OCEAN6 / EVIDENCE / 20 SEP 2026

Engineering evidence. Open to inspection.

Inspect what is implemented, reproduce the local checks and follow the release record.

Run the reference Download evidence bundle

9 / 9Python tests passed

80 / 80Cross-runtime comparisons

PASS → FAILOriginal / modified result

Self-published local checks, recorded 20 Sep 2026. These are not an independent security audit, network performance benchmark or AI inference result.

What is available today

Local text retrieval

Released · v0.2

Python and JavaScript implementations with a defined input contract. Inspect

Replay & byte integrity

Released · v0.2

Independent local replay and SHA-256 checks; unsigned receipts. Inspect

Node admission & scheduling

Proposed

Requirements and failure handling are documented; no deployed service is established by this release. Inspect

Remote model / GPU execution

Not demonstrated

Requires a deployed executor, environment details and reproducible workload records. Inspect

Identity, payments & settlement

Proposed

Authenticated execution records and payment reconciliation require separate implementation evidence. Inspect

Inspect the validation records

report.json ↓python-tests.txt ↓cross-runtime-tests.txt ↓task.json ↓result.json ↓receipt.json ↓replay-check.txt ↓modified-result.json ↓tamper-check.txt ↓SHA256SUMS.txt ↓README.md ↓

Reproduce the checks

Extract the evidence bundle, then run these commands in its directory. Python 3.9+; Node.js is needed for the cross-runtime test. Recorded environment: Python 3.12.14, Node.js v24.19.0, Linux x86_64.

python3 test_reference.py
python3 test_cross_runtime.py
python3 reference_runtime.py run task.json result.json receipt.json
python3 reference_runtime.py verify task.json result.json receipt.json
python3 reference_runtime.py verify task.json modified-result.json receipt.json
# Expected final command: FAIL REPLAY_MISMATCH (exit 1)

The example uses a synthetic shipping policy. The first result scores 2; the modified copy scores 9 and fails replay. This is a software example, not a customer case. Rerunning creates new local receipt timestamps.

Verification boundaries

Replay checks the result against the supplied task. File hashes check byte integrity against the supplied record. Neither authenticates a remote executor, proves model execution or proves payment. If all records are consistently replaced, these checks alone cannot establish the original truth.

Release history

2026-09-20

Evidence publication & product walkthrough

Published reproducible validation logs, downloadable task records, the 48-second walkthrough and a pilot enquiry workflow. Software reference remains v0.2.

2026-09-19

Reference 0.2 · Documentation 1.1

Released deterministic text retrieval, independent Python replay, JavaScript implementation and English technical documentation.

Evidence required for the next stage

Remote inference evidence must identify the model and version, runtime, hardware, task inputs, outputs, timing method and failure cases. Performance comparisons need a stated baseline and test conditions. Hardware availability, customer outcomes and team credentials will require their own attributable source records.

Discuss a pilot workload