OCEAN6 / DEVELOPER REFERENCE / 0.2
Run a task. Inspect the result.
Edit inputs, execute a reference workload in your browser, then replay independently in Python. Inspect the contract, source and versioned release files.
Local task lab
Turn a document into cited search results. Edit text and terms, inspect source lines and ranking, then replay the result. This is a minimal task-definition, execution and checking workflow.
Execution scope: deterministic line retrieval in this browser. No LLM, GPU, Ocean6 remote node or payment service is called. The lab code does not upload your document.
Run it on your own machine
Download and extract the v0.2 bundle. Use Python 3.9+ and the standard library; no account, wallet or third-party packages. Replace task.json with the browser export to execute the same task.
python3 reference_runtime.py run task.json result.json receipt.json
python3 reference_runtime.py verify task.json result.json receipt.json
python3 test_reference.py
# Compare a result exported by the browser (optional)
python3 reference_runtime.py compare task.json browser-result.jsonrun writes a result and unsigned local receipt; verify independently replays and checks SHA-256 of the input / output files. A receipt does not authenticate a remote executor or trusted time. Use compare to cross-check a browser-exported result.
A bounded workload contract
operation = literal-line-search, version 0.2. This is a retrieval baseline. It does not generate answers, measure semantic similarity or establish the truth of input content.
Scroll horizontally to view all columns.
| Field / rule | v0.2 behavior |
|---|---|
| Input | Accept only version, operation, document, terms and top_k; reject additional fields. |
| document | Unicode text, 1–20,000 UTF-8 bytes; reject lone surrogates. Normalize CRLF / CR to LF. One-based line numbers include empty lines. |
| terms | 1–16 strings of 1–80 Unicode code points, without newlines. Trim leading / trailing spaces and tabs, lowercase ASCII A–Z and deduplicate; normalized terms must be nonempty. |
| Matching | Literal substring matching per line. Only ASCII case folding; Chinese is matched literally without segmentation. Score counts distinct matched terms. |
| Ranking / top_k | Descending score, then ascending source line. top_k is an integer from 1–10. No matches returns an empty list. |
| Output | total_matches counts matching lines before truncation. matches contains line, score, text and matched_terms. Text comes from the input, not a model. |
Source, specification & release files
Inspect files individually or download the bundle. JavaScript and Python provide separate implementations, with an example task, boundary tests and replay checks.
2026-09-19 · v0.2: executable local task, independent Python replay, task schema, negative tests, admission / failure-handling proposals, threat model and release evidence gates. The v0.1 authored receipt fixture remains available.
v0.1 receipt format & fixed fixture
These earlier materials describe a different reference workload. Their events are not network records produced by the v0.2 local execution.
Reproduce the example
Use Python 3.9+ with no third-party dependencies. Extract the reference bundle and run the commands below from its directory. Change result.txt and rerun: the digest check should fail.
python3 verify_receipt.py example.json result.txt input.txt
# PASS: offline fixture integrity, correlation and lifecycle checksVerify the fixture in your browser
Calculate SHA-256 over the original UTF-8 bytes of result.txt, then check transitions, record correlation and timestamp order. The tamper demonstration changes an in-memory copy only.
Proposed data contract
These fields support integration design discussions. They are not an available API. Submission endpoints, authentication, limits and signing protocols require a separate engineering release.
Scroll horizontally to view all columns.
| Field | Definition |
|---|---|
| task_id / attempt_id | Stable task and independent attempt identifiers. Retries preserve history. |
| input_sha256 | Digest of the input file's original bytes, not arbitrary reserialized JSON. |
| acceptance_profile | A versioned acceptance method fixed before execution. |
| output_sha256 | Digest of the delivered file's original bytes. |
| provider_id / runtime_version | Declared provider and environment; this example does not authenticate either. |
| events | Ordered states and UTC timestamps for this attempt. |
| settlement | Proposed payment record; null in this example. No real payment. |
{
"task_id": "example-task-001",
"attempt_id": "example-task-001-a1",
"provider_id": "example-provider",
"runtime_version": "offline-fixture/0.1",
"output_sha256": "a84ccddc948c4d9400b7f3bf22ff5a0e6cf02b3e9b5b1a657d1922871de0df2a",
"delivered_at": "2026-09-19T00:00:03Z",
"status": "delivered",
"settlement": null
}What does the verifier establish?
It checks required reference fields, identifier correlation, supported transitions, UTC time ordering, delivery time and the output digest. It is not a general JSON Schema engine. It does not validate provider signatures, model execution or payment. If an attacker replaces both receipt and output, a digest alone cannot detect fabrication; a production system needs trusted signatures or another independent trust source.
Versioned downloads
Download individual files or the complete reference bundle. SHA256SUMS.txt checks download integrity; it does not independently authenticate the publisher.
2026-09-19 · v0.1: technical design proposal, reference data contract, fixed fixtures, Python and browser verifiers. This release covers documentation and a reference implementation, not completion of the core network.
Contribute a technical review
For workloads, integration requirements or existing technical evidence, include the reference version and section in your enquiry. Do not email private keys or sensitive customer data.
business@ocean6.ai