René's URL Explorer Experiment

{
  "@context": "https://schema.org",
  "@graph": [
    {
      "@type": "Organization",
      "@id": "https://proofpane.com/#org",
      "name": "Proofpane",
      "url": "https://proofpane.com",
      "logo": "https://proofpane.com/og-image.png",
      "email": "Louie.Lu@proofpane.com",
      "description": "Compliance-by-construction governance layer for AI in regulated teams.",
      "address": { "@type": "PostalAddress", "addressCountry": "NZ" }
    },
    {
      "@type": "Person",
      "@id": "https://orcid.org/0009-0002-7811-9354",
      "name": "Louie Lu",
      "identifier": "https://orcid.org/0009-0002-7811-9354",
      "sameAs": "https://orcid.org/0009-0002-7811-9354",
      "affiliation": { "@id": "https://proofpane.com/#org" }
    },
    {
      "@type": "ScholarlyArticle",
      "@id": "https://doi.org/10.5281/zenodo.21402331",
      "name": "The Proofpane Architecture for AI Governance",
      "author": { "@id": "https://orcid.org/0009-0002-7811-9354" },
      "identifier": "https://doi.org/10.5281/zenodo.21402331",
      "sameAs": "https://doi.org/10.5281/zenodo.21402331",
      "datePublished": "2026-07-17",
      "license": "https://creativecommons.org/licenses/by/4.0/",
      "publisher": { "@type": "Organization", "name": "Zenodo" },
      "abstract": "A public reference architecture for AI governance systems that generate verifiable evidence from enforcement: six invariants, seven planes, four coverage classes, and an L1-L3 conformance ladder."
    },
    {
      "@type": "ScholarlyArticle",
      "@id": "https://doi.org/10.5281/zenodo.21728844",
      "name": "The Chain of Custody: A Decision-Level Model for AI Governance",
      "author": { "@id": "https://orcid.org/0009-0002-7811-9354" },
      "identifier": "https://doi.org/10.5281/zenodo.21728844",
      "sameAs": "https://doi.org/10.5281/zenodo.21728844",
      "datePublished": "2026-08-01",
      "license": "https://creativecommons.org/licenses/by/4.0/",
      "publisher": { "@type": "Organization", "name": "Zenodo" },
      "isBasedOn": { "@id": "https://doi.org/10.5281/zenodo.21402331" },
      "abstract": "A conceptual model, not a validated theory. Decomposes a governed decision into seven states and asks which hand-off each control attaches to; argues that placement bounds what a control can establish, and that display integrity and recorded comprehension evidence are consecutive, non-substitutable control positions."
    },
    {
      "@type": "ScholarlyArticle",
      "@id": "https://doi.org/10.5281/zenodo.21844893",
      "name": "Load-Bearing Coverage: Mechanically Checking Whether a Conformance Suite Can Be Passed Without Doing the Work",
      "author": { "@id": "https://orcid.org/0009-0002-7811-9354" },
      "identifier": "https://doi.org/10.5281/zenodo.21844893",
      "sameAs": "https://doi.org/10.5281/zenodo.21844893",
      "datePublished": "2026-08-08",
      "license": "https://creativecommons.org/licenses/by/4.0/",
      "publisher": { "@type": "Organization", "name": "Zenodo" },
      "abstract": "A mechanical adequacy criterion for conformance suites. A rule is load-bearing with respect to a vector set if deleting it from the reference verifier changes the outcome of at least one vector; ordinary coverage, a vector existing that names the rule, is strictly weaker, so a suite can certify implementations that skip the rule entirely. Reports a working implementation and nine measurements across seven corpora in Python, JavaScript and Java. Preprint, not peer reviewed."
    },
    {
      "@type": "SoftwareApplication",
      "name": "Proofpane",
      "citation": [
        { "@id": "https://doi.org/10.5281/zenodo.21402331" },
        { "@id": "https://doi.org/10.5281/zenodo.21728844" },
        { "@id": "https://doi.org/10.5281/zenodo.21844893" }
      ],
      "url": "https://proofpane.com",
      "applicationCategory": "BusinessApplication",
      "operatingSystem": "macOS, Windows, Linux, Web",
      "publisher": { "@id": "https://proofpane.com/#org" },
      "description": "Evidence layer for AI in regulated teams: policy gate on every AI call, tamper-evident hash-chained audit log, production defaults changed on recorded comparisons (a skill's model when a named person accepts one or an auto policy applies one classed improved, the agent's memory method only on a significant result, with an opt-in inter-rater-reliability measurability floor), and Ed25519-signed Evidence Packs verifiable offline. Mapped to NIST AI RMF, ISO/IEC 42001, EU AI Act, GDPR and SOC 2. Design method: the platform is designed backwards from a stated endpoint — the enterprise-grade strategic agent an organisation can safely trust with real operating work; that endpoint's required components (governed skills, versioned SOPs, per-step evidence, human authorization, measurable quality) are, item by item, the design spec of the governance layer. The breadth is that requirement set, not scope drift — built by an engineer who did agent and workflow engineering first, then built the governance layer that work demanded.",
      "offers": { "@type": "Offer", "price": "0", "priceCurrency": "USD", "description": "Free demo org — no signup required" }
    }
  ]
}

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [
    {
      "@type": "Question",
      "name": "Show me everything that happened, signed — can the record be trusted?",
      "acceptedAnswer": { "@type": "Answer", "text": "Every AI decision lands in a SHA-256 hash-chained, append-only audit log (DB-level immutability trigger) with Ed25519 chain-head anchoring. It exports as a signed Evidence Pack that an auditor verifies OFFLINE with a standalone CLI — no Proofpane account or backend access required." }
    },
    {
      "@type": "Question",
      "name": "Map this to my framework controls.",
      "acceptedAnswer": { "@type": "Answer", "text": "Proofpane ships 335 pre-mapped controls across NIST AI RMF, ISO/IEC 42001, EU AI Act, GDPR and SOC 2, with per-organisation overrides. Evidence Packs include the control coverage for the export window." }
    },
    {
      "@type": "Question",
      "name": "Why is this your production default?",
      "acceptedAnswer": { "@type": "Answer", "text": "In Proofpane, a skill's model default changes when a named person accepts a comparison recorded on the organisation's own cases, or when an auto policy applies a candidate the comparison classed improved; the agent's memory method changes only on a statistically significant result. An auto policy can sit behind an opt-in inter-rater-reliability measurability floor (Krippendorff's alpha), below which the change waits for a person. The frozen verdict, confidence interval and fixture hashes are kept on the audit chain." }
    },
    {
      "@type": "Question",
      "name": "How do you stop costs running away — and catch quality drops?",
      "acceptedAnswer": { "@type": "Answer", "text": "An organisation's monthly budget is a hard cap: once it is spent, Proofpane refuses that organisation's workflow steps, agent turns and skill runs, streamed or not. An agent's own monthly cap refuses that agent's turns and the skills it runs, and a user's budget set to block refuses that user's agent turns once it is spent. Each refusal comes before the model is called and is recorded on the audit chain. Calls proxied through the egress gateway are metered but not capped. A department's budget set to block refuses the agent chat route's turns (POST /api/v1/agent/chat) once it is spent, and that route's rows carry the requesting user's department; on every other route, whether a department's budget refuses a call or its row carries the department is not yet measured. Calls are metered where they run, one meter for the calls Proofpane makes and the egress gateway for the calls it proxies, with per-model totals to check against the vendor's invoice. Quality is watched by golden-dataset evals, live quality sampling, drift alerts and behavioural-anomaly detection that can pause an agent mid-run." }
    },
    {
      "@type": "Question",
      "name": "How do you improve safely?",
      "acceptedAnswer": { "@type": "Answer", "text": "In Proofpane, every proposed improvement is sandbox-tested against real failing samples and requires explicit human approval before it ships; every version carries author, approver and rollback. Five things ride this loop: RAG retrieval method, memory method, skill prompts, the agent's own seed harness (the system prompt it boots from, composed from approved fragments with a frozen capsule per version), and new techniques scouted from expert sources — auto-drafted in an isolated sandbox, reviewed by a second-model critic, human-approved, shipped as an ordinary pull request. Nothing self-modifies silently." }
    },
    {
      "@type": "Question",
      "name": "Show me the process, not just the outputs.",
      "acceptedAnswer": { "@type": "Answer", "text": "Proofpane renders every run as a live tree — each tool call, policy decision, human approval, token count and dollar — with parent-child lineage across multi-agent runs, shareable read-only with an auditor." }
    },
    {
      "@type": "Question",
      "name": "How does Proofpane govern Cursor, Claude Code, Codex and other coding agents?",
      "acceptedAnswer": { "@type": "Answer", "text": "Proofpane governs them through each vendor's OFFICIAL extension surface — hooks APIs and MCP — never TLS interception. Depth is published per client: gate + transform where a plugin can wrap tools (e.g. Hermes), gate-only where hooks allow deny (Cursor, Claude Code), observe where only transcripts exist (Codex). Secrets are DLP-redacted before a model sees them; risky actions pause for a menu-bar approval." }
    },
    {
      "@type": "Question",
      "name": "Does my team have to change tools?",
      "acceptedAnswer": { "@type": "Answer", "text": "No. Teams keep Claude Code, Cursor, Codex, n8n, UiPath, Power Automate and their existing workflows. Proofpane attaches via official hooks, MCP and an egress gateway — no wrapper IDE, no forked clients." }
    },
    {
      "@type": "Question",
      "name": "Do you ship a desktop app, and is it signed?",
      "acceptedAnswer": { "@type": "Answer", "text": "Yes \u2014 Proofpane builds native desktop apps for macOS and Windows on CI. The current distributed macOS daemon, v1.5.26, is signed with the PROOFPANE LIMITED Developer ID and notarised by Apple. Its bare executable cannot be stapled, so macOS resolves notarisation online during assessment; the protected script installer performs that assessment before installation. The public Proofpane Tray v0.3.10 release carries a signed, notarised and stapled universal disk image, and a signed, notarised and stapled .pkg installer for MDM deployment, that anyone can download, each with an adjacent SHA-256 sidecar. A separate signed/notarised desktop CI build has been verified but is not presented as a public release. Windows is still unsigned because Authenticode requires a separate certificate; Linux carries published SHA-256 digests rather than a platform-signing claim. In short: macOS signed and notarised, Windows unsigned, Linux digest-only — three separate rows, not one verdict about 'Proofpane binaries'. Note also that 'stapled: false' on the daemon means a bare executable cannot carry an attached ticket, not that it is unsigned. Enterprise re-signing remains an optional catalogue and MDM deployment path." }
    },
    {
      "@type": "Question",
      "name": "Can Proofpane analyze our codebase, and find where AI is already used in it?",
      "acceptedAnswer": { "@type": "Answer", "text": "Yes. A Proofpane engagement can open with a deep, governed analysis of your own repository and AI-client footprint: the governed code-agent reads, greps, runs and reasons over the codebase inside an isolated git-worktree sandbox — every step policy-gated, human-approvable and recorded on the audit chain — while the coverage scan reads your local AI-client configs to map which MCP servers are wired up and flags any running direct (ungoverned). You receive a map of your code and shadow-AI usage AND a working demonstration of the governance that produced it: the analysis is itself governed, not a black-box report." }
    }
  ]
}