Dokumen ini mencatat tahap implementasi pada versi yang disebutkan. Untuk status akhir seluruh fitur, baca audit implementasi.
#34 — Calibration from recorded trips
Implemented on 2026-09-19 for Android 0.13.0-calibration. This delivers model fitting and chronological evaluation, not evidence that the road accuracy target is already met.
The recording panel exposes Perbarui kalibrasi dari riwayat. It reads completed native recordings and saves a versioned model report separately in models.json. Subsequent energy predictions, charging plans and native live replanning use the active matching model. Previously saved plans and pre-leg predictions retain their original model and values. The worker also checks the model document snapshot before publishing a result.
Evidence and model limits
Only explicitly non-synthetic completed recordings with matching vehicle/capacity/payload, manual endpoint SOC, continuous precise non-mock GPS and a matching pre-leg route qualify. The trace must span at least 1 km, stay within 10% of planned distance, match endpoints within 100 m, and sampled fixes must stay within a 75 m route corridor. Gaps over 30 seconds and implausible movement are excluded. Endpoint SOC must cover the leg within 30 seconds; a leg containing charging is excluded and must be recorded separately. Increasing SOC within a leg is currently excluded as ambiguous rather than attributed to regeneration. These strict filters can leave real recordings uncalibrated; the UI shows why.
Models use the median ratio of actual SOC consumption to the base scenario for the latest 20 eligible legs per vehicle, capacity, rounded 10 kg payload group and terrain band. At least three eligible legs from two distinct trips are required. Terrain uses matched planned-route elevation, not measured motor energy. The middle estimate is adjusted, while conservative consumption remains at least the original scenario plus its 20% margin. Historical model versions retain their support IDs and creation/training times; removed or changed supporting evidence invalidates them.
Evaluation
Every quality-qualified observation can be evaluated against its immutable pre-leg prediction, including large errors and small consumption that are unsuitable for training. A calibrated evaluation requires the model to have existed before the prediction and leg, with no support from that trip. Late predictions, inconsistent initial SOC, unknown models and invalidated evidence are excluded with reasons.
The UI reports predicted/actual arrival SOC and signed error per leg. Calibrated holdout MAE, maximum absolute error and the count exceeding 5 percentage points are separate from reserve breaches and conservative coverage. An uncaptured reserve is unknown, never assumed to be 20%.
The provisional evidence gate requires five holdout legs from three new trips. Any observed error over 5 pp reports not-met; insufficient evidence reports unverified. Passing this gate describes only those observations, not all profiles/terrain or a guarantee for future travel. No actual road dataset currently establishes the target.
Verification
- 64 Node tests pass, including ten calibration boundary tests: synthetic/incomplete exclusion, future prediction and restart, immutable old predictions, cohort isolation, detour, chronological holdouts, duplicate/broken/SOC-invalid recordings, invalidation, extreme and small-consumption errors, and malformed model persistence.
- 19 JVM tests, Android debug assembly and lint pass. No new location owner or wake lock is introduced.
- TECNO KF8 runs the bundled calibration module and native
OmoxModelsstorage. Synthetic histories produce zero models and explicit exclusions; force-stop/reopen preserves the report and displays both uncalibrated and unverified status. Fixtures are restored. Evidence/helper remain ignored atresearch/runtime/issue34-device.jsonandcheck-calibration-device.py. - Standards review found malformed model archives could escape the application error boundary; nested report validation and a visible fallback warning now preserve the document and use the base scenario. Spec review found training support filters could hide large evaluation errors; training eligibility is now separate and regression-tested.
Physical checks are integration tests. Generated fixtures are never retained as real training history or claimed as field accuracy evidence.