VibeOps · captured judgment · tldraw/tldraw · 2026-06-01

26 deterministic policies compiled from 203 captured judgments across 12 months of tldraw PRs.

Every policy is a rule extracted from the team's own past decisions - not a lint config, not a generic best practice. Each carries the owner class that would approve it, the historical compliance rate on PRs that triggered it, and the chain of judgments that compiled into it. This is the wall a senior engineer's brain looks like, made auditable.

Autonomy potential

72%

969 of 1,349 PRs trigger zero block-severity policies → auto-mergeable by captured rules.

Block-when-violated

Policies that route to owner approval when triggered. These encode irreversible judgments.

Advisory

Lower-severity policies. Flagged but not blocking - capture team norms with softer enforcement.

Distinct owner classes

Each policy is owned by the team that originally ratified the judgment. Package-derived where no CODEOWNERS file exists.

The wall

26 of 26 policies

How the wall was built

1Mined judgments.Ran the multi-agent miner across 12 months of tldraw PRs. Each agent (intent → history → convention → comment-pattern → reasoner) emits hypotheses. The verifier reads cited files at the PR's parent commit. Total surfaced judgments: 203.
2Bucketed by code area (package or top-level path). Within each bucket, an LLM clustered judgments sharing the same root concern and emitted policy specs (name + predicate + trigger keywords + supporting judgments). Lint-tier concerns (style, unused imports) are explicitly excluded.
3Computed coverage + compliancefrom train labels. For each policy, counted historical PRs touching the policy's files and computed the share labelled CLEAN at merge (no follow-up commit within 7 days). Revert history is checked separately.
4Assigned owner classesvia the repo's package structure. In your repo this hook is a CODEOWNERS lookup - tldraw has no CODEOWNERS file, so the demo uses a derived team-name proxy.
5Autonomy potential is the fraction of train PRs that trigger no block-severity policy at all. On tldraw, 72%. In your own repo this number replaces the gut-feel autonomy claim with a derived one bound to your team's captured rules.

Want this wall on your repo?

30-day pilot. We extract your team's captured judgments and hand you the wall on day 7.

We deploy inside your VPC, ingest your last 12 months of PRs + postmortems + Slack #incidents, and compile policies from your own decisions. CODEOWNERS becomes the owner class. The autonomy potential is whatever your captured rules say it is - no aspirational numbers.

Try with your org See the underlying judgments

26 deterministic policies compiled from 203 captured judgments across 12 months of tldraw PRs.

The wall

Validate required environment variables at startup before use

Feature flag changes must include paired adminRoutes update or explicit justification

Wrap all fetch/HTTP calls with explicit error handling and non-2xx response checks

Reverted PRs must document root cause and fix before re-landing

New CSS rules must be audited for RTL layout regressions

New CSS rules must not inadvertently create stacking contexts

Cache keys must include runner architecture segment

Consolidated workflow jobs must isolate runner-specific steps per matrix variant

Co-changed infrastructure template files must be updated together

Validate runtime theme-switch behavior before landing CSS theme changes

Require visual regression or snapshot test for theme/flash-related CSS changes

Publishing script changes must audit flag parity and CI auth config

Audit sibling ShapeUtil files when modifying any ShapeUtil

Regex changes in security-sensitive paths must include regression tests covering credential-URL patterns

Module-scope regexes with the global flag must not persist lastIndex across calls

New telemetry/RUM modules must include unit tests with fake timers and mock browser APIs

setInterval and addEventListener calls in module-level code must be cleaned up in a teardown path

Document/window access must be scoped to ownerDocument or container window, not bare globals

Non-obvious regex or logic decisions must include an inline explanatory comment

Attach rejection handlers to all async timeout helpers in tests

Clear polling intervals in afterEach/afterAll even on test timeout

Wrap all per-segment file operations in try/finally for guaranteed cleanup

Clear output/working directories at the start of each run

Use explicit stable sort key for segment ordering, never filesystem enumeration

Explicitly create all parent directories before writing output files

Ensure lockfile is updated when package.json dependencies change

How the wall was built

30-day pilot. We extract your team's captured judgments and hand you the wall on day 7.