The session behind the work

Copy this entire handoff into Claude Code:

# HANDOFF — SHIP WT-0090 END TO END

You are continuing an active, production-bound task in ~/miscsubjects-pages. Do not restart, redesign, or declare victory early. The target is a real, browser-visible, massive proof-of-work run that a cold ChatGPT, Grok, or Claude can independently verify from one article URL.

The owner wants one public article URL they can paste into a completely new ChatGPT, Grok, or Claude session. Without credentials, prior context, GitHub, or terminal access, that cold model must be able to traverse from the article into: the exact task-bound run; at least 1,000 evaluated organizations; every inclusion and exclusion decision; exact discovery queries, public sources, source quotes, reasons, and skill versions; at least 100 contacts verified from organizations' own public websites; 100 individualized email bodies tied to the selected evidence; 100 lawful provider-accepted sends, after exact owner review; public send receipts and proof IDs; raw or appropriately redacted payloads; hash commitments for redacted private values; four rival-model audits that were actually executed; a signed checkpoint; and public verification and challenge/comment doors.

[The full handoff also specified: the current task WT-0090 and its acceptance tests; why WT-0089 failed (WF-0006); the code already shipped; the live surfaces; the candidate->invocation binding bug to repair; the discovery, enrichment, drafting, owner-review, send, rival-audit, checkpoint, article, and cold-walkthrough requirements; and the rule to never declare a false finish. Verbatim instruction preserved; secrets and local paths stripped by the transcript sanitizer.]

Observe this feedback, also: I want hyperlinks to every supporting article or skill; tokens and the OIP tied in explicitly; the tracking (the justcloakit cloaker, to track model visits) detailed explicitly; models invited to ALWAYS mint tokens and verify work; everything revised so you track opens and clicks and responses and then do iterative version testing; the skill versions associated with the artifact scored, versioned, tracked; a ton of widgets in the articles; the article canonical, assuming zero context, with dense hyperlinks to all associated parts of the build; a new widget that takes the state-card turns and turns them into widgets the way the ledger has turns as widgets; generously source and showcase the evidentiary ledger; articulate exhaustively all competitors and the variances and thoroughly research to show where this is distinct; everything in place for much more scraping, scraping at large, of the AI field; a new arbitrage equation calculating the optimal number of people per day to email starting today (day 0, day 1, one more day); companion Twitter posts, where every outbound email always tags the relevant handle, as a new build rule; take the raw prompts involved (my words, these words, any errors, all tool calls) and show them; every party emailed inspectable on the ledger, do not redact who is being emailed, I am preparing to post the entire build publicly; and build a new feature whereby these work objects take the state-management cards and make them publicly inspectable while observing security practices, so red-team to see if you can break it.

Table Web's edits (the cold verifier): The three WT-0089 defects are fixed and visibly fixed. But do NOT approve the 103 yet — three NEW defects: (1) a fake 'verified' contact, GV at v...@rjt6iungs.smae, a garbage domain stamped verified; (2) the same firm gets contradictory verdicts (Betaworks, Glasswing, Boldstart, Work-Bench, BoxGroup, 2048; Khosla 5x, Lightspeed 4x, Felicis 4x) because passes never deduped; (3) the inclusion rule isn't applied consistently (Accel excluded for no official quote yet included on a listicle; Sequoia's evidence is a Forbes line, not its own thesis). Fix the verifier, dedupe by org, tighten the inclusion source rule, restage, then approve. The first 103 to inspect should find zero of what I found.