Set the experiment
Define success before the first iteration.
Choose the parameters, metrics and evaluations. Approve the experiment's budget, iteration limit, promotion criteria and saturation threshold before it runs.
The local-first workbench
Build AI you can stand behind.
Bring intelligent automations into real work without assembling a governance system around every workflow. PathLAB Studio connects authoring, enforced runtime controls, durable Pathproof records and controlled experiments in one local-first workbench.
Explore the free experience01 / COMPILE
Compiled runtime
Defined before execution
02 / RECORD
Pathproof
Recorded throughout execution
Cryptographically linked evidence
03 / EXPERIMENT
One workbench connects the whole process.CurrentCandidate
Compare. Retain or reject. Record why.
Explore the free experience
See what you can control, inspect and improve.
The Free plan includes the open-source runtime, Pathproof and limited working PathLAB access. Start here with three website demonstrations: inspect a workflow, test a receipt and compare candidate versions.
These are website demonstrations, not a workbench download or a live model run.
01 / Write
Define the work. Compile the controls.
Set the actions, legal routes, tools, checks and budgets in a Flight Plan. Pathsum binds them into one sealed execution contract. Your boundaries become rules the runtime enforces, not instructions you hope a model follows.
HVAC quote / Flight Plan
WORKFLOW Prepare a customer quote
TOOLS Read catalogue · Draft quote · Send quote
CHECKS Pricing · Required details
GATE Human approval before send
BUDGETS Tokens · Cost · Wall time
The permitted pathCompiled before the run
02 / Run
Keep each action within its declared authority.
The runtime enforces legal routes, input and output validation, business rules, permissions, budgets and retry limits. Sealed Tools restrict each step to its declared operations. Human Gates pause designated actions for approval. The model works within those controls; it cannot grant itself more authority.
Declared tool access for the quoting workflow
- Parts catalogue
- Read
- Quote authoring
- Draft
- Send quote
- Human approval
Checks at the boundaryEvents enter Pathproof
03 / Review
Keep the record. Verify what happened.
Pathproof records execution events as they happen in a durable, tamper-evident ledger. Follow tool calls, retries, validation results and human decisions back through the run. Verify the cryptographic receipts independently, rather than relying on the automation’s own account of its work.
Recorded execution of the quoting workflow
- Workflow
- HVAC quote
- Recorded events
- 19
- Recorded outcome
- Passed
- Run identity
- 01M0J98NQD3P6PX45H0S3D95CN
04 / Improve
Change one thing. Find out whether it helps.
Try a different prompt, model or retrieval setting. Change one variable, compare it with the retained version on the same cases and criteria, and inspect the result. Keep the change when it meets your conditions; keep the evidence either way.
- What can change
- Choose one variable, such as a prompt, model choice or retrieval setting, and define the alternatives the experiment may try.
- What counts
- Choose what matters and how to judge it: a measured result, a rule, a human review or a model-based assessment. Define the rubric and what earns a better result. Keep the test cases and scoring rules fixed across candidates.
- When to stop
- Set a spend limit, a maximum number of attempts or a limit on consecutive attempts without improvement.
- What stays with you
- Candidates, results and retain-or-reject decisions stay linked to their run records, including unsuccessful attempts. Select examples for evaluation or training datasets.
Configured experimentation
Know which changes earned their place.
Compare a candidate with the current version using the same reference inputs. Set the metrics, evaluations and promotion criteria before the experiment. A separate verifier assesses the results, and the record preserves the decision to retain or reject the change.
Selected trial
run_042
↑ +7 · Promoted
- Candidate
- 91
- Incumbent
- 84
- Retained
- 91
- Difference
- +7
One variable changedprompt.versionv7 → v8
The candidate scored 91, above the incumbent's 84. This comparison retains the candidate.
Illustrative Pathproof lineagerun_042 derived_from run_041 · 87bd…4ee1
- 0 gain
- Saturation stop
- 3 no-gain runs
- Patience stop
Keep the comparison
Review the decision, including the rejected changes.
Inspect the candidate alongside the retained version. Rejected and no-change results remain connected to the run records that explain the decision.
Use the record
Turn recorded work into evaluation data.
Export inputs, outputs and labelled comparisons for evaluation and machine-learning datasets, with each example linked to its source run and experiment.
Beyond a single run
Start with one automation. See where it can take you.
Your first automation can become the starting point for something more specialised. Review its work, turn useful examples into training data, and explore fine-tuning or LoRA adapters for your own tasks.
Could a different prompt, retrieval strategy or trained model do better? Follow a code-search example from reviewed runs to a candidate you can compare. PathLAB keeps the evidence and comparisons connected; you choose which direction to pursue. Use each comparison to shape your next experiment.
Explore learning from your automations
Local-first by design
Keep the work in your environment.
Write workflows, inspect runs and retain their records in a local-first studio. Choose the models and integrations that suit your work while keeping controls, evidence and experiments together.
Choose the plan that fits your work.
Start with the open-source runtime, Pathproof and limited working PathLAB access in Free. Choose PathLAB Studio for the full workbench, or discuss enterprise licensing for your organisation.
Compare plansAsk about a PathLAB licence ↗