# KhelVision public technical overview

Version 1.0 - 30 August 2026

This document explains the public product and evidence boundary without exposing credentials,
provider endpoints, private corpus identities, participant details, storage locations, or internal
network topology.

## System model

1. **Qualify the source.** Establish authority, sport, original timebase, capture profile, and the
   agreed intake and deletion boundary.
2. **Preserve source identity.** Work against the original recording timeline. Do not silently
   change the timebase or treat a proxy as the source of record.
3. **Index candidate play windows.** A named model release produces temporal evidence which is
   decoded into candidate rally intervals under an explicit operating point.
4. **Review the result.** A human-visible review boundary remains part of the assisted-pilot shape;
   detection is not automatic match or game splitting.
5. **Deliver a versioned result.** Boundaries, source duration/timebase, schema identity, and the
   actually served model release travel together in the pilot handoff.

## Result invariants under exploration

- Integer millisecond boundaries use the source recording's timebase.
- Every end is greater than its corresponding start.
- Windows are source ordered and non-overlapping.
- Schema and served-model identity are explicit.
- No confidence field is published until its semantics and calibration form a real contract.
- Partial, ambiguous, or unsupported results must remain visible rather than being reported as a
  silent success.

The downloadable JSON Schema and synthetic example on the API preview page are a direction for a
qualified pilot, not a generally available endpoint.

## Evaluation discipline

Evaluation results, including tolF1, are available on request. A qualified brief keeps the
following context attached to any disclosed figure:

- candidate identity remains separate from the model actually served to a future pilot;
- locked operating points remain separate from calibration selected on development evidence;
- aggregate and sport-specific interpretation stays attached to representation and imbalance;
- reused, adaptive, model-selecting, fresh, and independently held-out evidence are not conflated;
  and
- metric semantics, evaluation scope, operating points, selection status, and limitations travel
  together.

The open website intentionally omits standalone performance figures, dataset counts, evaluation
substrate details, and model-selection deltas. Private-only commits and artifact digests are not
presented as public reproducibility.

## Golden-label chain

1. Preserve an authorised original and its source timeline.
2. Have a human mark rally starts and ends against a frame-identical annotation view.
3. Validate schema, timing, ordering, and suspicious boundaries, then perform human review.
4. Freeze the labels and their training, development, or held-out role.
5. Score predictions against one agreed ruler while keeping failures visible.

The capture session, not an extracted clip, is the split unit because clips from one sitting share
camera, venue, lighting, and players. Automated checks find malformed or suspicious labels; they do
not certify that a human missed no rally. Exact inventory, sport distribution, split roles, and
quality-control findings are part of the request-only evaluation brief.

## tolF1 in plain language

- A human marks each rally's start and end on the source timeline.
- The system predicts intervals on the same timeline.
- Maximum-cardinality one-to-one matching prevents a prediction from claiming multiple human
  rallies.
- A match must satisfy the agreed start and end tolerance windows.
- F1 is calculated per recording and aggregated across the agreed panel so a long recording cannot
  silently dominate the result.

Exact tolerances, aggregation details, panel construction, operating points, and results are
supplied together on request. Scorer conformance does not prove model quality, representative data,
media rights, venue generalisation, or a production serving release.

## Public, pilot, and not-yet-public surfaces

| Surface | Current status |
| --- | --- |
| Product, capture, technical, legal, and evaluation-policy pages | Public beta material |
| Synthetic rally-index example and JSON Schema | Downloadable preview |
| Evaluation brief | Available by email; no form on this marketing site |
| Video upload or URL import on this site | Not available |
| Authenticated media intake and processing | Arranged only for an accepted assisted pilot |
| General processing API, callback, batch, or live-stream contract | Not available |
| Named candidate serving claim | Not made |

## Public-site data path

This marketing site serves static descriptions, synthetic examples and downloads. It accepts no recordings, account credentials or form submissions and starts no inference. Cloudflare processes ordinary request metadata to deliver and protect these pages. To request a contextual evaluation brief, email hello@khelvision.com; use the KhelVision app site for product access and account policies.

## Publication boundary

Public documentation intentionally excludes secrets, authenticated endpoints, private hostnames,
provider identifiers, corpus and participant locators, local storage paths, internal incident
details, and commands that can mutate paid infrastructure. Those details are unnecessary to
evaluate the public product and would weaken, rather than strengthen, operational assurance.

## Inspectable resources

- Evaluation method and request path: /evidence
- Result-shape preview: /schemas/rally-index.preview.v1.schema.json
- Synthetic example: /examples/rally-index.preview.v1.json
- JSON Schema: /schemas/rally-index.preview.v1.schema.json
- Evaluation availability statement: /evidence/evaluation-availability.json
- Camera recording guide: /capture
- Resource library: /library