ab_test/.
1. Enumerate every claim: scan quality, the acoustic match, the reactivity, and the A/B result. Mark each verified, unverified or false.
2. Open the .splat yourself and look for holes in the performance area. Load it in Gausseous and confirm it reacts.
3. Recompute the processed take's RT60 and compare it with the measured room RT60. Is it within 10%?
4. Confirm the audio in the final edit is the convolved room take, not the dry take or a preset reverb.
5. Check the A/B: are the posts matched for time and length, and was the metric pre-registered?
6. List what is missing: scan holes, a clap-not-sweep IR, no dry take, an underpowered A/B.
7. Rank the gaps by effect on the result. Fix the top three now where possible (re-balance the wet/dry, raise the reactivity gates). Report what needs a reshoot and why.
Return audit.md with the columns claim, status, evidence and fix, plus 'Remaining gaps' and an A-F grade on whether the format earned its complexity.