Add drop-test protocol, Edison synthesis, and first-data analysis#86
Conversation
|
@me-madsen claude pings will probably work here now, but you might want to point back to specific comments or areas from the previous 2 PRs: #82 and #67 (now the work is split across three PRs) |
|
We tested the combination of 60 in of drop height and 5 felt sheets. We used two specimens, 7xadt6 and 9GMQYQ. The test setup was the same as I used in #82 (comment), and a single-axial accelerometer and a tri-axial accelerometer were used for this test, which means that we still got rid of the bottom tri-axial accelerometer. Ch 5 measured data from a single accelerometer attached to the bottom acrylic plate, and Ch 2~4 measured data from a tri-axial accelerometer attached to the top vertex of the tensegrity structure. Ch2 is connected to the X axis, Ch3 to the Y axis, and Ch4 to the Z axis. This test is to prove the validation of the experiment setting using a 60-in drop height and 5 felt sheets. Here are the data |
|
Claude finished @ctrhjk's task in 9m 42s —— View job 60 in / 5 felts validation analysis (7xadt6 & 9GMQYQ) ✅
TL;DRThe 60 in / 5 felt setting is validated — keep it — but the felt is a consumable. 201/201 captures triggered cleanly, the output is the strongest and most repeatable of any condition tested (461–475 G, CV ≤ 1.7 %), and the two geometries separate decisively. The one campaign-scale effect in the data is felt compaction: the raw CH5 spike tripled (2.1 → 6.5 kG, 22 % → 69 % of full scale) across the two back-to-back sessions, which is what drives every drift term below. Transmissibility Capture health
OLS regression (stabilized phase, burn-in = 5 drops)The burn-in scan finds no k ≤ 20 with a non-significant trend on either specimen — the drift is campaign-scale felt wear, not a seating transient — so the SOP 5-drop burn-in is used. 7xadt6 (drops 6–100, n = 95):
9GMQYQ (drops 6–101, n = 96):
Reading it: the TOP output climbs ~0.03–0.06 %/drop, but in lockstep with the input (per-drop corr(TOP, CH5) = 0.999 for 7xadt6) — so Specimen discrimination (stabilized drops, Welch t): TOP 461.4 vs 475.2 G (−2.9 %, p = 3.7e-30, |d| = 2.1); T 1.034 vs 1.027 (+0.7 %, p = 6.4e-27, d = 2.1). These two geometries happen to be close responders, yet they still separate cleanly — sub-1 % differences in T are resolvable at n ≈ 5 drops. The felt-wear findingCH5's raw spike grows monotonically all evening: 7xadt6 starts at the fresh-felt level the sweep measured (~2.1 kG ≈ 22 % FS) and crosses the FS/3 head-room target around drop 67; 9GMQYQ (run second) is above FS/3 from drop 1 and ends at ~6.5 kG (69 % FS). The CFC-180 input barely moves (446 → 463 G) — compacted felt mostly adds high-frequency spike content, which is exactly what eats raw-peak head-room. Signal 101, recorded after the 12.9 min pause, drops back to 56 % FS — consistent with viscoelastic felt recovery. (See Recommendation on height / felt count
One question back, @ctrhjk: was the felt stack changed between the two specimens? The analysis assumes the same stack all evening (the monotonic CH5 trend suggests so). If it was changed, the 9GMQYQ starting level (~42 % FS) means felt sheets vary unit-to-unit — which makes logging felt state and normalizing by the input even more important. |
Stabilized-phase OLS drift, felt-wear saturation audit, and specimen discrimination for the two 100-drop validation campaigns at the felt-sheet sweep's recommended operating point. Co-authored-by: Jinkwan Han <282969777+ctrhjk@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
The stack of felt sheets wasn't changed |
|
Claude finished @ctrhjk's task in 1m 5s —— View job Felt-stack confirmation ✅
Thanks @ctrhjk — that confirms the analysis. With a single felt stack all evening:
The doc now records the confirmation in §3 and closes the corresponding caveat in §6. --- · Branch |
|
@claude recommend a minimum number of drop tests per specimen to get accurate data. How much variance have we had in our data per specimen being tested so far? Also provide how long each set of tests would take given that it currently takes ~42 seconds/drop at 60 inches with automatic dropping. For evaluating this, reference the data from #86 (comment) and #82 (comment) and #67 |
|
Claude finished @me-madsen's task in 4m 45s —— View job Minimum drops per specimen, variance so far, and set timing ✅
TL;DRMinimum: 5 recorded drops per specimen, after 2 discarded warm-up drops → 7 total ≈ 4.9 min/specimen at 60 in. This was the recommendation from the earlier meta-analysis, and folding in the two datasets you referenced (the 60 in/5 felts validation and the 5-vs-10-in comparison) strengthens it — the newest, largest datasets are among the tightest we've recorded. 1. How much variance per specimen so far?Every dataset where one specimen was dropped repeatedly, with the within-specimen coefficient of variation (CV) of the go-forward metrics (top-vertex CFC-180 output peak, and transmissibility
* the 60 in TOP/CH5 CVs are inflated by the campaign-scale felt-wear drift (both channels climb in lockstep); Pooled: output CV 0.31–3.5 %, median ≈ 1.0 % (90th pct ≈ 2.3 %); 2. Minimum drops per specimenTwo requirements, both computed from the observed CVs (
3. Time per set at ~42 s/drop (60 in, auto)Your 42 s figure is confirmed at campaign scale — the 201-drop 60 in/5 felts session logged a median cadence of 41 s.
Batch scale at 60 in, baseline plan: 10 designs ≈ 0.8 h, 20 designs ≈ 1.6 h, 48 ≈ 3.9 h, 96 ≈ 7.8 h (conservative plan ≈ 1.7× those). One caveat carried over from the writeup: these CVs are repeat drops of one physical article — print-to-print reproducibility of the same geometry isn't characterized yet, so plan on ≥3 replicate prints per geometry before treating peak-g/ --- · Branch |
Co-authored-by: Jinkwan Han <282969777+ctrhjk@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Refreshes the drops-per-specimen variance/sample-size/timing meta-analysis with the two datasets referenced in the PR #86 re-ask: the 60 in / 5 felts validation (7xadt6/9GMQYQ, ~95 stabilized drops each) and the 5-vs-10-in comparison (30 drops/height). Pooled output CV 0.31-3.5% (median ~1.0%); T median CV tightens to ~1.2% (0.12-0.45% at the 60 in / 5 felts operating point). Recommendation unchanged: 2 warm-up + 5 recorded = 7 drops/specimen. Co-authored-by: Marcus Madsen <265197858+me-madsen@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
These are the slo-mo videos of the two specimens' drop tests. 7xadt6: 9GMQYQ: Can you do a video analysis? |
|
Claude finished @ctrhjk's task in 8m 56s —— View job Slow-mo video analysis (7xadt6 & 9GMQYQ) — done, with one limitation
|
Good to note that there seems to be relatively low variance in our testing sessions per specimen. We will currently plan to do 50 drops per specimen as a baseline for our testing, and may increase this up to 100 in the future; simply to have a relatively large dataset for our research. |
YouTube blocks full downloads from CI (sign-in/bot gate), so this commits the eight real preview frames YouTube serves publicly (25/50/75% + poster per video) with frame-level observations: bungee-free rig verified, key-seat mount + tie-down visible, specimen intact in every frame, felt impact-zone mottling consistent with the compaction finding, and carriage/ specimen parked to <=1 px between the two mid-run 9GMQYQ frames. Co-authored-by: Jinkwan Han <282969777+ctrhjk@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
@claude threw those videos into the branch, top-level dir. Rename and organize as you see fit. Not sure what the fps, but I'm certain that this is available somewhere else in this repo, perhaps in one of the earlier PRs. You may need to do an extensive sweep through all PR comments or do a repository-level search or equivalent to find it. |
|
Claude finished @sgbaird's task in 30m 8s —— View job Videos organized + full frame-by-frame kinematics done ✅
TL;DRThe committed files unlock exactly what the preview-frame pass couldn't do, and the headline result is that the video independently corroborates the accelerometer record: the impact pulse is ≤2 capture frames ≈ 1–2 ms (DAQ says ~1.6 ms), and the mystery "parked carriage" from the preview frames turns out to be an anti-rebound brake catch — no secondary impacts ever reach the specimen. Time base — solved exactly, no fps guesswork leftThe uploads are YouTube's 30 fps re-encodes of the camera's 24p/960 fps HFR clips. Instead of assuming a mapping (the burn-in-wax pass had to carry a ±25 % caveat here), the script detects and removes the 24→30 pulldown duplicate frames: 19.8 % duplicates in 7xadt6 (textbook 1-in-5) and 14.1 % in 9GMQYQ. The surviving frames are the camera's capture frames, so real time = Kinematics (one drop per specimen)
Figures: velocity/trajectory · impact zoom · montages: 7xadt6, 9GMQYQ The legacy preview JPGs stay in --- · Branch |
Move @sgbaird's root-level video uploads to data/drop-tests/60in-5felts-validation/video/{7xadt6,9GMQYQ}_slomo.mp4 and run the frame-by-frame pass that was previously blocked by YouTube's bot gate. fps recovered from the repo record + PR #67 camera spec: Sony RX100 IV HFR at 960 fps; the script deduplicates the 24->30 pulldown so real time is exactly unique-frame/960 (19.8%/14.1% duplicates found). Findings: impact pulse <= 2 capture frames (~1-2 ms) corroborating the DAQ's ~1.6 ms CFC width; anti-rebound brake catch at ~2.1-2.4 g holding the carriage 130-150 mm above the felt from ~86-89 ms after impact (no secondary hits); top-vertex elastic snap-back at ~0.7x impact speed with visible strut flexure fully recovered; cross-framing scale consistency (82 vs 78 mm specimen extent) under the free-fall anchor (DAQ dv 5.53/5.69 m/s corroborates). Compression itself falls between frames - >=5000 fps DIC still needed. Co-authored-by: Sterling G. Baird <45469701+sgbaird@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Given these results the tower should be fine to resume regular testing. Depending on the results of this inquiry we may be changing how we run this campaign to collect data soon #97 (comment) |
|
Note the following conversation regarding beginning testing. 
|
|
Claude finished @me-madsen's task in 1m 51s —— View job Accelerometer calibration settings — what's recorded in the repo ✅
Yes — the TP4 per-channel calibration settings are recorded, in many places, and your instinct is right: the source is @ctrhjk. The fullest record traces back to @ctrhjk's channel-map post on PR #67, which was committed into The most complete record —
|
| Channel | Sensor | Full scale (G) | Sensitivity (mV/G) | Coupling | ICP | Waveform analysis |
|---|---|---|---|---|---|---|
| CH2 | tri-axis (X) | 14,492.8 | 0.69 | AC | yes | Half Sine |
| CH3 | tri-axis (Y) | 14,992.5 | 0.667 | AC | yes | Half Sine |
| CH4 | tri-axis (Z) | 13,624.0 | 0.734 | AC | yes | Half Sine |
| CH5 | single-axis | 9,442.9 | 1.059 | AC | yes | Half Sine |
The axis assignment (CH2 = X, CH3 = Y, CH4 = Z) was stated by @ctrhjk in the 60 in/5 felts comment and is recorded in the READMEs of every recent dataset (abc123-blind, 60in-5felts-validation, print-defects, pu-configs, etc.).
The retired low-range tri-axis (bottom-vertex "BOT", CH6–CH8) is recorded too, e.g. data/drop-tests/100drops/README.md: 1,002.0 / 991.1 / 989.1 G at 9.98 / 10.09 / 10.11 mV/G.
These sensitivities have been identical in every dataset from June through the current campaigns — nobody has changed the per-channel calibration entries on the TP4 since @ctrhjk recorded them.
Capture-setting history (also recorded, per dataset README)
| era | record / rate | pre-trigger | trigger |
|---|---|---|---|
| June–early July (125 kHz era) | 200 ms / 125 kHz / 25,000 samples | 2 % = 4 ms | 1000 G (CH4, later CH5) |
| late July (PU sweeps) | 20 ms / 1.25 MHz | ~0.39 ms | 300 G, then 150 G |
| current SOP (since abc123 blind test) | 100 ms / 1.25 MHz / 125,000 samples | 2 % = 2 ms | 150 G |
The current SOP is pinned in docs/drop-test-ab-blind-protocol.md and data/drop-tests/abc123-blind/README.md.
What is not recorded anywhere — the actual gaps
- Sensor make/model/serial numbers and factory calibration certificates. The protocol doc says only "Accelerometer Get started on this repo #1 (PCB / similar)" (
docs/drop-test-protocol.md:23). So we know the numbers entered into the TP4, but not whether the sensitivities came from each sensor's factory cal sheet or were typed from something else — and there's no way to check drift against a certificate we don't have. - Where the full-scale values come from. They're consistent with the TP4's ±10 V input divided by each sensitivity (e.g. 10 V ÷ 1.059 mV/G = 9,443 G), i.e. derived ceilings, not independent settings — but that derivation isn't confirmed in any committed record.
- Mechanical shock-survival ratings of the sensors (separate from measurement range) — flagged as missing back in the accelerometer-ranges answer, still missing.
- No screenshot of the TP4 Channel Settings page is committed — everything we have is transcribed.
Suggested 5-minute close-out (for @ctrhjk, next time at the tower)
Photograph (a) the TP4 Channel Settings screen and (b) the model/serial labels on the three sensors — plus the calibration sheets if they're in the case — and post them here. I'll fold them into the protocol doc as the canonical calibration record. That would close gaps 1–4 in one pass and make the sensitivity entries auditable against the factory values.
No code or docs were changed — this was a search/report task.
· Branch
|
@claude note that these are what I believe are the right settings for future reference. The settings were cleared and so I want to ensure the settings are giving the right sensitivities and so on. Compare the following data of the same specimen before the reset to see if the sensitivities are right: Before clearing of setting (101 drops): https://byu.box.com/s/it5499hkyw24twg7179smsn0fv0bodal 
|
|
Claude finished @me-madsen's task in 13m 55s —— View job Post-reset sensitivity verification — the re-entered settings check out ✅
TL;DRThe re-entered settings are right — resume testing on them as-is. The screenshot matches the June-era channel map digit-for-digit, and the data confirms the entries took effect: every observable that a wrong sensitivity must move is continuous across the reset to ≲1 %, an order of magnitude below the smallest plausible typo (3.4 %). The one thing that did change (+9–11 % on the peak levels) has the signature of the rested/re-seated mat, not of a channel gain — a gain error can't narrow a pulse, and the pulse narrowed −8 % while impulse, Δv and hop timing stayed put. 1. Screenshot vs the canonical record — exact matchCH2/CH3/CH4 = 0.690 / 0.667 / 0.734 mV/G (14,492.8 / 14,992.5 / 13,624.0 G), CH5 = 1.059 mV/G (9,442.9 G) with the 150 G trigger, all AC + ICP + Half Sine; 100 ms / 1.25 MHz / 125,000 samples / 2 % = 2 ms pre-trigger. Identical to @ctrhjk's channel map recorded since June ( 2. The data-level checkA mis-entered sensitivity is a per-channel multiplicative factor: it must scale that channel's peak, Δv and noise floor by the same amount while leaving all timing/shape metrics untouched. So the test is scale-coherence, not raw level matching:
Candidate typos, each excluded: a channel left at the 10 mV/G default (×0.07–0.10 — nowhere close), CH5 entered as 1.000 (×1.059 on Δv and e_rebound — both flat), a Z-axis value swap (would drag T by 6–10 % — T moved 0.3 %). The one candidate below the check's noise floor is a CH2↔CH3 value swap (±3.4 %, opposite signs on the two small lateral axes): the observed lateral shifts are same-sign (a physics pattern, not a swap pattern), and even if it existed it would move the campaign metric (the resultant, dominated 95 % by CH4/Z) by ≲0.3 %. Figures: continuity across the reset · per-channel ratios vs candidates. Bonus continuity: the 08-17 session's input (197.5 G, 2.48 ms) lands right on the 08-11/12 speed-decay record at this arrangement (195–201 G, 2.5–2.6 ms), so the chain runs unbroken from the pre-reset era through the reset. 3. What the visible +10 % shift actually isThe after session hits harder and shorter at the same arrival energy — rested mat (2 days idle after 101 drops) plus whatever re-seating happened during the settings work. The mirror image is inside the before session: its Δv slid 3.96 → 3.11 m/s over 101 drops as the mat's rebound faded, the exact within-session mat drift the speed-decay campaign documented. Practical corollary: after any mat rest/re-seat, budget ~5–10 warm-up drops (the after session takes that long to reach its plateau, a bit more than the usual 2). Two small things back
|
…ere it matters; commit the settings screenshot as the canonical record Co-authored-by: Marcus Madsen <265197858+me-madsen@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
With the complete SOBOL +S0 batch, the campaign has officially begun. 4/9 of our SOBOL + S0 batch have been tested at the following conditions:
The data will be analyzed for average transmissibility per specimen, and any other performance measurements that @claude can extract from our data as has already been set as a precedent in this PR, #94, and #97, and anywhere else as discussed in this repository. The data is being uploaded to box. It will then be posted here and analyzed. The BO will most likely be handled in a separate thread. The initial transmissibility analysis may also be handled in that separate PR. @sgbaird will handle implementation of the BO at this time. It takes roughly 90 minutes to run each specimen through the testing. We can expect to have this data at some point tomorrow. |
|
Claude finished @me-madsen's task in 6m 45s —— View job Campaign kickoff — batch pipeline ready for the data ✅
TL;DRNothing to analyze yet (data lands tomorrow), so I used the window to make the analysis a one-command turnaround: What gets extracted per specimen (the full precedent stack)
The one thing the BO hand-off still needs: the specimen-ID → design-parameter mapThis has been an open gap since the first input-output run ("IDs not yet tied back to design parameters"). The script takes Asks for the upload (all cheap, all bite later if missing)
One connection worth confirming: the 08-18 settings screenshot showed the TP4 database sitting on a session named When the Box links are posted, ping me and the turnaround is: fetch → run → per-specimen writeup + campaign ranking + · Branch |
…pload One command per Box upload: per-specimen stabilized metrics (T in both CFC bands, dv rig-health verdict, t_second, e_rebound, r2-gated ringdown fn/zeta), invalid-capture and warm-up handling, cross-specimen ANOVA/pairwise stats, figures, and a BO-ready campaign_summary.csv with optional specimen-id -> design-parameter passthrough. Smoke-tested on Box captures from the committed speed-decay manifests (loose CSVs and zip archives both exercised). Co-authored-by: Marcus Madsen <265197858+me-madsen@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @me-madsen's task in 23m 3s —— View job SOBOL + S0 campaign analysis — 8/9 specimens, 942 drops ✅
TL;DRThe campaign discriminates decisively, and there's a headline: The ranking (T = TOP/CH5, CFC-180; stabilized drops)
† Every adjacent pair separates at |d| ≥ 2.8, but against the known ~2 % print-to-print floor the honest read is four design tiers:
|
… fix the pre-trigger contact-foot baseline, confirm the tower healthy via the TP4 series tables, rank the designs (6lhxfy T=0.893 is the first strong attenuator), and emit the BO summary Co-authored-by: Marcus Madsen <265197858+me-madsen@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Claude finished @sgbaird's task in 3m 51s —— View job Mass variation explained — the constraint was implemented, but it holds solid volume constant, not printed grams ✅
TL;DRThe constant-mass constraint was implemented and converged — the batch CSV predicts 30.90–30.97 g for all 9 designs (±0.1 %). The measured spread (18.50–22.04 g, CV 5.9 %) comes from a definitional gap: the constraint equalizes solid CAD volume, but the printer gives PLA ~57 % effective density (walls + sparse infill) while the thin TPU cables print ~solid. Since the PLA/TPU split varies strongly by design (TPU fraction 12–36 %), a constant-solid-mass batch deterministically becomes an ~±9 % printed-mass batch. Regression on the generator's own per-material predictions explains R² = 0.78 of the variation; the residual ±0.3–0.8 g (~3 %) is ordinary print-to-print variance. 1. The constraint exists and did its jobYou remembered right: 2. Where the grams wentThe CSV also predicts the per-material split ( measured ≈ 0.565 · m_PLA,solid + 0.986 · m_TPU,solid (R² = 0.78, residual sd 0.64 g)
Reading it: the slicer prints PLA at ~57 % of solid and TPU at ~99 % (thin cables are all walls — no room for infill). The mass projection preserves each design's shape, so the thick-strut/thin-cable corner ( 3. What to do about it (if anything)For this batch, nothing is lost: the actual masses are recorded in For future batches, two one-file fixes in
Which is "right" depends on the constraint's purpose: if the point is attenuation-per-gram-of-printed-structure at fixed mass budget, printed mass is the physical quantity to hold constant, and option 2 is the honest version. 4. A useful side effect: the fit can help identify
|
…s constraint held solid volume constant (30.90-30.97 g predicted), but PLA prints at ~57% effective density vs ~99% for TPU, so the design-dependent PLA/TPU split maps it onto the observed +/-9% printed-mass spread (fit R2=0.78) Co-authored-by: Sterling G. Baird <45469701+sgbaird@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Aside: seems like 100% infill of the PLA struts would have been better. Was there not a conversation around this in the repository somewhere about setting the infill to 100%? |
…suggest round-2 batch Multi-objective (t180, e_rebound), fully Bayesian, batch, existing-data Honegumi template adapted for the physical campaign: Sobol init step dropped (PR #35 batch was the init), results ingested from the PR #86 campaign_summary.csv snapshot, specs 03/06/07 attached as pending trials, amdjwm skipped until its spec mapping is resolved. Records 9 suggested base-space designs with posterior predictions, the AxClient state, and a Pareto + parameter-space figure. Co-authored-by: Sterling G. Baird <45469701+sgbaird@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Task 3e398131: Edison derives its own objective set from the per-drop campaign data before reading ours, then attacks the three legs of the hand-off claim (t180+e_rebound objective pair, the n=8 trade-off/Pareto framing, and the per-drop-SEM noise model that ignores the ~2% print floor). Bundle: 23 files across three branches (per-drop metrics, series tables, analysis doc+script, print-defect floor study, #97 energy review, BO script). Follows the d9092c5a precedent from PR #86. Co-authored-by: Sterling G. Baird <45469701+sgbaird@users.noreply.github.com>
|
Note this about databases when using the droptower. Databases can fill up quickly due to the high fidelity data we are gathering. This shows how to create new databases and change between them: https://youtu.be/BSK_UcERTVw |
…33 #94 #97 #85 #86 #98 #101) Pivot from planned-methods SEA/eta_c framing to the executed campaign: drop-tower objectives t180 (filtered peak-acceleration ratio) and rebound energy per drop, SAASBO round 1 on the printed Sobol seed batch (real results table, real Pareto/feature-importance/LOOCV figures with labels regenerated for naming consistency), round-2 batch in fabrication, constant-solid-mass projection + printability screens, as-printed fabrication record (manual painted supports, TPU dry box, high-flow nozzle), corrected J211 filter provenance, simulation screening ladder from PR #33 with honest scorecard, metal-analog metric switched to t180, Edison adversarial objective review reflected in Discussion/round-3 plan. SI rewritten: print key, drop-tower protocol/rig characterization, printed-mass model, simulation ladder. Em-dash sweep per style guide. Rebuilt all four PDFs. Co-authored-by: Sterling G. Baird <45469701+sgbaird@users.noreply.github.com> Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Issue is operational (coordinating the first crush/drop tests on Jeff Hill's tower) and this repo is the LaTeX MRG proposal with no code or existing test-protocol surface. This PR consolidates the moving parts from the issue thread, adds a literature synthesis, and analyzes the recorded accelerometer data.
Added
docs/drop-test-protocol.md— single source of truth for the drop-test setup, covering:g_max, SEA, full ~10 s ringdown (not just the 200 ms shock), reusability, slow-mo framing from t=0edison-trajectories/drop-test/— Edison Scientific LITERATURE_HIGH synthesis (task653d7d39) on drop-tower troubleshooting for small 3D-printed lattice/tensegrity specimens: ~57 KB report, full JSON dump, submission record, and README. Idempotent driver atscripts/edison/submit_drop_test.py. Surfaces a standards stack (ASTM D5276/D7136/D3332, ISO 6603/1683/5347, MIL-STD-810 method 516, SAE J211) and recommendations (linear sleeve bearings, magnetic/elastic top-plate hold-down, ≥10 s ring-buffer DAQ, SAE J211 CFC filtering, n ≥ 5 + CV, ≥5000 fps DIC; closest analogues Pajunen 2019, Dwyer 2023).First drop-test data analysis of the five TP4 accelerometer exports posted by @me-madsen (
Signal 10–14, 4-channel, 125 kHz, 0.2 s window):data/drop-tests/raw/— committed raw export filesscripts/analysis/drop_test_analysis.py— loader, SAE J211 CFC-1000 / CFC-180 filtering, peak/pulse/PSD metrics, and figure generationdata/drop-tests/figures/— full-window CH1 overlay, per-run impact zoom (raw vs filtered), peak-g bar chart, PSD, and CH4 trigger-artifact plotdocs/drop-test-analysis.md+data/drop-tests/README.md— findings: the "audrey" tensegrity specimen reduces CFC-180 peak acceleration ~74–79 % vs the no-specimen control (~370–463 G vs ~1,792 G), while the PETG run's raw peak is within ~1 % of the control (≈ direct plate-on-plate hit), strong evidence of the bungee-driven lift-off; CH4 carries a fixed ~1.4 kG trigger/release artifact at t≈4.2 ms in every run. Caveats noted: unconfirmed channel map, 200 ms window only, n = 1 for control/PETG, no Δv/SEA quoted yet.Vertex vs. acrylic-plate T3-prism drop-test data and analysis posted by @ctrhjk (PR Add drop-test protocol, Edison synthesis, and first-data analysis #67), under
data/drop-tests/vertex-acrylic/:raw/— eight TP4 exports{n0jdwk, m6cyoq, T3_0103, T3_0000}_Signal{1,2}.csv(Signal1= vertex-mounted,Signal2= acrylic-plate), single drop per configuration at 13 ft, 200 ms / 125 kHzREADME.md— channel map (CH1 removed; CH2–CH4 tri-axis; CH5 single-axis; CH4 = 1000 G trigger) with full-scale/sensitivity per channel, per-specimen file index, and @ctrhjk's observations (clip-height/no-trigger issue, hot-glue z-axis mount limitation,m6cyoqstrut andT3_0103TPU-tendon damage after the acrylic test, and the invalidT3_0000acrylic run where the accelerometer fell off)scripts/analysis/drop_test_vertex_acrylic_analysis.py— locates the impact via the triggered CH4 channel (windowed ±1.5 ms peak search in the first 10 ms, not a global max), baseline-corrects, and reports raw / SAE J211 CFC-1000 / CFC-180 peaks for the single-axis CH5 (primary go-forward sensor) and tri-axis CH4, auto-flagging invalid / no-clean-impact runsdata/drop-tests/vertex-acrylic/figures/— vertex-vs-acrylic CH5 impact windows, CFC-180 peak-g bar chart, and vertex CH5 PSDdocs/drop-test-vertex-acrylic-analysis.md— findings + an explicit SOP / test-method section: vertex mounting is repeatable (4/4 clean, CFC-180 229–284 G, CV ≈ 9 %) while the acrylic configuration is not (3/4 runs registered no clean impact — clips too low so the plate seats on the specimen, plus the fell-offT3_0000); the vertex peaks do not yet discriminate geometry so fresh intact distinct-geometry samples (vertex-only, n ≥ 5) are needed before peak-g is a trustworthy BO objective; replace the hot-glue mount with a z-axis-aligned seat; and the single-axis sensor's raw peaks reach 70–90 % of its 9,442.9 G full scale (near saturation on the m6cyoq-acrylic run). Caveats: n = 1 per (specimen, mount), 200 ms window only, partial-pulse Δv, unconfirmed CH4/CH5 axis correspondence.Clip-height sweep & base-plate accelerometer-check diagnostic posted by @ctrhjk, under
data/drop-tests/clip-height/— drilling into why the acrylic-plate configuration repeatedly fails to trigger:raw/Accelerometer_check_Signal1.csv+README.md— the one triggered base-plate CSV (tri-axis on the bottom plate, 13 in drop) plus a setup README documenting both experiments: the clip-height sweep (extra bungees cured fly-off; tri-axis on the acrylic plate; clips at 0.5/1/1.5/2 in, two drops each; 0/8 drops triggered, video only) and the base-plate accelerometer check, with the shared channel mapscripts/analysis/drop_test_clip_height_analysis.py+data/drop-tests/clip-height/figures/— windowed CH4 impact location, SAE J211 CFC-1000 / CFC-180 peak/pulse/Δv metrics, and figures (base-plate impact window, full-window CH4, PSD)docs/drop-test-clip-height-analysis.md— findings: the base-plate hit triggers cleanly (CH4 raw 3072 G ≈ 3.1× the 1000 G trigger, CFC-180 280 G, Δv ≈ 3.3 m/s; CH4 dominates the off-axis channels ~23–55×), so the acrylic-plate "no trigger" failure (0/8 across the clip sweep) is a load-path problem — the plate seats on / is damped by the bungee-restrained specimen — not the sensor, DAQ, or trigger level.Input-output (transmissibility) drop-test data and analysis posted by @ctrhjk (PR Add drop-test protocol, Edison synthesis, and first-data analysis #67), under
data/drop-tests/input-output/— @ctrhjk's input-output instrumentation design: a single-axis accelerometer on the bottom plate = input (now the triggered channel CH5), a tri-axis accelerometer hot-glued to the top vertex = output (CH2–CH4), bungees removed, four distinct-geometry specimens (practice,n0jdwk,yqpmx1,h8Lbev) each dropped five times at 13 in:raw/— 20 TP4 exports{practice,n0jdwk,yqpmx1,h8Lbev}_Signal{1..5}.csv(Signalindex = drop number) +README.mdwith the channel map (trigger moved to the single-axis input CH5) and @ctrhjk's setup notesscripts/analysis/drop_test_input_output_analysis.py— locates the impact on the triggered CH5 (windowed ±1.5 ms peak), baseline-corrects, and reports raw / SAE J211 CFC-1000 / CFC-180 peaks for the input (CH5) and the tri-axis output resultant, the transmissibilityT = output/input, pulse width and Δv, with per-specimen mean ± 1σ / CV aggregatesdata/drop-tests/input-output/figures/— input-vs-output impact windows (5 drops overlaid), transmissibility bar chart, input repeatability, output PSDdocs/drop-test-input-output-analysis.md— findings: the input-output design works — 20/20 drops triggered cleanly, removing the bungees makes the input nearly constant (235–248 G CFC-180, ≤1.7 % CV), and transmissibility now discriminates geometry (yqpmx1≈ 0.96 is the only attenuator,h8Lbev≈ 1.09,practice/n0jdwk≈ 1.17–1.19), makingT(or output-peak-at-fixed-input) a usable BO objective; a mild within-run drift across the five cyclic drops is flagged as most likely hot-glue-mount-driven. Caveats: n = 1 specimen per geometry (5 repeat drops), 200 ms window, unverified tri-axis orientation, IDs not yet tied back to design parametersedison-trajectories/input-output/— Edison Scientific ANALYSIS (taskfe044079) that independently reproduced the transmissibility values exactly to two decimals, confirmed the within-run drift is statistically real and mount-driven (pooled +0.015/drop, p = 0.0001), endorsedTas a first-pass screening objective (recommending FRF / SRS-band metrics and output-peak-at-fixed-input as it matures), and gave a prioritized SOP (rigid z-aligned keyed sensor seat, keep bungees removed, extend capture past 200 ms, n ≥ 5 distinct prints per geometry with randomized order, anchor in SAE J211 / ISO 5347 / ASTM D3332). Idempotent driverscripts/edison/submit_input_output.py+ fetchscripts/edison/fetch_input_output.py; a cross-ch...Drops-per-specimen variance, sample-size, and timing meta-analysis answering @me-madsen's question (PR Add drop-test protocol, Edison synthesis, and first-data analysis #82 comment 5026945744: minimum drops per specimen, variance so far, and set duration at ~42 s/drop at 60 in), under
data/drop-tests/sample-size/(derived — no raw data of its own):scripts/analysis/drop_test_sample_size_analysis.py— aggregates the within-specimen coefficients of variation already emitted by the committed per-dataset analyses (large-campaignstabilized_ols/burn_in_drops, the felt-sheet per-condition CVs, and the n = 5 mount-validation writeups), then computes precision-based sample sizes (n = (t·CV/MoE)²), two-design discrimination power, and timing; emitsfigures/sample_size_metrics.json+ two figuresdocs/drop-test-sample-size-analysis.md+ folderREADME.md— findings: the go-forward output (top-vertex tri-axis CFC-180 peak / transmissibilityT) has pooled within-specimen CV 0.31–3.5 % (median ≈ 1.0 %), best ≈ 0.3–0.6 % on the wax key-seat mount, while the base-plate/input channel is the noisy near-saturation one. Recommendation: baseline 2 warm-up (discarded) + 5 recorded = 7 drops/specimen (5 recorded gives a ±1.2 % 95 % CI on the per-specimen mean and resolves the ≥10 % between-design differences), stepping to 10 recorded for the noisier input/Tchannel or a hot-glue mount. Timing at ~42 s/drop (60 in, auto): baseline ≈ 4.9 min/specimen; a 20-design BO batch × 7 drops ≈ 1.6 h (lower drop heights log ~12–20 s/drop, ~2.5× faster)data/drop-tests/README.md) and.gitignore(__pycache__/,*.pyc) updatedNote: I cannot open new PRs (only push commits to this branch), so @sgbaird's request to create a new PR is not actioned here; work continues on this branch.