Research Reports
← All reports

2026-09-23 · audited · confidence medium · verified 2026-09-23 · stable · 13 sources · keyboard-acoustics-evidence-audit-and-measurement-programme.md

  • keyboards
  • acoustics
  • measurement
  • fft
  • gasket-mount
  • mahogany
  • gmmk-tkl
  • marbly

Keyboard Acoustics — Evidence Audit of the Build Corpus and the First Measurement Programme

Date: 2026-09-23 | Sources: 13 primary + two-round cross-model audit (three models round 1, two deep-research models round 2) | Overall confidence: Medium Supersedes mahogany-founding-profile-2026-05-28 and gmmk-tkl-v1-ultimate-marble-wood-build-2026-05-26. Synthesis: 11-SYNTHESIS-R2.md. Pre-registered predictions: 03-PREDICTIONS.md §R2.

Executive Summary

The keyboard corpus (twelve logs, two reports, a cluster of stubs) was written as a decision system: build briefs were derived from it, wood was specified to its numbers. Two audit rounds have now tested it against physics and against everything a web search can reach. The result is blunt. No controlled measurement exists, in the corpus or anywhere either deep-research model could find, for any variable the corpus ranks: case material, plate material, tape mod, PE/IXPE foam, plate foam, cavity fill or keycap profile. The number the May reports called "the most important in this entire report", a 12–17 dB cut at 4.2 kHz for wood cases, has no primary source. It is vendor copy that propagated, and it is removed here as a decision input.

Three things change what the operator does next.

  1. The two May reports are superseded. Their central conclusions, that the mahogany case is the dominant filter and that marbly is structurally unattainable on wood, rested on the unsourced figure and on a case-first hierarchy that the operator's own keycap A/B overturned. The as-built description of the G87 sound survives. The attainability matrix does not.
  2. The planned GMMK hybrid (LOG-021) is acoustically a different machine from what its brief says. The aluminium top rests on a silicone strip under its own weight. At that load the strip squeezes about 0.3 %, the whole unit bounces on it at roughly 170–250 Hz, and the strip isolates the 1–4 kHz band from the wood by an estimated 30–50 dB. So the wood can neither kill nor preserve the marble ring through the joint, and the "woody warmth" the brief credits to mahogany may instead be that bounce mode. Both claims are now pre-registered predictions with thresholds. Separately, the brief's pocket-depth rule assumed a 25 % squeeze and would leave the top about 2 mm proud; it is withdrawn, and the carpenter must not cut to it.
  3. The measurement programme is specified and costs almost nothing. A phone checked for gain stability, a gravity-drop striker, an MDF rim and a Python script. Session 0 measures the rig's own noise floor, because nobody has published one. The operator's own boards supply the anchors: the LOG-029 board with Cherry caps is the recorded definition of creamy, with SA caps of marbly.

Codebase Context

This report is the published form of the keyboard-acoustics audit workspace. Round 1 (2026-09-06) audited the physics of twelve logs without browsing and demoted the case-first hierarchy, the "permanently blocked" rule, the numeric confidences and the CLD label. Round 2 (2026-09-23) sent a deep-research brief to GPT and Gemini asking for sources only browsing could reach and for a measurement design. It also carried a correction: round 1 had been told the GMMK top is bolted into the wood, and it is not.

Documents changed alongside this report:

  • LOG-021 v6: joint description, withdrawn coupling argument, bounce-mode caveat, magnets flagged as preload, and the 2.25 mm pocket-depth rule withdrawn as a cut-once error.
  • LOG-006 §3.4a: silicone and hydraulic-lock wording marked unsourced; compression exists only if something clamps the plate.
  • LOG-023 §1.1 and LOG-001 §6.2: the 12–17 dB figure struck.
  • LOG-029: keycap-cavity explanation checked and bounded; cap mass must be weighed.
  • STUB-068: capture protocol and sourced data written in.
  • STUB-066 and STUB-074: tape-mod and per-material claims marked as unmeasured consensus.

1. What the corpus claims, and what stands behind it

Claim Where Evidence after two rounds Status
Wood cuts 12–17 dB at 4.2 kHz (widened to 3–6 kHz, reassigned to mahogany) LOG-001 §6.2, LOG-023 §1.1, mahogany report No primary source. Both deep-research models, plus a July search, found only vendor re-assertions. Removed
Case material is the dominant filter LOG-023 §3.1, mahogany report Contradicted by the operator's keycap A/B. All three round-1 models put it near the bottom of causal ordering. Withdrawn
Marbly is permanently blocked on wood LOG-023 §1.3, LOG-001 §6.4 The claimed kill band and the target band overlap only at 3–4 kHz. Direct radiation from switch and keycap never passes through the case. Attenuated, magnitude unmeasured
Hybrid top radiates up while wood damps down, two independent paths LOG-021 §0.2 v4 Round 1 rejected it on a wrong premise (bolted joint). Round 2 found the real joint isolates the marble band from the wood (§2). The paths are nearly independent above ~300 Hz, but for a reason the brief never gave. Mechanism replaced; P3′ tests it
Tape mod is a low-pass that makes a board poppier LOG-005 §6C, STUB-066 RQ3 No controlled before/after spectrum found. Community consensus only. Unmeasured
PE/IXPE under switches adds pop; Poron plate foam cleans the ring STUB-066, LOG-029 No controlled spectrum found. Unmeasured
Silicone relaxes under 5 %; 25 % confined squeeze causes "hydraulic lock" LOG-006 §3.4a (from round 1) No 30–40 A sheet datum found. "Hydraulic lock" is not a real term. The real mechanism is shape-factor stiffening of a near-incompressible elastomer (ν ≈ 0.499, sourced). Unsourced; moot for gravity-seated mounts
Poron relaxes ~20 %, so pre-compress 25 % LOG-006 v4 Rogers publishes compression set ≤10 % at 70 °C and no stress-relaxation figure for PORON 4701-30. No source
SA caps sound marbly because tall chambers radiate a ring LOG-029 v4 The A/B is real. A cavity resonance at the 1–1.8 kHz a reviewer proposed is ruled out by arithmetic (§4). Cap mass has not been weighed. Result stands; mechanism open
Cavity axial modes 486 / 1,289 / 1,143 Hz LOG-006, LOG-011 Arithmetic verified in round 1. Their relevance to a few-millisecond transient is unproven. Arithmetic correct
Tenting gap gives a 2–3 Hz Helmholtz mode LOG-006, LOG-011 Refuted in round 1. A real gap gives 82–163 Hz. Corrected

2. The GMMK hybrid mount, computed

LOG-021 drops the aluminium top and its screwed-on PCB into a mahogany pocket. The top's 2 mm perimeter wall rests on a 3 mm silicone strip in a routed ledge. Nothing screws into the wood. The only preload is the unit's own weight.

Taking the unit at 0.65 kg on a 970 mm by 2 mm landing, and converting Shore hardness to modulus with Gent's relation:

Silicone Contact pressure Static squeeze Bounce mode Estimated isolation at 1 / 2 / 4 kHz
Shore 30 A 3.3 kPa 0.29 % 170 Hz, about 210 Hz with dynamic stiffening 31 / 43 / 55 dB
Shore 40 A 3.3 kPa 0.19 % 206 Hz, about 250 Hz 27 / 39 / 51 dB

The model is a single mass on a spring. It ignores the air path through the cavity and wave motion inside the strip, whose through-thickness resonance lands near 5.4–6.5 kHz. Script and output: measure/r2-verify.py.

Four consequences follow.

  • The wood cannot reach the plate's marble ring through the joint. The earlier arguments that wood drains the ring, or that wood cannot touch it, were both reasoning about a path this mount mostly cuts. What the case still does is bound the air cavity under the PCB.
  • The warmth band has a rival explanation. A whole-unit bounce at 170–250 Hz sits inside the 200–600 Hz band the brief assigns to mahogany. That mode depends on mass and gasket stiffness, not on wood species. Prediction P8 separates the two.
  • Ledge depth does not set gasket squeeze on this build. Gravity does, at about 0.3 %. The 25 % versus 10–15 % question cannot arise unless something clamps the unit.
  • The brief's pocket depth would leave the top about 2 mm proud. LOG-021 sized the pocket as alu wall height minus 2.25 mm, a 25 % gasket sink. Under gravity the sink is about 9 µm. The flush rule is mechanical: pocket wall height above the ledge equals the calipered wall height plus the calipered strip thickness. LOG-021 v6 withdraws the old figure in every place it appeared. Nothing should be cut until both parts are measured.
  • The optional retention magnets are a clamp. Any magnet preload comparable to the unit's 6.4 N weight moves the bounce mode and the squeeze. Treat magnets as an acoustic variable, and test with and without.

A reviewer's surviving point is directional: the integrated top is the visible upper surface, so it radiates straight at the listener where a drop-in plate sits partly behind a bezel. No owned hardware can test that. It is recorded as hypothesis P3b and carries no number.

3. Measurement programme

3.1 Rig

Part Specification Why
Recorder Android phone, 48 kHz WAV, the least-processed audio source the app exposes, on a rigid stand Adequate for relative band comparisons at fixed geometry. Android gain control and noise suppression depend on source and device, so the path must be proven stable, not assumed raw.
Stability check A fixed reference tone from a second device, recorded at the start and end of every block If the tone's level or spectrum changes while nothing physical moved, discard the block. Never compensate after the fact.
Optional mic A USB measurement mic with an individual calibration file, such as a miniDSP UMIK-1 or Dayton iMM-6 Buy only if the phone fails the tone check, or when P4's low band is needed.
Striker Gravity-drop arm on a hinge, fixed release stop, small hard tip, single strike with the rebound caught No published repeatability exists for any cheap striker. This one has no electrical timing to drift.
Key A non-stabilised 1u alpha near the board centre, such as G or H A stabiliser adds another mechanical system to the measurement.
Contact sensor A piezo disc taped to the rim or case, recorded on a second channel if possible Needed to separate "the top got quieter" from "less energy entered the case". Required for any structure-borne claim.
Fixture Routed MDF or plywood rim sized for the GMMK unit, with the same silicone strip LOG-021 specifies Legitimate for A/B mount comparisons, because the substrate is common to both arms. It says nothing about absolute decay in mahogany.

3.2 Protocol

  1. Fix board, deskmat, stand and phone. Mark every position. Nothing moves within a block except the variable under test.
  2. Record the reference tone, then 30 strikes, then the tone again.
  3. Record 10 seconds of silence at the same position. A peak present without keystrokes is the room, not the board.
  4. Vary one thing. Randomise the order of conditions where practical.
  5. Repeat one condition at the end of the day. Its drift is part of the floor.

3.3 Analysis

The side-channel literature is the only large body of measured keystroke recordings. Asonov and Agrawal (2004) found a keystroke lasts about 100 ms. Its push peak is about 10 ms and contains two separate events of 2–3 ms each: the finger touching the cap, then the key hitting the plate. The release peak is much weaker. Their best spectral features came from a 2–3 ms window, not an 8–10 ms one. Zhuang, Zhou and Tygar (2009) put keystroke energy mainly between 400 Hz and 12 kHz. Both used non-mechanical keyboards, so their magnitudes do not transfer. The event structure does.

The pipeline follows from that:

  1. Detect each strike from short-time energy. Keep a 10 ms pre-trigger.
  2. Locate the hit peak. Take its spectrum with a Hann window of 512 samples, about 10.7 ms at 48 kHz. A 2048-sample window would blend the touch, the hit and the tail.
  3. Integrate band energy in 100–400 Hz, 250–500 Hz, 1–4 kHz and 3–6 kHz.
  4. Band-pass the waveform, take its envelope, and measure the time from peak to 20 dB down in each band. Do not use T20 or T30, which assume a diffuse reverberant room.
  5. Compute the spectral centroid over the first 20 ms.
  6. Report medians and bootstrap 95 % intervals over the 30 strikes, never a single averaged waveform.

A difference counts only if it meets the pre-registered effect size and its 95 % interval excludes zero.

3.4 Session order

Session Board Conditions Decides Can run now?
M0 floor LOG-029 board, SA caps Same configuration, two blocks, hours apart Whether anything below can be scored. A spread above ±1.5 dB at 1–4 kHz means fix the rig first. Yes
M1 anchors and P1 LOG-029 board Cherry caps, SA caps. Then plate and PCB out of the case on foam blocks, only if the operator accepts reopening the kit. The creamy and marbly anchors. P1: keycap versus case. Yes
M2 cap mass Both cap sets Weigh ten 1u caps of each on a 0.01 g scale Whether the SA effect could be mass. Replaces an unsourced reviewer table. Yes
M3 P3′ and P8 GMMK unit on the MDF rim Rim, stock ABS tray, foam blocks. Then +200 g taped at centre. Then magnets or a 1 kg preload. Whether the support changes the marble band. Whether a bounce mode explains the warmth. After the MDF rim is cut
M4 P4 GMMK unit on the MDF rim Sealed, then one side shimmed off the strip to open a 20–100 mm² gap The Helmholtz correction, off the built G87 After M3
M5 P5 RK61 Stock, butyl between plate and PCB, same added mass without butyl Mass-loading versus real damping Yes
M6 P7 deskmat half Any built board Deskmat on, off How big the desk boundary is. If it is below the floor, species tracking stays frozen for good. Yes
M7 WDT anchor Built G87 As-is The thock anchor Yes
— P2, P7 species half — Need a free-air G87 or two species cases. Not testable without cutting or dismantling. No
— P6 — Suspended until a design names its clamp No

3.5 Profile definitions

No published spectral definition of marbly, creamy, thock or poppy exists. The operator's own boards supply the anchors, recorded in M1 and M7 before any test board and never moved afterwards.

Label Anchor Rule
Creamy LOG-029 board, Cherry PBT caps A board is on the marbly side only if both the 1–4 kHz to 250–500 Hz energy ratio and the 20 ms centroid exceed the midpoint between the creamy and marbly anchors.
Marbly LOG-029 board, SA PBT caps As above
Thock / WDT Built G87 Reference point only
Poppy None owned Undefined until an anchor exists

3.6 Listening

Thirty blinded forced-choice trials per pair, with randomised, opaquely named clips. Twenty or more correct out of thirty is significant one-sided at p = 0.049. Run one block at original level and one level-matched, because matching can erase part of what "marbly" is. The same test measures the operator's own just-noticeable difference. No published figure exists for keyboard transients, and the one a reviewer cited was about reverberant music, not impacts.

4. Keycaps

The Cherry-to-SA swap on the LOG-029 board, "super creamy" to "marbly", remains the corpus's only empirical acoustic result. Its mechanism is still open.

One proposed mechanism is ruled out. A reviewer put an SA cap-cavity resonance at 1.2–1.8 kHz. A 12.5 mm closed-top cavity has its quarter-wave mode near 6.9 kHz. Treating the cap and switch top as a Helmholtz resonator with a realistic skirt gap gives 2.9–5.2 kHz. Neither reaches the band claimed. A cavity mode could still shape the top of the marble band, but it cannot be the 1–2 kHz ring.

Cap mass is the cheapest open variable. No sourced per-cap mass exists for SA PBT, and set weights cannot be divided into per-cap figures. Session M2 weighs them.

5. What survives from the superseded reports

  • The G87 as-built description. A silent tactile board with a full damping stack sounds muted, cushioned and dead-tailed with a woody colour. That is an observation of a built board, and it becomes the WDT anchor in M7.
  • The GMMK TKL v1 external dimensions, 356 × 129 × 35.9 mm from Glorious, and the Shopee PH sourcing leads. Both live on in LOG-021 and STUB-066.
  • Shopee PH has no dedicated PET plate-pad listing. This was still true at the May search and is unaffected by the audit.

Everything else in the two May reports that ranks, predicts or forbids a profile is superseded.

6. Sourced material data

Item Value Source
PORON 4701-30 compression force at 25 % 21 / 35 / 62 kPa for 240 / 320 / 400 kg/m³ Rogers product page
PORON 4701-30 compression set ≤10 % at 70 °C, ASTM D3574. No stress-relaxation figure published. Rogers product page
Silicone rubber Poisson ratio ≈0.499 Cui et al., elastomer characterisation paper (Penn State)
O-ring static squeeze guidance Up to about 30 %, for round sections only. Does not transfer to a flat strip. Parker O-Ring Handbook
Philippine mahogany viscoelasticity DMA study exists. Species and damping values not confirmed here. Montano et al., Applied Acoustics, 2023
Philippine mahogany density ≈570 kg/m³, small clear air-dried samples MatWeb
Plate E and density Aluminium 6061 69 GPa, 2,700 kg/m³. Brass C26000 ≈110 GPa. Stainless 304 193 GPa, 8,000 kg/m³. FR4 18–24 GPa in-plane, 1,850 kg/m³. PC ≈2.2 GPa, 1,200 kg/m³. POM ≈2.9 GPa, 1,410 kg/m³. Handbook values. Loss factors deliberately omitted: none were sourced.

7. Not found

Searched for by two deep-research models and not found. Each item is an open question, not an assumed zero.

  • A primary source for the 12–17 dB at 4.2 kHz wood figure.
  • Any controlled keystroke spectrum varying case material, plate material, tape, PE/IXPE, plate foam, cavity fill or keycap profile.
  • Strike-to-strike variability of keystroke band energy, with a finger or a striker.
  • Loss factors for lauan, Khaya or Swietenia suitable for comparison, or their variation with moisture.
  • Stress relaxation or compression set for 30–40 A silicone sheet at room temperature.
  • A just-noticeable difference for level or decay in keyboard-like transients.
  • GMMK TKL v1 top thickness and screw counts. The operator owns the board and can count them.
  • SA PBT per-cap mass.

Sources

  1. Asonov, D. & Agrawal, R. (2004). Keyboard Acoustic Emanations. IEEE S&P. https://web.eecs.umich.edu/~genkin/teaching/fall2019/EECS598-12_files/kbdacoustic.pdf
  2. Zhuang, L., Zhou, F. & Tygar, J. D. (2009). Keyboard Acoustic Emanations Revisited. ACM TISSEC. https://people.eecs.berkeley.edu/~tygar/papers/Keyboard_Acoustic_Emanations_Revisited/TISSEC.pdf
  3. Rogers Corporation. PORON 4701-30 Polyurethane. https://www.rogerscorp.com/Elastomeric-Material-Solutions/PORON-Industrial-Polyurethanes/PORON-4701-30
  4. Rogers Corporation. Real PORON. https://www.rogerscorp.com/elastomeric-material-solutions/Real-Poron
  5. Montano et al. (2023). Unraveling viscoelastic properties of Philippine Mahogany and manufactured acoustic guitar. Applied Acoustics. https://www.sciencedirect.com/science/article/abs/pii/S0003682X23004723
  6. Android Developers. AutomaticGainControl. https://developer.android.com/reference/android/media/audiofx/AutomaticGainControl
  7. Parker Hannifin. O-Ring Handbook ORD 5700. https://discover.parker.com/Parker-ORing-Handbook-ORD-5700
  8. Cui et al. Characterizing nonlinear response of elastomeric material systems. https://bpb-us-e1.wpmucdn.com/sites.psu.edu/dist/1/118779/files/2020/08/Cui_characterizing_nonlinear_response_elastomeric_material_systems_under_critical_point_constraints.pdf
  9. Stockwell Elastomerics. SG-series solid silicone. https://www.stockwell.com/data-sheets/sg-series-solid-silicone.pdf
  10. MatWeb. Philippine Mahogany. https://www.matweb.com/search/datasheet.aspx?matguid=5e029ab08277442ebaf3d9c49b4f512a
  11. Dayton Audio. iMM-6 calibrated measurement microphone. https://www.daytonaudio.com/product/1117/imm-6-idevice-calibrated-measurement-microphone
  12. miniDSP. UMIK-1 setup with REW. https://www.minidsp.com/applications/acoustic-measurements/umik-1-setup-with-rew
  13. Gent, A. N. (1958). On the relation between indentation hardness and Young's modulus. Trans. IRI 34, 46–57. Used for the Shore-to-modulus conversion in §2.