Public technical boundary · version 1.0
Serious video systems make uncertainty inspectable.
This is the architecture a prospective operator can evaluate without exposing credentials, paid-provider controls, private corpus identities, participant details, storage locations, or internal network topology.
System model
Source identity in. Versioned boundaries out.
The assisted-pilot shape deliberately separates source qualification, model evidence, interval decoding, review, and output identity. Rally detection does not silently become automatic match or game splitting.
Contract direction
Small invariants are more useful than a large vague API.
Source timebase
Integer millisecond boundaries are expressed against the original recording timeline, not an unnamed proxy.
Ordered intervals
Every end follows its start; windows are source ordered and non-overlapping.
Explicit identity
Schema revision and the actually served model release travel with an accepted pilot handoff.
No decorative confidence
No confidence field is published until its semantics and calibration form a real public contract.
Visible partiality
Ambiguous, unsupported, or partial output must remain visible rather than being reported as a silent success.
Evidence method
The result is never separated from the ruler.
Results on request
Evaluation results, including tolF1, are available on request with the applicable context attached.
Candidate vs served
A development candidate is not represented as the model serving a future pilot. The actual served release is identified per accepted run.
Locked vs selected
Locked operating points remain distinct from calibration selected on development evidence.
Scope stays attached
Aggregate and sport-specific interpretation travels with representation, imbalance, and panel-history notes.
Scorer semantics
Matching rules, tolerances, aggregation, operating points, and limitations accompany any disclosed result.
Fresh confirmation
Fresh recordings remain a gate before broader performance language.
Owned golden labels
Human boundaries become the ruler through a controlled chain.
The capture session—not an extracted clip—is the split unit because clips from one sitting share camera, venue, lighting, and players. Automated checks can find malformed, overlapping, or suspicious labels; they cannot certify that a human missed no rally.
Dronacharya evaluation uses owned, human-labelled rally boundaries with session-aware data roles. Exact training inventory, sport distribution, split roles, and quality-control findings are supplied only with a contextual evaluation brief; the public site does not turn dataset counts into a performance claim.
Plain-language scorer
What tolF1 means here.
- A human marks each rally start and end against the source timeline.
- The system predicts intervals on that same timeline.
- Maximum-cardinality one-to-one matching prevents one 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.
Evaluation results, including tolF1, are available on request. Exact tolerances, aggregation details, panel construction, operating points, and limitations are supplied with the figures.
Conformance is not model quality. A consistent scorer does not prove representative data, media rights, venue generalisation, or a production serving release. Those remain separate evidence questions.
Release map
What exists, what is assisted, and what is not public.
| Surface | Status | Boundary |
|---|---|---|
| Model method, capture, resources and evidence | Public beta | Inspectable pages and downloads |
| Rally-index example and JSON Schema | Preview | Synthetic direction, no public endpoint |
| Evaluation brief | On request | Email KhelVision; no submission form on this site |
| Media intake and processing | Assisted | Only after fit, rights, transfer, retention, and deletion are agreed |
| General API, callbacks, batch, live stream | Not public | No compatibility or service-level promise |
| Named candidate serving claim | Not made | Development candidate evidence remains separate from serving identity |
Marketing site
Read the method. Request the context.
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.
Failure semantics
No silent success.
- A source that cannot be qualified does not quietly become a valid run.
- A model candidate, deployment readiness signal, and actually served release are separate facts.
- Missing or ambiguous intervals remain reviewable; absence is not rewritten as zero.
- A partial output is not described as complete, and a liveness check is not treated as proof of an accepted media job.
- Release receipts and public evidence state what they prove and what they do not.
Safe publication
Transparency does not require an infrastructure leak.
Public material intentionally excludes credentials, authenticated endpoints, private hostnames, provider and billing 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 operational assurance.
What remains public is the useful layer: capture guidance, system boundaries, schema direction, synthetic examples, the evaluation-disclosure policy, scorer semantics, limitations, privacy and terms, and the public labeling contracts behind owned golden labels.