# Glossary — LLM quality and AI sprint terms we use

Short definitions as used on arviser.com. Not legal advice.

| Term | Meaning here |
| --- | --- |
| Narrow pipeline gate (24h PoC) | Verifying in principle within 24 senior engineering hours whether a working AI pipeline can solve the core or narrow bottleneck of a task across 20–50 cases and 2–3 models. See [24h PoC Spike](/ai-poc-sprint). |
| 50% Risk-Shared Track | Optional track on the 24h PoC (€2,400): if none of the 2–3 tested setups meet the agreed pass marks on the narrow pipeline gate (Kill), the second 50% payment is waived (€1,200 total). See [Risk-Shared PoC Protocol](/guides/risk-shared-poc-and-ip-protocol). |
| Sanitization protocol | Four rules for any research note published after a Risk-Shared Kill: zero client identifiers, 100% synthetic data, generic engineering framing, and a 48-hour client review window. |
| Smoke evaluation suite (40h Prototype) | A 15–25 case automated Promptfoo test suite shipped inside every [40-Hour Working AI Prototype Sprint](/ai-prototype-sprint) to verify happy paths, refusals, and schema compliance. |
| VC Technical Due Diligence Memo | A 2-page architecture, p95 latency, unit-economics, and evaluation baseline memo signed by Founder Vadim Nareyko for startup investor data rooms. Included on request in our [Startup Program](/startups). |
| Golden set | Labeled cases for the live surface. Breadth, label quality, freshness. See [golden set](/guides/golden-set). |
| Groundedness and faithfulness | Answers tied to allowed sources. Inventing vs refusing when evidence is missing. |
| Refusal and safety | Jailbreaks, policy breaches, unsafe tool use where agents act. |
| CI harness | Evals on change with pass or fail gates, not only on a laptop. Built in our 32h Eval & Guardrail CI Harness sprint (from €3,100) or 2-week Audit. |
| Production sampling | Live traffic feeding the suite, or a frozen demo set if live access is blocked. See [production sampling](/guides/production-sampling). |
| Go, Hold, No-go | Executive verdict after scores and failure modes. Hold means remediate before unsupervised scale. See [go, hold, or no-go](/guides/go-hold-no-go). |
| Under-refusal and over-refusal | Answered when it should decline, or declined when it should answer. Measure both. See [refusal both ways](/guides/refusal-both-ways). |
| Kill switch | Written stop or throttle path with owner, triggers, action, and logs. See [kill switch](/guides/kill-switch). |
| RAG evaluation | Gate retrieval and generation separately so the generator cannot hide weak recall. See [RAG evaluation](/guides/rag-evaluation). |
| Indirect prompt injection | Instructions that change model behavior via retrieved or fetched content, not only via the chat box. See [indirect prompt injection](/guides/indirect-prompt-injection). |
| Lethal trifecta | Untrusted input + sensitive access + outbound channel. See [lethal trifecta](/guides/lethal-trifecta). |
| Cross-tenant retrieval | Retrieval returns chunks another tenant or role must not see. See [cross-tenant retrieval](/guides/cross-tenant-retrieval). |
| AI Security Review | Scoped review of injection, data access, and agent permissions for one LLM app. See [AI Security](/ai-security). |
| Cost and latency gates | Budget per successful task and p95 latency next to quality metrics. See [cost and latency](/guides/cost-latency-gates). |
| Evidence Pack | Outline of inventory, golden set, evals, logs, handoff, kill switch, and decision. |
| Working baseline | Prioritized roadmap plus a working eval baseline for the agreed surface. Without both, you do not pay the Audit fee. |

**Next:** [How we score](/guides/how-we-score) · [AI Sprints](/ai-engineering) · [Evidence Pack](/evidence-pack) · [Request scope](/#scope)
