DexHoldem

Benchmark submission

Submit a model.
Prove the interface first.

Register a dexterous policy or embodied agent for DexHoldem. Every packet must pass an interface contract, a safety screen, and an abuse-risk gate before it can enter the evaluation queue.

01 Validate locally 02 Challenge safely 03 Evaluate in queue

Admission workspace

Prepare the submission packet.

Credentials stay in memory and are never written to browser storage. This static page demonstrates the admission contract; a production deployment must repeat every check at the edge and on the evaluation worker.

01

Choose a track

The selected track changes the required runtime contract.

02

Identity & provenance

Enough information to reproduce and audit a result.

03

Runtime endpoint

Only public HTTPS inference endpoints are admitted.

04

Benchmark contract

Confirm the required agent capabilities.

Complete the required fields, then run the preflight.

Abuse & DDoS sentry

Six discriminators. Four outcomes.

Except for hard protocol violations, no single soft signal blocks a submission. The gateway combines identity, traffic, replay, payload, endpoint, and resource-cost evidence into a risk score, then applies the least disruptive safe action.

01 / Volume

Token-bucket pressure

Score per account, organization, IP, and ASN. Separate short bursts from sustained volume.

1 / 15 min · 5 / day / org
02 / Behavior

Interaction plausibility

Compare completion time, field order, cadence, and device continuity. Escalate; do not silently fingerprint.

fast + repeated + low entropy
03 / Replay

Duplicate packet detection

Hash normalized manifests and enforce single-use idempotency keys and signed nonces with a short TTL.

body hash + nonce + timestamp
04 / Payload

Shape and size bounds

Reject non-JSON bodies, deep nesting, long strings, unknown fields, decompression bombs, and oversized manifests.

64 KiB · depth 8 · strict schema
05 / Endpoint

SSRF and egress safety

Resolve only public IPs, pin DNS results during a probe, cap redirects, and block link-local or private ranges.

HTTPS · public IP · 0 redirects
06 / Cost

Evaluation budget guard

Reserve quota before work, use one active job per team, bound response time and bytes, then trip the circuit breaker.

queue · lease · hard timeout
0–29AllowIssue signed probe nonce
30–59ChallengeCAPTCHA or small proof of work
60–79QuarantineManual review, no evaluator slot
80–100Reject429/403 with retry window

Portable handoff

A reviewable manifest, without secrets.

The packet records benchmark compatibility and operational limits. Authentication material is attached only to the server-side probe job and never copied into the manifest.

Read agent contract Read policy contract
submission.json