Sports Lib combines automated tests with testing against real activity and route exports. This guide documents the corpus validation through 4 November 2026, alongside the project's earlier testing history.
The corpus has driven fixes for exporter quirks, incomplete recordings and malformed values that previously caused import failures, lost observations or incorrect summaries. Sports Lib includes tested recovery for supported file irregularities as well as standard-format imports.
| File irregularity | Supported handling |
|---|---|
| Vendor-specific GPX and TCX layouts | Recover recorded speed, heart rate and altitude from supported alternate locations and namespaces, using established exporter units and preserving standard-field precedence. |
| Nonstandard FIT record headers | Recover the evidenced reserved-header-bit variation while retaining the other format and integrity checks. |
| Missing or conflicting session timing | Retain credible session bounds and recover later observations when independent timer, record and distance evidence establishes their ownership and timing. |
| Missing final multisport session | Recover a final Unknown Sport leg only when the file explicitly establishes a missing session and independently supports its boundaries. |
| Pool and multisport export quirks | Preserve supported length/rest records, canonical pool units and valid cross-sport laps; guard against incomplete generated distance and source-proven impossible length durations. |
| Invalid values in older native JSON | Restore activity and route data without promoting non-finite or null scalar summaries into valid measurements; preserve valid zeros and independently usable data. |
| Impossible FIT power-balance percentages | Treat invalid balance as missing on both sides, preserving total recorded power and valid 0–100% values instead of manufacturing negative side watts. |
| Suunto diving header summaries | Apply the existing terrain-summary exclusion after the header merge, preserving raw altitude/depth streams and consistent native JSON roundtrips. |
These are scoped compatibility rules backed by source evidence and regression tests. Recovery preserves meaningful zeros, missing readings, pauses and sport boundaries. Unproven values are not invented to make a file pass, and unsupported non-activity files remain rejected by the activity importer.
See Import activities and Work with routes for the supported layouts, recovery conditions and historical reparse requirements. The known limitations below remain separate from these completed fixes.
As of 4 November 2026, the regression corpus contains 212,905 distinct source files. Validation covers activity and route imports and file-type validation. Coverage combines complete replays with focused follow-ups; the dated results below identify each run's scope.
Files are deduplicated by their contents. Different exports of the same workout can remain distinct files, so this count does not represent unique workouts or athletes.
The corpus includes FIT, GPX, TCX, provider JSON and SML, with exports from Garmin, Suunto, COROS, Polar and other sources.
The complete 22 October 2026 replay covered 195,562 source files and included these successfully imported activity and route cases:
| Coverage | Count |
|---|---|
| Activity source files | 188,277 |
| Activity legs | 189,455 |
| Multisport source files | 429 |
| Laps | 814,607 |
| Swimming lengths, including rests | 341,872 |
| Route source files | 1,288 |
Legs, laps and lengths are nested in activity files; these counts must not be added together as unique activities.
Validation included:
Checks vary by format and scenario. Intentional changes were reviewed against source evidence; the completed comparisons found no unexpected differences or newly introduced rejections.
Focused follow-up replays checked all 2,629 JSON/SML originals for the Suunto diving-summary correction and 2,337 balance-related FIT originals for record-balance validation. The latter corrected 100 invalid balance readings in nine files, eliminating 60 negative left-side power samples; the other 2,328 files stayed unchanged. Total power and TSS were unchanged. Both follow-ups checked native JSON roundtrips and repeated generation. These focused runs supplement the complete snapshot above; they are not additional unique source files.
The 4 November 2026 FIT refresh added 16,203 distinct originals, expanding the corpus from 196,702 to 212,905 files. All additions and three existing controls received fresh decoder and library checks, covering 16,206 inputs. The activity imports passed 32,408 native JSON restoration cycles, with imported TSS preserved and no unexpected output differences or newly introduced rejections. This incremental replay supplements the earlier coverage; it does not represent a new complete replay of every retained source.
The 22 October 2026 route validation covered 1,287 GPX files and one FIT file.
Each was imported through the route APIs, compared with the published implementation and restored from native JSON. Comparisons covered the complete serialized output, including geometry, metadata and statistics. No failures or unexpected differences were found.
This provides substantially broader corpus coverage for GPX than FIT routes. Exporting every route through every supported format and manually reviewing every map were outside this audit.
The Strava suites use saved reference streams already in the repository. They are repeatable comparisons with those exports, not a claim of exact equivalence to today's Strava service. The larger regression corpus does not contain paired Strava results for every source, so corpus replay and Strava agreement answer different questions.
Comparisons use the same elapsed timestamp, preserve missing observations, and validate the reference clock. They check sample coverage separately from error: one matching point cannot make an otherwise missing stream pass. Distance comparisons must not accept a value merely because it appears somewhere else in the stream. Grade errors are percentage points of slope, not relative percentage errors. No time shift is chosen to improve agreement.
The October 2026 sparse-distance change retains the existing grade look-ahead and GAP formula. The eight active
legacy Grade comparisons remain within their existing error limits, with 97.6–99.99% reference-point coverage.
Separate checks now measure both Grade and Grade Smooth against Strava's grade_smooth, including these previously
excluded comparisons:
| Saved reference | Grade mean absolute difference | Grade Smooth mean absolute difference |
|---|---|---|
| Suunto uphill | 7.68 points | 5.68 points |
| COROS trail | 2.66 points | 3.03 points |
Both discrepancies predate the sparse-distance change. Their altitude and heart-rate observations match at the reference timestamps. Suunto's reference distance also differs from recorded distance (mean absolute difference 0.68 m, maximum 11.7 m); COROS's distance differences beyond 0.051 m are limited to six startup observations. Correcting the uphill test's unrelated flat-activity reference does not eliminate its grade discrepancy. These observations locate remaining disagreement in derived terrain processing and reference-distance handling; they do not establish which smoothing model is more accurate. Applying a blanket time shift would misalign sensors that already agree.
The saved hilly-running reference also contains pace and GAP. Its new check compares the unitless adjustment
GAP / pace with Speed / Grade Adjusted Speed, requiring at least 4,811 matched observations out of 4,988.
It accounts separately for 176 zero-speed observations and the remaining unsupported observation. Mean absolute
adjustment-factor difference is about 0.0612, with a 95th-percentile difference of 0.2025; this is essentially unchanged
by the distance fix. This isolates the grade adjustment from different speed smoothing and does not establish absolute
GAP pace or whole-workout GAP equivalence. Coverage is one paired GAP reference, not all providers or sports.
The new smoothed-grade and known-discrepancy bounds record existing behavior and catch worsening agreement. They are regression ceilings, not newly relaxed accuracy claims; the stricter legacy Grade expectations remain unchanged. Strava describes its GAP model as fitted to real-world running data. Sports Lib's polynomial approximation and terrain smoothing should be evaluated independently before changing either to chase reference numbers.
Run the focused reference checks with:
npm test -- --runInBand src/specs/strava-streams-compliance.spec.ts src/specs/strava-derived-streams.spec.ts src/specs/helpers/stream-reference-comparison.spec.ts
A successful regression test establishes expected software behavior. It does not independently establish sensor accuracy or physiological validity.
Open investigations cover power-data quality, questionable MET inputs, conflicting FIT timing, missing multisport ownership, remaining swimming chronology and vendor-specific TCX cadence units. These remain tracked in the issue tracker.
Original files and account-level evidence remain outside the public repository. Public documentation reports aggregate results; regression fixtures derived from private findings use synthetic examples.