# How the 50% risk-shared PoC and sanitization protocol work

A 24-hour Technical PoC Research Spike exists to verify whether a working AI pipeline can be built for the core or narrow bottleneck of your product idea before you commit budget to a full build. To align engineering incentives and remove downside risk, Arviser offers two transparent tracks at kickoff.

## Two tracks for the 24-hour PoC Spike (from €2,400)

| Dimension | Risk-Shared Open-Research Track | Standard Closed-NDA Track |
| --- | --- | --- |
| Kickoff deposit | €1,200 (50%) | €1,200 (50%) |
| Balance when narrow pipeline is feasible (Go) | €1,200 on delivery (total €2,400) | €1,200 on delivery (total €2,400) |
| Balance when pipeline is not feasible (Kill) | **€0, second 50% waived (total €1,200)** | €1,200 on delivery (total €2,400) |
| Code and IP when feasible (Go) | 100% exclusive IP transfer under NDA | 100% exclusive IP transfer under NDA |
| Code and IP when not feasible (Kill) | Non-exclusive license for client; Arviser retains reusable harness rights | 100% exclusive IP transfer under NDA |
| Fee credit on Go (within 60 days) | 100% (€2,400) toward 80h MVP or Audit; 50% (€1,200) toward 40h Prototype | 100% (€2,400) toward 80h MVP or Audit; 50% (€1,200) toward 40h Prototype |

## Objective feasibility criteria in Hours 1 to 4

Before any code is written, a senior engineer (10+ years of production experience) and the founder lock the written PoC scope with your team:

1. **Target slice:** We isolate the core technical bottleneck or a narrow, high-value sub-task of the pipeline (for example, structured extraction from noisy PDFs, hybrid retrieval groundedness, on-device PII stripping, or sub-1.5s agent tool routing).
2. **Pass marks:** We agree on explicit numeric thresholds across 20–50 representative test cases: task accuracy, groundedness, refusal behaviour, p95 latency, and unit cost per 1,000 runs.
3. **Multi-model evidence bar:** Every 24-hour PoC tests at least 2–3 distinct model or pipeline configurations inside an automated Promptfoo or DeepEval harness.

If any tested configuration achieves the agreed baseline thresholds and confirms that a working pipeline is feasible in principle, even on a focused slice of the problem, the PoC is a confirmed **Go**. Upon payment of the remaining 50%, your company receives 100% exclusive IP ownership and full credit eligibility toward the next build stage.

## What happens on a Kill verdict under the Risk-Shared Track

If none of the 2–3 tested architectures can meet the agreed feasibility thresholds on the 20–50 case benchmark, the PoC delivers a **Kill** verdict.

- **Second payment waived:** You pay only the initial €1,200 deposit (50% off the €2,400 package price).  
- **Full engineering deliverable:** You still receive the complete repository, the Promptfoo evaluation logs across all tested models, the root-cause engineering memo, and a non-exclusive perpetual license to use the code internally.  

Under the Risk-Shared Track, Arviser retains rights to the generic evaluation harness and may publish an anonymized technical benchmark study under four sanitization rules:

1. **Zero client identifiers:** No company name, brand, industry clue, or employee names ever appear.  
2. **Zero proprietary data:** All published benchmarks and examples use 100% synthetic datasets generated separately from your materials.  
3. **Generic task framing:** Domain-specific business rules are abstracted into general software engineering problems, such as multi-table document parsing or latency-bounded tool calling.  
4. **48-hour pre-publication check:** Your team receives a 48-hour window to verify that the sanitized technical note contains no confidential information.  

**Next:** [24h Technical PoC Spike](/ai-poc-sprint) · [Startup & Scale-up Program](/startups) · [What an AI PoC must prove](/guides/what-an-ai-poc-must-prove) · [Request scope](/#scope)
