Skip to main content
CerebraTech AI
/compatibility · evidence registry

Evidence-backed technology compatibility matrix

See which model, runtime and hardware combinations were tested, are in support scope or remain untested—without turning a status into a compatibility guarantee.

This is a decision aid sourced from the Evidence Registry, not a promise that every device fits every site. Where benchmark values do not exist, the record shows a test plan and limitations instead.

Compatibility records

Every record carries an evidence ID, method, owner and review date so a team can trace it before using the combination on a real workload.

6 records

Compatibility records
ModelRuntimeVersionHardwarePrecisionCheckedStatusEvidenceMethod / scopeLimitationsOwner / review
Custom YOLO detectorTensorRTModel and engine version pending evaluationJetson Orin Nano Super with industrial camera and LED barFP16 to be confirmed2026-09-12not-testedEV-2026-001Run the versioned vision acceptance plan with recorded camera geometry, lighting, line speed and stream count.
  • This internal note has no publishable performance result.
  • A new installation requires a repeat evaluation before compatibility is claimed.
lab-team2026-09-122026-12-12
Workload-specific modelEdge runtime selected per workloadNot pinnedEdge appliance selected after site assessmentNot specified2026-09-12supported · in scopeEV-2026-002Confirm the deployment boundary and runtime/model selection during Solution Architecture review.
  • Supported describes the review scope, not a performance or portability guarantee.
  • Remote support, updates and backups remain separate operating paths.
engineering-team2026-09-122027-03-12
YOLOv12 prototypeCustom stereo calibrationPrototype revision not releasedTwo CCTV cameras, Raspberry Pi 5 and ESP32 sensorsNot recorded2026-09-12testedEV-2026-004Internal stereo-camera sampling run using the stated prototype hardware and calibration flow.
  • Tested means this prototype ran; it does not prove weight-prediction accuracy.
  • Sampling design and accuracy require a new approved evaluation.
lab-team2026-09-122026-12-12
Custom vision detectorRuntime under evaluationPending measured evaluationIndustrial camera and LED bar; final line hardware pending assessmentTo be selected with the acceptance plan2026-09-12not-testedEV-2026-005Freeze the dataset, model/runtime versions, line conditions and acceptance threshold before measuring.
  • No precision, recall, false-positive or throughput number is published.
  • Results cannot be transferred to another line without a repeat test.
lab-team2026-09-122026-12-12
Document retrieval and grounded-answer modelPaddleOCR + llama.cppVersions to be pinned at evaluationRTX 4090 edge workstation; final deployment hardware pending assessmentTo be recorded with model package2026-09-12supported · in scopeEV-2026-006Use the versioned corpus and technician-authored question set with OCR, chunking and human-review notes.
  • Supported is an in-scope architecture path, not evidence of answer quality.
  • No retrieval recall, grounded-answer rate or citation coverage value is approved.
lab-team2026-09-122026-12-12
Model/input package pending benchmarkTensorRTRuntime and model version pending benchmark runJetson Orin Nano Super; exact SKU and thermal profile pending evaluationTo be recorded with power mode and input shape2026-09-12not-testedEV-2026-007Run repeated latency and thermal measurements with resolution, stream count, preprocessing, power mode and ambient temperature recorded.
  • No latency, FPS, power or temperature value is published by this plan.
  • A hardware or runtime change triggers a new evaluation before replacement claims.
engineering-team2026-09-122026-12-12

How to read and update the matrix

  • tested means the stated setup was run; it does not mean every workload is compatible.
  • supported means the combination is inside the scope the team is willing to review and operate; it is not benchmark evidence.
  • not-tested means there is no compatibility evidence yet and must never be read as compatible by default.
  • When the model, runtime, version, hardware or precision changes, open a new evaluation and review record before making a replacement claim.
  • If no benchmark exists, publish the test plan, conditions and limitations instead of a number.

Canonical source: content/registry.json → compatibility and evidence. The same EV IDs are used by Labs, Work and Trust.

How to update a record

The owner pins model/runtime/hardware/precision, links a method or test plan, checks the evidence ID and sets the next review date before publication. Any component change invalidates a replacement claim until the combination is evaluated again.

Evidence to inspect: EV-2026-001EV-2026-002EV-2026-004EV-2026-005EV-2026-006EV-2026-007

Related test method example: Lab Note: vision camera lighting

This site uses very few cookies

We use only the cookies necessary to remember your language choice and save this consent preference (your chosen theme is remembered via browser local storage, not a cookie). Our analytics tool uses no cookies and collects no personal data.

Read the cookie policy