Versalist Assurance
Run the skill your agent uses today and the revision you want to ship on the same cases and model. See which cases got worse, then record Ship, Hold or Needs work.
Plus models from OpenAI, Anthropic and Google.
The problem
An overall score can rise while individual cases get worse. A release review shows both versions case by case, so you see the drop before your users do.
How it works
Turn a real workflow into fixed test cases and a scoring rubric your team agrees on.
Run both versions on the same cases and model. Read the scores case by case and dimension by dimension.
The review owner records Ship, Hold or Needs work, and can export the review.
Features
Scores for both versions, the regressions between them, and every decision made on the review.
Scores for both versions, per case and per rubric dimension. A review holds up to 500 case rows and 100 dimensions per run.
A comparable dimension score that drops is marked as a regression.
Both runs must share the challenge, skill, rubric and case identifiers. A missing model identifier does not prove the models match.
Ship, Hold or Needs work, saved with the evidence and the full decision history.
Print a review or export it to Markdown for your release notes and sign-off.
Share a review with your organization, or create a public link for eligible reviews.
Built for engineers
Versalist checks the challenge, skill, rubric and case identifiers before creating a review. Model identifiers are compared when present. Missing model identifiers do not prove a model match.
| Field | Baseline | Candidate | Match |
|---|---|---|---|
| challenge_id | 3f9a… | 3f9a… | same |
| skill_bundle_id | 8c21… | 8c21… | same |
| challenge_content_hash | e4d0… | e4d0… | same |
| rubric_content_hash | b07c… | b07c… | same |
| cases_content_hash | 91fa… | 91fa… | same |
| model_contract_hash | 6e2b… | 6e2b… | same |
| agent_model_id | model-a | model-a | same |
skill_version2 to 3
the change under review (not part of the key)
| Field | Baseline | Candidate | Match |
|---|---|---|---|
| challenge_id | 3f9a… | 3f9a… | same |
| skill_bundle_id | 8c21… | 8c21… | same |
| challenge_content_hash | e4d0… | e4d0… | same |
| rubric_content_hash | b07c… | 5d93… | differs |
| cases_content_hash | 91fa… | 91fa… | same |
| model_contract_hash | 6e2b… | 6e2b… | same |
| agent_model_id | model-a | model-a | same |
Select two completed episodes with the same challenge, skill, rubric, cases, and model contract.
The platform
Use Assurance on its own, or with the rest of Versalist. Challenges fix the cases. Skills hold the revisions. Assurance compares them.
A private registry for your team's agent skills. Publish a version once and everyone pulls the same one. Roles and review policy decide who can publish, and every rollback is recorded with a reason.
revision · file digests · rollback reason
Challenge programs on your team's real workflows. Each challenge fixes the task, the cases, the rubric and the execution policy, which is what lets two runs be compared.
cases · rubric · execution policy
A free directory of tool contracts. See what a tool can do and what access it needs before your agent calls it, and verify contract exports in your browser.
Continuous learning
The model behind your agent does not learn from yesterday’s failure. Its skill can. Versalist turns a failure into a proposed amendment and tests it before anyone promotes it.
Import the execution receipts from a completed episode and read what failed.
An improvement cycle drafts an amendment: the skill version it starts from, the new content and the reason.
Baseline and candidate run on a random sample of acceptance cases. Cases that produced the proposal are excluded.
A person promotes the amendment when the comparison supports it. A model proposal alone cannot.
The skill changes, not the model weights. A proposal alone cannot authorize promotion: the comparison has to support it. The workflow runs on personal skills today.
Read the improvement workflowSecurity
Reviews start private. Owners control how they are shared.
Reviews start private. Owners can share with an organization or create a public link for eligible reviews.
Model-provider keys are stored per user in Google Secret Manager.
Give each integration only the access it needs, and revoke it at any time.
Trace capture stores prompts as hashes, never as text, and never records model reasoning.
The evidence behind a review can’t be edited, and decisions are added, never overwritten.
Versalist never operates your production agent.
Get started
Our team works with yours to scope one workflow and review the results.
Pick the workflow and agree on what good looks like.
Prepare your agent’s skill and model, and turn real cases into a test suite.
Score every candidate version against the one you run today.
Make the release call and share the results with leadership.
Need to make the case internally? Build your business case