The board is gone
Pathmode used to present evidence as a board: cards you grouped, clustered, bulk-selected, and turned into a spec in one move. That surface has been removed. Grouping made a pile look like a finding, and bulk selection let a hundred uninspected rows enter a decision as if someone had read them.
What replaced it is smaller and stricter.
The Evidence surface today
Evidence has two views inside a product:
- Inbox holds every signal nobody has reviewed yet. You work it one item at a time, and each item gets exactly one of three judgments: Use as backing, Keep unreviewed, or Dismiss. Evidence opens here when unreviewed items are waiting.
- Browse is a flat, searchable record of everything ever captured. Selecting a row opens a detail panel with its provenance, its raw content, its source, and read-only links to the exact claims it backs. Browse is where you audit; it is not where you judge.
Trust tiers
One classifier assigns every item exactly one tier:
- Backing: a signed-in person reviewed it and anchored it to an exact claim.
- Unreviewed: it was captured and nobody has judged it. Pasting, importing, and agent writes all land here.
- Assumption: an AI-drafted hypothesis. An assumption stays visible and can never become backing.
- Dismissed and archived items are excluded from the working set. They never leave the record.
Evidence types
Type describes what a signal is; tier describes whether anyone has judged it. The five types:
- Friction — a specific user pain point observed in behavior or reported in feedback ("users click the back button 3 times trying to find account settings")
- Quote — a direct statement from a user interview, support ticket, or review ("I gave up after 10 minutes and just emailed your team instead")
- Observation — a pattern noticed by the team through analytics, session recordings, or direct observation ("40% of users never complete onboarding step 3")
- Metric — a quantitative signal from analytics or monitoring ("checkout abandonment rate increased from 12% to 23% after the redesign")
- Request — a feature request or enhancement suggestion from users or stakeholders ("can we export reports as PDF?")
How evidence reaches a spec
Inside an IntentSpec, the Evidence section lists every claim in the spec: the objective, each outcome, each constraint, each edge case. Each claim reads Backed, Unreviewed, Assumption, or No evidence. Expanding one shows the exact items behind it with their tier, origin, source, and latest reviewer.
Use as backing is the only action that creates backing. It reviews the signal and anchors it to one exact claim in a single attributed operation. Items linked to an intent without an exact claim show as General / not claim-anchored and never make a claim Backed.
No evidence is descriptive, not a readiness failure. Some claims are decisions rather than findings, and the spec says so instead of pretending otherwise.
Who can do what
Agents can capture. Through the API or MCP an agent can create and query signals, and everything it writes arrives Unreviewed with an agent origin. Only a signed-in person can review a signal, attribute the judgment, and turn it into backing for a claim.