Rubric-Scoped Evidence
Screenshots and DOM state are grouped against specific checks so subtle failures can be caught without noisy context.
WootzApp turns sensitive enterprise data and workflows into privacy-preserving W8-RL environments. Our browser captures the evidence a verifier needs. W8 controls what leaves the enterprise boundary.
A forked renderer is a better verifier: it captures the browser evidence needed to turn private enterprise data into trainable, privacy-reviewed W8-RL gyms.
Identify enterprise data boundaries, sensitive surfaces, and what evidence can safely become part of an RL environment.
Capture browser-visible state, screenshots, DOM evidence, and actions without handing over uncontrolled raw systems.
Constrain, redact, or keep sensitive fields inside controlled deployments while preserving training utility.
Score process and outcome separately using browser evidence, rubrics, and trajectory replay.
researcher@lab:~$ # Browser evidence pipeline
$ w8 snapshot --workflow claims-admin --mode private
✓ Boundary map ..................... LOADED
✓ Sensitive fields ................. CONSTRAINED
✓ Browser evidence ................. CAPTURED
$ w8 verify --trajectory run_0421
process_reward: 0.86
outcome_success: true
environment_blocker: false
$ cat privacy-ledger.json
"retained_fields": "approved"
"exported_artifacts": "reviewed"
The browser is the verifier. WootzApp records screenshots, DOM state, actions, and outcomes so W8-RL can score both execution quality and user-visible success without uncontrolled raw data movement.
Evidence
Screenshots + DOM
Privacy
Reviewed artifacts
Process
Rubric rewards
Outcome
Goal success
W8 assesses, sanitizes, documents, and packages sensitive systems so useful signal can move into model-training pipelines as defensible RL assets.
Reliable RL starts with reliable verifiers. Browser evidence shows what actually happened: what was visible, what changed, what was submitted, and whether the user outcome was satisfied.
W8-RL uses the renderer fork to preserve verifier evidence while documenting what sensitive fields were removed, constrained, or kept inside the deployment.
Process rewards, outcome labels, failure separation, and privacy decisions travel with every environment.
Screenshots and DOM state are grouped against specific checks so subtle failures can be caught without noisy context.
Execution quality is scored separately: incomplete subtasks, hallucinated actions, and side effects are not hidden by a lucky final state.
The final user-visible goal is checked independently from the agent's intermediate process.
Agent reasoning errors are separated from environment blockers like login state, missing inventory, or unavailable test data.
Retained, redacted, and constrained artifacts are documented alongside the environment.
Training loops, inference servers, and orchestration are becoming commoditized. What still determines RL quality is the private data asset: realistic systems, privacy-reviewed evidence, replayable trajectories, and reliable rewards.
W8-RL output
Every environment ships with documentation of what was kept, what was removed, and why the asset is usable downstream.
Our forked browser renderer captures the verifier evidence RL needs: screenshots, DOM state, actions, outcomes, and replayable trajectories. W8 controls what leaves the enterprise boundary.
Real workflows, business logic, edge cases, and operational context do not show up in public corpora.
Data owners need to see what is retained, redacted, constrained, and documented before an asset can move downstream.
W8 packages approved evidence into task specs, reward traces, rollouts, and environments where process quality and outcome success are scored separately.
W8-RL expands the portion of private enterprise systems that can safely move into RL pipelines. Codebases and workflows that were too sensitive to license become documented, replayable assets with verifier evidence attached.
A claims, ERP, or card-operations workflow becomes a W8-RL package: approved screens, sanitized browser evidence, rubric-scored process checks, and outcome labels tied to the final browser state.
Fields, documents, and records are tagged before they become environment evidence.
Process checks stay specific and non-overlapping so failures do not cascade.
The final browser state proves whether the user's request was actually completed.
Retained snapshots, redactions, and delivery constraints are attached to the environment.
Bring us a private browser workflow. We will help map sensitive boundaries, capture the evidence, build process and outcome verification, and package the environment for model training.
WootzApp W8 - The privacy pipeline for browser-based RL.