VIRIDIAN WORKBENCH

Viridian
Workbench.
Local-first semiconductor software.

Inspect physical designs, investigate timing and trace engineering findings to their source. A local-first workstation built around established EDA tooling.

Proposed licence · Distribution qualification and written terms pending.

Physical-design inspectionTiming diagnosticsTraceable evidenceIndependent AI assurance

01 / SOFTWARE

Viridian Workbench.

A local engineering workspace connecting design inspection, analysis and reproducible evidence.

SEMICONDUCTOR ENGINEERING

From physical layout
to a traceable finding.

DEF/LEF placement and routing visualisation, OpenSTA-backed diagnostics, reproducible OpenROAD execution and evidence-linked technical reports.

InspectCells, pins, nets and available physical geometry.
InvestigateTiming context, design weaknesses and engineering evidence.
RecordFindings, engine evaluations and their supporting artifacts.
Product scope and qualification

02 / AI ASSURANCE

AI evidence reviews.

Independent examination of the evidence behind AI evaluation, model-comparison and release decisions. Three services, with an explicit scope.

£750

Experimental Audit

Does the evidence support the claim?

One principal claim. One evidence package. One written assurance finding.

YOU PROVIDE
One principal claim and one evidence package.
WE EXAMINE
Comparability, provenance, evaluator context and the claim-to-evidence relationship.
YOU RECEIVE
One written assurance finding with supported conclusions and material limits.
Request an Experimental Audit
£150

Diagnostic Review

What is compromising the experiment?

One defined technical issue. An agreed review boundary.

YOU PROVIDE
One defined technical issue and the artifacts available.
WE EXAMINE
The failure boundary: inconsistent runs, evaluator drift, provenance gaps or broken comparison context.
YOU RECEIVE
A focused diagnosis and targeted next checks; remediation is separately scoped.
Request a Diagnostic Review
£49

Benchmark Verification

What does the reported result establish?

One benchmark result or model-comparison claim.

YOU PROVIDE
A single reported benchmark result or model-comparison claim and its evidence.
WE EXAMINE
Comparability, evaluator consistency, configuration identity and obvious confounding.
YOU RECEIVE
A concise determination and the maximum defensible claim.
Request Benchmark Verification

Benchmark Verification excludes remediation, experimental redesign and extended root-cause analysis. Those belong in a Diagnostic Review.

  1. 01

    Describe the case

    State the question, the decision it informs and the artifacts available.

  2. 02

    Agree the scope

    Confirm the evidence boundary, handling requirements, fee and delivery timing.

  3. 03

    Receive the finding

    Written conclusions and next checks, limited to the material reviewed.

What evidence should I prepare?

Scores and per-example results, run logs, model and dataset identities, experiment configuration, evaluator configuration and the rubric where relevant. Identify missing artifacts at intake; they may limit the conclusion.

How should I handle private evidence?

Describe the case in the intake form and indicate that artifacts are private. Keep confidential material and credentials out of the form. Agree evidence handling and any local-first requirements by email before sharing files.

03 / AI SOFTWARE

Assurance infrastructure.

Software for retaining experimental context and examining the evidence behind model and evaluation decisions.

AI ASSURANCE

Osmium

Experimental assurance across repeated AI development cycles.

Measurement identity, provenance and reproducibility connect reported results to the experiments that produced them. Regression and release conclusions remain limited to the scope justified by that record.

Deployment requirements and the applicable assurance scope are established for the engagement.

AI ASSURANCE

Adamantine

Local experiment records, controlled execution and evidence reconstruction.

Connect experiment, attempt, evidence, verification and report records. Keep execution state, local consistency and independent authority distinct; a successful verification does not supply scientific validity or release approval.

Qualification remains specific to the assessed candidate, runtime and gate policy.

04 / RESEARCH

Technical research.

Published technical records, with the experimental setup and the boundary of each conclusion kept visible.

AI evaluation, assurance & progress

Open the research library
001

Response-order sensitivity in LLM judging

Technical manuscript · Not peer reviewed

Reversal preserved the preferred response in 19 of 20 comparisons. One pair showed a second-position pattern.

Specific to this frozen response set and judge configuration. The evidence does not establish systematic position bias or general robustness.

002

Evaluator Contract Stability

Exploratory execution · Not peer reviewed

No recorded winner disagreements across the tested rubric formulations on 20 reused frozen pairs.

All conditions shared one judge conversation; specified fresh-context isolation was not followed. Shared state may suppress disagreement.

003

Presentation / Serialization Sensitivity

Exploratory execution · Not peer reviewed

The exact tested boundary wrapper preserved recorded winners on all 20 reused frozen pairs.

All conditions shared one judge conversation. The result does not establish general presentation robustness or an independently estimated wrapper effect.

K2

· TRIAL 001

When Fine-Tuning Made a 5B Model Worse

Revised evidence review · Publication draft

The reported 300-case exact-pass score fell from 238 to 106 after one QLoRA intervention.

A bounded, single-seed result. The source’s audit bundle, case and manifest identities remain unresolved; this is not independent audit qualification.

RETAINED RESEARCH

Semiconductor physical design & EDA

SILEX’s OpenROAD GCD / Nangate45 investigation examines sampled path structure and physical constraints. The shared segment identifies a region for investigation, not a proven causal bottleneck or optimisation gain.

Explore SILEX research

05 / APPROACH

Keep the claim
within the evidence.

Benchmark scores, passing tests and successful tool runs answer different questions. Our work keeps those distinctions explicit.

Defined scope

State the question, assumptions and evidence required before interpreting a result.

Traceable context

Retain artifact, configuration and evaluator identities alongside measured outcomes.

Qualified conclusions

Refuse or narrow unsupported claims. Missing evidence remains a limitation.

06 / PARTNERSHIPS

Specialist assurance
for consultancy teams.

For AI engineering, governance, MLOps and risk consultancies whose client decisions depend on experimental evidence. Bring Viridian into a defined review through a specialist subcontract or client referral.

Keep the client relationship. Agree responsibilities, access and scope before work starts. No upfront partnership fee; commercial terms apply when paid client work is generated.

Discuss a partnership

CONTACT

Technical enquiries.

Tell us about the engineering question, the decision it informs and the evidence available.

derrymitchell@viridianassurance.com

All enquiries and engagements are asynchronous.

EXPERIMENTAL AUDIT · £750

Submit an audit enquiry.

One principal claim, one evidence package and one written assurance finding. Describe the claim you want examined; we will confirm scope and evidence handling by email.

State the result or decision, and briefly describe the evidence available.

A public report or repository, if available. Private evidence can be arranged after scoping.

PRODUCT PARTNERSHIP

Discuss purchase or integration.

For customers considering a Viridian product licence or integration into an existing system. Tell us what you need and which product you are interested in.

Describe your use case, existing system and any technical requirements.

Fields are required unless marked optional. Viridian uses your details to assess and respond to your enquiry; Netlify processes form submissions. Please do not include confidential evidence, credentials or personal datasets. All enquiries are asynchronous. You can also email derrymitchell@viridianassurance.com.