What's new

Film analysis now consistently uses folkstyle terminology and result logic instead of switching rulesets based on a model's guess. Scoring remains subject to visible-board evidence and coach review.

• Tell Gemini and sampled-frame fallbacks to analyze all film as American scholastic/collegiate folkstyle.
• Always label analyzed reports as folkstyle and apply folkstyle decision, major-decision, and technical-fall thresholds.
• Avoid guessing near-fall points, riding-time awards, or win method when competition level or official evidence is unclear.
• Add regression coverage for folkstyle behavior when a model returns an international style label.

A timed-out Gemini request for a public video link now advances to the next configured Gemini model instead of retrying the same slow model and exhausting the analysis time budget.

• Treat network timeouts and aborted provider requests as retryable for linked-video analysis.
• On timeout, move directly to the next Gemini model rather than repeating the same request.
• Explain before analysis that Hugging Face sampled-frame fallback requires uploading a video file.
• Add a regression test proving a timed-out linked-video request can recover with a configured alternate model.

Sampled-frame fallback reports no longer dress up generic model guesses as scouting findings, and Hugging Face requests use its documented provider router.

• Remove unverified strengths, improvements, scouting and move-rate claims from sampled-frame fallback reports.
• Route Hugging Face vision requests through its documented OpenAI-compatible endpoint with automatic provider selection.
• Use fewer fallback provider attempts and avoid response-format parameters that some hosted vision providers reject.

When the sampled-frame analysis needs Gemini as backup, the report now explains why Hugging Face did not return a result and omits move-success percentages unless scoring is reconciled.

• Expose a redacted Hugging Face failure summary (or missing-token notice) alongside Gemini fallback results.
• Log sanitized Hugging Face provider failures for deployment diagnosis without exposing credentials.
• Hide inferred move attempts and completion rates whenever the official-score ledger remains disputed.

Uploaded-film retries now exercise Hugging Face before Gemini, and a final score that conflicts with confirmed scoring events is no longer presented as official.

• Try configured Hugging Face vision providers first after a native-video token-limit failure, with Gemini sampled-frame analysis as backup.
• Suppress an uncorroborated final score, winner, and win method when the verified event ledger does not reconcile with it.
• Add regression tests for fallback provider order and conflicting scoreboard totals.

The vision fallback now tries explicit Hugging Face provider routes as well as the shared router, with smaller sampled-frame payloads and clearer network diagnostics.

• Route each configured model directly to its selected Hugging Face inference provider before retrying through the shared router.
• Retry through the shared router when an explicit provider route loses its network connection.
• Send fewer chronological image sheets to reduce fallback request size while retaining the reference and match coverage.

Uploaded matches that exceed Gemini's native video token limit now retry as sampled images through the configured Gemini API before using external vision fallbacks.

• Use browser-extracted contact sheets for a Gemini image-based retry when native video input exceeds the token limit.
• Keep fallback results clearly marked as low-confidence and requiring coach review.
• Expose the reason and provider when sampled-frame fallbacks also fail.

When a primary video read exceeds its token limit, uploaded matches can now retry through the Hugging Face vision fallback using the sampled frames already prepared by the website.

• Pass browser-extracted contact sheets to the Hugging Face fallback when native video analysis fails, instead of checking only the primary request payload.
• Show a clearer failure reason when the fallback token or sampled frames are unavailable, or when a Hugging Face provider rejects the request.

Public YouTube match analysis now uses Gemini's adaptive video mode to inspect relevant moments more closely, while disputed scoring remains flagged for coach review.

• Route public linked videos through Gemini's agentic video-understanding API instead of fixed-rate video sampling.
• Use Gemini 3.8 Flash by default, with adaptive-capable Flash models as fallbacks.
• Preserve independent score audits and coach-review safeguards; adaptive analysis improves inspection but does not guarantee correct calls.

When independent film reads disagree, match reports now hide disputed running scores and label those actions unverified.

• Do not show a running scoreboard transition for a scoring event that needs coach review.
• Clearly distinguish possible scoring actions from verified events in the timeline.
• Label displayed match totals as AI-read final scores so they are not mistaken for reconciled event ledgers.

Scoring events that disagree between independent video-analysis passes are now flagged for coach review instead of being reported as verified.

• Cross-check scoring category, wrestler, score transition, and timestamp across two independent analysis passes.
• Downgrade disputed scoring calls and overall confidence when the passes disagree.
• Mark timestamps approximate when sampled footage cannot establish the exact action moment.

Match reports now use athlete names read from film, and wrestler links are limited to matches owned by the signed-in coach.

• Show film-identified wrestler names in the match title and score labels instead of generic upload placeholders.
• Restrict wrestler-link reads to the signed-in user's own matches.

Improved public sharing details and made the product experience clearer before sign-in or purchase.

• Added Terms of Service and Privacy Policy draft pages, plus sitewide contact and legal links.
• Added an illustrative sample scoreboard and scouting report to the homepage.
• Updated social preview metadata and made AI score estimates clearer about review and uncertainty.
• Improved match loading feedback and OAuth sign-in error handling.

A new home for product changes and improvements.

• Added a public updates page with release dates and version numbers.
• Made release notes accessible from the public site and signed-in navigation.