anthropic/claude-sonnet-5
Amazon Bedrock · amazon-bedrock/global
UntestedObserved 2026-09-08. Endpoint price snapshot: $2 / $10 per million input / output tokens.
These are vendor metadata, not measured TENX quality, latency, availability or spend.
TENX / AI CAPABILITY LABORATORY
Capability is a tested work system, not a model name.
Model, provider, sources, tools, memory and authority form one versioned configuration. A model must earn permission for each workload.
Current release boundary
Not activatedThis page describes the engineering laboratory. It is not live spending telemetry or an operational AI assessment. No model has been qualified by this release.
The existing controlled missions remain scripted. Public real applicants, hiring verdicts and advanced scoring stay disabled.
The operator-only status endpoint requires the existing administrator credential. This public page neither reads nor accepts credentials.
Dated model screening
google/gemini-3.8-flash
Google · google-vertex/global
UntestedObserved 2026-09-08. Endpoint price snapshot: $0.75 / $3.75 per million input / output tokens.
These are vendor metadata, not measured TENX quality, latency, availability or spend.
qwen/qwen3.8-flash
Alibaba · alibaba
Privacy screen refusedObserved 2026-09-08. Endpoint price snapshot: $0.15 / $0.47 per million input / output tokens.
These are vendor metadata, not measured TENX quality, latency, availability or spend.
Twelve independent workload gates
36 development probes, 36 qualification probes and 36 separate final-holdout probes. These small analytic tasks do not represent the full diversity of enterprise work. Final holdout has not been executed.
role design
human accountability · little or no ai allowed · review burden · uncertainty
Independent review requiredoccupational synthesis
evidence not employer fact · source fidelity · transfer assumptions · unknowns
Independent review requiredscenario authoring
feasible authority · multiple solution paths · no hidden answer · draft only
Independent review requiredscenario qa
numeric reconciliation · constraint conflict · solvability · publication refusal
Independent review requiredsimulated stakeholder
limited knowledge · role consistency · no invented approval · ordinary disagreement
Independent review requiredcandidate collaborator
visible information only · source fidelity · verification · no authority escalation
Independent review requiredskeptical reviewer
error detection · false alarm control · bounded review · no ranking
Independent review requiredsource grounding
exact citations · no invented sources · conflict preservation · abstention
Independent review requiredartifact critique
current revision · constraint check · specific correction · no inferred cognition
Independent review requiredevidence extraction
event fidelity · missingness · observation not inference · no scores
Independent review requiredevidence summary
balanced coverage · contradiction preservation · provenance · no verdict
Independent review requiredstructured transform
schema · numeric fidelity · units · no extra fields
Independent review requiredWhat the laboratory enforces
- Exact dated model and provider request, no fallback, no private reasoning storage.
- Explicit current-revision sharing before an actor receives a document.
- Bounded read-only tools; model proposals never grant approval or database authority.
- Durable operator-local reservation before dispatch, a $2 maximum laboratory budget, and no automatic retry of an ambiguous call.
- Seeded draft instances, hidden research fault receipts and observable-event measures with missing-data definitions.
A local laboratory journal is not a production database adapter. Production AI dispatch remains closed until its durable budget, canonical release and qualification gates are verified.