# Experiment 003 — Weight Placement & Flight Stability

## Provenance and configuration
- Owner supplied independently decoded numeric evidence from `btfl_all(2).bbl`. Raw BBL is not bundled. Lovable did not decode or independently validate that source.
- Entire combined source SHA-256: `14a6a462a64864c34aad54ecff0e84753aec78b572a752a380d3ac2260b44797` (not a per-session hash).
- Four decoded BBL sessions; owner reports zero parsing continuity errors, recognizable end markers and approximately 407.2 Hz.
- BETAFPV Meteor75 Pro, Betaflight 2025.12.5, Angle Mode (owner-reported setup).
- November 15, 2026 is owner-reported, not independently time-stamped.
- Sessions 1, 2 and 3: weight on side(s), total added mass unspecified. Session 4: 2.3 g added at the front, not total aircraft mass. After initial processing, the owner confirmed session 3 had the same side-weight configuration as session 1. Four recordings exist although the owner describes three tests; no session is merged or discarded. Matched mass and measured center of gravity are not established.

## Reproduction from the original source
1. Obtain the original file from the owner and verify the entire-file SHA-256 above. Decode all four sessions independently in recorded order using a Betaflight-compatible Blackbox decoder. Record decoder/version and unit conversion used; these details were not supplied here, so complete independent reproduction remains conditional on them.
2. Use decoded gyroADC axes in deg/s and rcCommand[3] in microseconds. Per-sample gyro vector magnitude is `sqrt(x² + y² + z²)`.
3. Overall gyro RMS is `sqrt(mean(gyroADC[0]² + gyroADC[1]² + gyroADC[2]²))` across the session's samples. Gyro p95 is the 95th percentile of vector magnitude. The percentile interpolation convention was not supplied.
4. Apply the existing baseline screen inclusively: throttle rcCommand[3] 1140–1400 µs AND each absolute roll/pitch/yaw setpoint <=10 deg/s simultaneously. Low-command RMS and p95 apply ONLY to this screened population. Proxy seconds = screened sample count / approximately 407.2 Hz; supplied rounded seconds are retained unchanged.
5. Calculate battery/throttle medians and throttle p95 from decoded samples, not from bin averages. Retain the supplied summaries verbatim; do not dynamically reconstruct RMS or percentiles from means.
6. Bin samples into elapsed recording seconds starting at zero, using arithmetic means of instantaneous gyro vector magnitude, logged battery voltage and throttle. Keep the partial last bin for each session. Exact timestamp origin/boundary convention was not provided; the report uses the supplied bin assignments unchanged.

## Download contracts and rounding
`weight-placement-003-summary.csv` contains the 14 supplied fields in supplied order: durations at 3 decimals, gyro statistics/proxy seconds/battery medians at 2 decimals, integer counts and throttle summaries.

`weight-placement-003-timeseries.csv` uses the exact requested header: `session,placement,second,mean_gyro_magnitude_deg_s,mean_battery_voltage_V,mean_throttle_us,sample_count,low_command_samples`. All 94 supplied rows (5/14/13/62) are preserved. Means retain 2 decimals for magnitude, 3 for voltage and 1 for throttle, including trailing zeros. Counts are integers. CSV provenance is this companion methods document and the hash above; no extra comment rows alter the requested headers. Both CSVs are generated from the same browser-safe evidence module used by the report.

Lines connect supplied bin means as visual guides, not raw higher-frequency samples. The gyro curve is MEAN MAGNITUDE, not RMS. Voltage/throttle axes are truncated and each session/signal auto-fits; chart scales are not identical across selections. Gyro axes start at zero and include spikes. Final bins contain 149/22/346/278 samples respectively.

## Interpretation limits
Recorded duration is not verified airborne time. Rates in deg/s are angular velocity, not tilt angles, GPS position, velocity over ground, altitude or verified hover/position holding. Low-command screening is a proxy, not proof of hovering.

Session 4 low-command RMS is 9.05 deg/s versus 17.83 and 14.44 deg/s in side sessions 1 and 2. This is descriptive only. Side session 3 records 12.56 deg/s; its configuration was owner-confirmed after initial processing. Session 2's large late spike could be a collision or maneuver, but neither is independently verified; its overall RMS must not be ranked as flight steadiness. Overall and screened statistics cover different populations.

Side recordings are short. Recording windows, battery state (session 2 median 4.05 V versus session 4 3.87 V), pilot inputs and conditions differ. Side mass is unspecified, so this is not a matched-mass experiment. No causal, statistically significant stability, efficiency or endurance conclusion is established. Logged voltage is not independently calibrated. Raw samples, exact decoder/version, percentile convention and timestamp semantics are not bundled or supplied; precision is not evidence of independent verification.

Experiments 001 and 002 remain separate and unchanged; this report does not reclassify their evidence.