ADAMANTINE / EXPERIMENTAL AI ASSURANCE
Adamantine.
A locally integrated application for registering experiments, executing supported local model workloads, applying declared evaluators and reconstructing results.
Working local integration is documented. Independent release authorisation and commercial distribution remain separate decisions.
AUTHENTIC INTERFACE
The application, with actual records.

The interface’s Experiments workspace is the registry of experiment definitions. Execution attempts, retained artifacts, evaluator outcomes and verification records remain related but distinct objects. Independent approval sits outside that processing workflow.
HOW THE RECORD IS BUILT
A technical walkthrough.
STAGE 01
Register the question before interpreting the result.
Specify tasks, source declarations, input identities and evaluation rules. Retain the registered protocol state so later execution can be examined against it.
CONCEPTUAL WORKFLOW · NOT SOFTWARE OUTPUT
- 01Definition
Question, tasks and configuration
- 02Registration
Retained protocol and identity context
- 03Execution boundary
Changed or stale context stays visible
Experiment Registry showing a saved definition, declared evaluator and protocol state.
STAGE 02
Capture what ran and what returned.
Supported local model workloads produce attempts and response records. Completion, failure, timeout and cancellation are execution states; none is automatically a scientific outcome.
CONCEPTUAL WORKFLOW · NOT SOFTWARE OUTPUT
- 01Attempt
Supported local workload and configuration
- 02Response
Retained output or explicit absence
- 03Lifecycle state
Completed, failed or cancelled within scope
Execution workspace with a completed local attempt and its retained response.
STAGE 03
Keep the evaluator attached to the result.
Apply the declared scorer without collapsing malformed output or an unevaluable attempt into an incorrect answer. Retain evaluator identity, response and artifact context alongside the outcome.
CONCEPTUAL WORKFLOW · NOT SOFTWARE OUTPUT
- 01Declared evaluator
Rule identity and supported input
- 02Outcome
CORRECT, INCORRECT or a non-score state
- 03Evidence
Response and source-linked records
Evidence inspection showing an actual retained artifact and evaluator identity.
STAGE 04
Reopen the record and examine its limits.
Check retained local consistency and scorer replay, reconstruct the result from saved records and generate a report. These checks do not authenticate model origin or confer scientific validity.
CONCEPTUAL WORKFLOW · NOT SOFTWARE OUTPUT
- 01Retained record
Source identities and saved artifacts
- 02Reconstruction
Local consistency and supported scorer replay
- 03Report
Outcome, qualification and unresolved limits
Verification workspace showing an actual reconstruction result and report link.
STATUS SEMANTICS
Different statuses answer different questions.
CORRECT/INCORRECT- Outcome under the declared scorer and task contract; not a universal capability judgement.
- Malformed output
- The response does not satisfy the expected representation. It must not be silently reinterpreted as a valid scored answer.
NOT_EVALUABLE- A supported score cannot be established, for example when generation is cancelled without a response.
VERIFIED_LOCAL_CONSISTENCY- Retained local records and supported scorer replay agree within the verification scope. Model origin and scientific validity are not authenticated by that status.
- Scientific qualification
- Requires the experimental design, evaluator validity and evidence needed for the particular scientific claim.
- Independent authorisation
- Separately controlled authority and release decisions. They cannot be supplied by the application’s own passing checks.
INTERNAL ENGINEERING QUALIFICATION
Reported integrated engineering evidence.
Regression tests passed on the final integrated candidate. This is a software engineering result, not independent scientific certification.
Evidence cut-off: 17 October 2026 · Source revision 79cd69593b922d99ee993c11d75aa3ac8c453115 · Reported in the supplied test-and-logistics digest.
- Browser-driven local workflow
- The installed application completed experiment handling, one local-model generation, evidence retention, verification and report generation.
- Reopen after restart
- A new browser session retrieved retained experiment, task, attempt, evidence and report records. Fresh-process local verification replayed one response without mismatches.
- Bounded generation example
- The frozen task
31 + 11returned raw42and scoredCORRECT / EXACT_INTEGER_MATCH. A single arithmetic example does not establish broad model capability. - Model identity as recorded
Jackrong/DeepSeek-V4-Pro-Qwen3.5-4B, revision3453701a8b55cddf80f8bdc7fe30557e705cf80b. The record does not independently authenticate the model’s origin.- Qualification limits
- The retained protocol snapshot was
STALEafter execution followed registration. Local consistency and scorer replay do not establish scientific validity or independent release authority.
The digest reports local engineering checks, not an independent rerun performed for this website. Results from different revisions and workloads are not combined into one qualification.
DEPLOYMENT & ENQUIRIES
Local operation and release boundaries.
- Operating model
- Local application with a loopback service and a browser interface. Browser-local persistence is not the authoritative evidence backend.
- Native inference
- Requires the supported local runtime, dependencies and an eligible local model. Core-only operation must not imply inference availability.
- Records and output
- Experiment definitions, attempt states, retained artifacts, evaluator outcomes, verification and technical reports.
- Current commercial action
- Request technical information. Public distribution, independent release authorisation and licence terms are not presented as settled.