# Permadyn AI: sample hardware acceptance plan

Blank planning template. This is not a benchmark result, customer case study, or certification.
Reviewed: September 15, 2026.

## 1. Define the decision

- Application and representative tasks:
- Business owner and technical reviewer:
- Equipment budget and separate implementation/support budget:
- Proposed hardware and alternative to compare:
- Procurement decision date and supplier quote expiry:
- Acceptance thresholds agreed before testing:

## 2. Pin the complete configuration

- Hardware model, number of systems, memory per device, storage:
- Network topology, switch, cables, and measured link behavior:
- Power, cooling, physical location, and recovery access:
- Operating system, driver, inference runtime, and container versions:
- Model repository, exact revision, license, and quantization:
- Context length, cache precision, output length, and batching:
- Per-node model placement and memory reserve:
- Retrieval, embeddings, reranking, and application processes:

Record each device separately. Aggregate installed memory is not a single usable memory pool.

## 3. Evaluate quality

- Approved, non-sensitive evaluation set and version:
- Task-specific scoring rubric and expected outputs:
- Baseline system and minimum acceptable result:
- Reviewers, disagreement handling, and failure examples:
- Comparison of quantized and higher-precision behavior where relevant:
- Unsupported questions, source citations, and escalation behavior:

Report the tested sample size, scoring method, and limitations. Do not substitute vendor benchmarks for application results.

## 4. Measure application performance

For each concurrency and context setting, record repeated runs:

| Configuration | Context | Concurrent requests | Time to first token p50/p95 | End-to-end latency p50/p95 | Output tokens/sec | Peak memory per device | Error rate |
| --- | --- | --- | --- | --- | --- | --- | --- |
| To be tested | | | | | | | |

- Separate cold starts from warm runs.
- State whether prefix caching is enabled and whether prompts share a prefix.
- Report input-processing speed separately from output generation.
- Record test duration, request count, queue delays, and measurement method.
- For document/image/speech workloads, add the appropriate task throughput and quality metrics.
- Record wall-power measurements only when measured with an identified method.

## 5. Verify the complete system

- Authentication, authorization, and source-access restrictions:
- Permitted outbound connections and verified local-data boundary:
- Tool-action validation and human approval where required:
- Unavailable integrations, interrupted requests, and retry behavior:
- Model/runtime update and rollback:
- Node/link failure behavior for multi-node deployments:
- Backup restore and service recovery rehearsal:
- Monitoring, alert routing, incident ownership, and support coverage:

A cluster is not automatically highly available. Test the actual application behavior when a node or link fails.

## 6. Record the decision and handoff

- Measured results against each agreed threshold:
- Remaining issues and accepted limitations:
- Recommendation: buy / change configuration / retest / defer:
- Final bill of materials, supplier quote, and lead time:
- Source-code rights, accounts, and infrastructure ownership:
- Operating documentation and named support responsibilities:
- Reviewer approval and date:

Published by Permadyn AI.
https://permadyn.ai/services/on-premises-ai