Identity & consent
You choose what Proof may see — extension capture, share links/exports, and optional API keys for drafting. Nothing ships without your later confirmation.
Proof methodology · connect & capture
Proof starts where the work already happened.
Outcomes via OAuth. AI work path via an optional browser extension, share links, or exports — with coverage gaps (mobile, desktop apps, blocked extensions) documented honestly, not papered over.
Capture the work path
Optional Chromium live capture on claude.ai / ChatGPT / Gemini / Copilot web, or paste share links and exports when the extension cannot see the chat. GitHub and HubSpot can also bring outcome receipts in.
Intelligence drafts — rules grade
Related sessions become draft Traces. Judgment moments, outcomes, and tool lineage stay attached. The LLM may propose; JUDGMENT_RULES set the legal grade.
Confirm, verify, share
You confirm or fix drafts in Inbox (facts move grades; tone never does). A witness can attest. Only then does a Value Card or portfolio link leave your private record — employers see the same grades, not a second algorithm.
What you can bring in
Live capture (browser extension)
Manifest V3 content scripts on explicit web hosts. Off by default; pause stops read and send. Mobile and desktop apps are not covered.
Manual exports & share links
Enterprise fallback when extensions are blocked: weekly ZIP/JSON/Markdown exports or Claude/ChatGPT share links.
Outcome systems + MCP (planned)
Gmail, Drive/Docs, GitHub, and HubSpot OAuth today — read-only, candidate signals only. MCP connectors for IDE/agent sessions are the path that does not depend on provider DOMs.
Private by default. Source chats and exports are not served from a public URL. Proof shows approved excerpts and receipts on a Value Card — never your raw model history. Live capture is off until you enable it; counsel ToS review per provider is still pending and is shown as such on Connections.
Intelligence Layer. Grades are computed by deterministic rules in /intelligence — same engine as Value Cards, mining, and admin extraction. LLM document parsing plugs in behind the same types.
Coverage is incomplete by design. The extension cannot see mobile or desktop-app chats, and many corporate machines block extensions. Fallbacks are share links, export cadence, outcome OAuth, and planned MCP connectors — not a pretend 100% capture story.
After the history is in
Every outcome still separates what happened from why this work gets credit.
Connecting an LLM shows the work path. Grading still requires Work → Artifact → Outcome → Confirmation. An outcome can be real while the contribution link remains uncertain, so the card shows the evidence and the limit.
How to read the evidence
Claimed
A structured professional assertion with its limits visible.
Supported
The correct minimum evidence for this value form is present.
Established
Full evidence threshold for this value form is met.
Contested
A tagged party disputes the link, number, role, or share.
Grading model
Evidence—not writing quality—sets the grade.
01Draft
Proof proposes a claim and identifies the evidence it can see.
02Rules grade
Value-specific evidence thresholds set Claimed, Supported, or Established.
03You confirm
Fact changes re-grade the outcome. Tone changes never do.
Six forms of attribution evidence
01Temporal precedence
The work existed before the outcome moved.
02Artifact lineage
The produced artifact is the thing that shipped.
03Counterfactual baseline
The before-state and measured delta are documented.
04Contribution attestation
The people involved confirm roles and shares totaling 100%.
05Independent corroboration
An outside system confirms the outcome near the work.
06Adoption telemetry
A durable asset shows sustained use over time.
07Reinvestment documented
Witnesses do not create causation by themselves. A live response updates the badge immediately. The grade changes only when the correct evidence threshold for that value form is met.
Built here: Value Cards, evidence depth, tokenized witness response, session mode, C3 judgment-motion detection, and P2 directed Value Mining with learned toss suppression.
Next up: deeper connector verification, full provenance across tools, and normalized production outcomes storage.