Experiment Lab
VeriLoop does not blindly optimize prompts. It evaluates targeted candidate interventions against 60/20/20 splits.
Targeted intervention modalities
RETRIEVAL
Change source priority, recency decay & vector index weighting
PROMPT
Modify instructions & meta-prompting constraints with XML boundaries
TOOL
Fix tool schema prerequisites, routing rules and disambiguation
GUARDRAIL
Add input/output safety validation and prompt injection wrappers
MEMORY
Fix multi-turn entity retention and explicit state amendment rules
WORKFLOW
Enforce conditional policy check gates before sensitive actions
Loading Candidate Experiments...
Fetching 60/20/20 generalization splits and regression verifications
Generalization Split Breakdown
60% OPT (60) · 20% VAL (20) · 20% HOLDOUT (20)| Split | Baseline v1.0 | Candidate A | Candidate B (Promoted candidate) |
|---|---|---|---|
| Optimization 60 cases (60%) | 95% (57/60) | 98% (59/60) | 98% (59/60) |
| Validation 20 cases (20%) | 95% (19/20) | 85% (17/20) | 100% (20/20) |
| Holdout 20 cases (Strictly isolated from tuning) | 95% (19/20) | 80% (16/20) [BLOCKED] | 95% (19/20) [VERIFIED] |
VeriLoop proves that Candidate B does not simply overfit to optimization probes. It generalizes cleanly to unseen holdout cases with zero regressions.
Generalization check
Did the candidate fix actually generalize?
Optimization · PASS (98%)Validation · PASS (100%)Holdout · PASS (95%)Regression · 0 (CLEARED)FINAL VERDICT: VERIFIED