Add production AML pilots and an ethical public-site cohort

[aruintelligence]

Summary

Runs the AML real-screen harness against five live production screens and completes one full detect → decide → ship → recapture → verify loop.

Owned-screen pilots:

1. https://aruintelligence.com/
2. https://chargergogoplacements.com/ before and after shipped remediation

Public open-source cohort:

3. https://react.dev/
4. https://vite.dev/
5. https://developer.mozilla.org/en-US/

Each package binds production viewport artifacts by SHA-256, uses separate labels with provenance, applies only render_allowed = restoration_value >= attention_cost, and records whether a deterministic result is appropriate to ship.

Owned-screen results

• ĀRU Intelligence: 6 ALLOW / 1 SUPPRESS; do not ship suppression on this evidence alone; receipt root 9e77955425864e91557f2adb9c9473d4461d56408aecce5bc3cafb3b32b83294.
• ChargerGoGo before: 6 ALLOW / 2 SUPPRESS; receipt root c22acf30c095b3ff5f1a8befa0c735f6c00c106d736dfa361469b56c3a5f7409.
• ChargerGoGo after shipped remediation: 7 ALLOW / 0 SUPPRESS; receipt root 74bd6a8d84c5dde4412b970aaa0d98ce011da58138f36a656d0c7c83915bf30a.

The ChargerGoGo remediation removed the redundant sales-hours strip and retained the domain-important bailment disclosure in a calmer Venue note.

Public open-source cohort

• React: 6 ALLOW / 0 SUPPRESS; receipt root 2d9e18781a4951ee07d06cb3367c3e807a3b9939b06fcbfaa78fd86a1669191b.
• Vite: 6 ALLOW / 1 SUPPRESS; receipt root caba52b8dde98101a292d4484737dac5b2af6512594e2c53658c38ab2696de2a.
• MDN: 6 ALLOW / 3 SUPPRESS; receipt root eaa2da67526039242eecaa526d8855704defaf6085b3eb686b61174015e986d0.

The cohort used one ordinary unauthenticated page load per site. No crawling, form submission, access-control testing, circumvention, or third-party outreach occurred. Exact screenshots and rendered artifacts were preserved and hashed outside Git; the repository contains hashes plus small semantic extractions to avoid redistributing site assets.

Evidence and claim boundary

All labels are project-directed operator reviews by OpenAI Codex. Numeric values are declared reviewer judgments, not objective measurements or independent validation. Every receipt correctly reports real_screen_evidence_eligible: false.

A SUPPRESS result on a third-party screen is only a hypothesis for maintainer review. It is not authorization to alter a site and may omit sponsorship, advertising, accessibility, legal, or product constraints. The maintainer notes propose questions and reversible experiments, not verdicts.

Verification

• all receipts reproduced byte-for-byte across two runs
• repository test suite: 494 passed, 0 failed
• Base44 production build passed before the ChargerGoGo publication