Correct the org card: remove protected engine internals, drop false public-availability claims, fix demo links and status

README.md

@@ -10,192 +10,171 @@ license: other
 
 # ARF AI
 
-**Trusted decision infrastructure for AI‑driven operations.**
+**Control plane for autonomous AI.**
 
-> 🔒 **The core ARF engine is access‑controlled and not open source.**  
-> Public specification and demo UI are Apache 2.0.  
-> Pilots: outcome‑based pricing – pay for *verified risk reduction*, not API calls.
+ARF sits between an agent's intent and its execution. Every proposed action is
+evaluated against deterministic policy before anything happens, and the
+decision is recorded before the action runs — not after it succeeds.
+
+> 🔒 **ARF is commercial software and is not open source.** The core engine,
+> API control plane, gateway and enterprise extension are access-controlled.
+> Access is through the sandbox, a pilot, or an enterprise agreement.
 
 ---
 
-## Why ARF AI? Why now, not tomorrow?
+## The problem
 
-Every AI‑assisted decision you *don’t* govern is a liability you *already own*.  
-Unreviewed, unaudited AI actions in production are like signing blank checks.
+An agent that can restart a service, delete a volume, move money or approve a
+claim is not a chatbot. It is an operator without a review process.
 
-- **Loss aversion** – The pain of a regulatory fine or operational outage is twice as powerful as the pleasure of moving fast. Avoiding a $500k incident is more valuable than gaining $200k in speed.
-- **Ambiguity aversion** – Teams hate unclear approval processes. ARF replaces tribal knowledge with deterministic, explainable outcomes.
-- **Urgency of compounding risk** – Each ungoverned AI decision increases your exposure. The longer you wait, the harder to retroactively audit.
+The usual answer is to ask a model to check its own work. That produces a
+judgment which can differ between two identical requests — so it cannot be
+audited, cannot be appealed, and cannot be shown to anyone as a rule.
 
-**ARF AI gives you control today.** Not after the breach. Not next quarter. **Now.**
+ARF's answer is that some decisions must not be delegated to judgment at all.
 
 ---
 
-## What ARF AI does
+## What ARF does
 
-ARF sits between AI intent and execution. It evaluates every request in real time and returns one of three outcomes:
+Every request returns one of three outcomes:
 
-- ✅ **Approve** – action is safe, within policy, and uncertainty is low.
-- ⚠️ **Escalate** – insufficient confidence or high impact → human reviewer.
-- ❌ **Deny** – violates policy, exceeds risk tolerance, or data incomplete.
+- ✅ **APPROVE** — within policy, and uncertainty is low enough to act.
+- ⚠️ **ESCALATE** — the system could not establish that this is permitted. An
+  input is missing, or the action is above the bar this policy clears alone.
+- ❌ **DENY** — violates policy or exceeds risk tolerance.
 
-Every decision is:
-- **Logged** with a full audit trail (who, what, when, inputs, output).
-- **Explained** in plain language – no black boxes.
-- **Signed** (Ed25519) for non‑repudiation.
+**ESCALATE is not a softer DENY.** Collapsing the two destroys the difference
+between *this is forbidden* and *this needs a person*, and that difference is
+the whole of what a reviewer needs in order to act.
 
----
+Every decision is:
 
-## For whom?
+- **Deterministic** where it matters — identical inputs, identical policy
+  version, identical answer, every time.
+- **Explained** — a refusal nobody can explain is not auditable, so a refusal
+  without reasons is rejected rather than passed downstream.
+- **Recorded in a hash-chained audit log** — each entry carries the hash of
+  the entry before it, so modifying any past entry breaks every entry after
+  it. Entries are additionally **Ed25519-signed when a signing key is
+  configured**; without one they are still written and still chained, just
+  unsigned. Chaining is integrity, not authorship — the two are not the same
+  claim and we do not merge them.
 
-| Role | Value |
-|------|-------|
-| **VP of Engineering** | Reduce operational risk from AI‑generated actions. |
-| **CTO / CIO** | Maintain speed while adding governance. No full redesign required. |
-| **Head of Compliance** | Complete, tamper‑evident audit trail for regulators. |
-| **Security Officer** | Deterministic policy gates that cannot be silently overridden. |
-| **AI/ML teams** | Get your models into production *with* trust and oversight. |
+Risk scoring is Bayesian and reports its own uncertainty, which is what drives
+escalation rather than a coin-flip at the threshold.
 
 ---
 
-## Core engine (protected – no public access)
-
-ARF’s core is a Bayesian governance engine. Key technical components:
-
-### Bayesian risk fusion
-Combines online conjugate priors, offline HMC logistic regression, and optional hierarchical hyperpriors.
-
-$$
-\text{risk} = w_{\text{conj}}\cdot\frac{\alpha}{\alpha+\beta} + w_{\text{hmc}}\cdot p_{\text{hmc}} + w_{\text{hyper}}\cdot \mu_{\text{hyper}}
-$$
+## Who this is for
 
-Weights adapt with data volume. Posterior → 90% HDI for uncertainty.
+ARF is built first for teams where **an agent can already change production
+state and nobody owns governance yet**. A practical test:
 
-### Expected loss minimisation
-Chooses the action that minimises expected cost:
+- Can one of your agents take an action a customer would notice, without a
+  human in the loop?
+- If it did that wrongly at 3am, would you be able to show what was decided,
+  by which rule, and on what inputs?
+- Is there a named owner for that question today?
 
-$$
-\begin{aligned}
-L_{\text{approve}} &= \text{COST\_FP}\cdot R + \text{COST\_IMPACT}\cdot b_{\text{mean}} \\
-L_{\text{deny}} &= \text{COST\_FN}\cdot(1-R) + \text{COST\_OPP}\cdot v_{\text{mean}} \\
-L_{\text{escalate}} &= \text{COST\_REVIEW} + \text{COST\_UNCERTAINTY}\cdot\psi
-\end{aligned}
-$$
+If the first two are *yes* and the third is *no*, that gap is what ARF closes.
 
-### Execution ladder (Rust)
-Mechanical gates: `license` → `confidence` → `risk` → `rollback` → `causal`. Deterministic, auditable.
+| Role | What they get |
+|------|---------------|
+| **Founder / CTO at an AI-native team** | Governance before the first incident, without a redesign or a compliance hire |
+| **Platform & SRE leads** | Deterministic gates on agent-initiated production change |
+| **Security & compliance leaders** | A tamper-evident record of what was decided and why |
+| **AI/ML teams** | A route to production that survives review |
 
-### Lyapunov stability
-Quadratic candidate $V(x,r) = \alpha r^2 + \beta\|x - x_{\text{des}}\|^2$ guarantees healing actions converge.
-
-### Cryptographic signing
-Ed25519 signatures for `HealingIntent` – non‑repudiable governance.
+Larger and regulated deployments are served too — through the enterprise tier,
+with SSO, multi-tenancy and an SLA — but the sandbox and pilot are designed for
+small teams shipping agents now.
 
 ---
 
-## Public specification (Apache 2.0)
-
-- [Data models & API contracts](https://github.com/arf-foundation/arf-spec) – `InfrastructureIntent`, `HealingIntent`, `RiskScore`, `GovernanceLoop`
-- [Mathematics](https://arf-foundation.github.io/arf-spec/mathematics/) – Full derivations, Lyapunov proof
-- [Governance loop](https://arf-foundation.github.io/arf-spec/governance/) – Constants (`COST_FP`, `COST_FN`, `EPISTEMIC_ESCALATION_THRESHOLD`)
-
----
+## Live demos
 
-## Live demos (mock data only)
+Both run on mock data. They demonstrate the decision surface; they do not
+enforce anything.
 
-| Demo | Description | Link |
-|------|-------------|------|
-| **Risk Dashboard** | Adjust priors, see HMC simulation, semantic memory retrieval | [Launch](https://huggingface.co/spaces/A-R-F/Agentic-Reliability-Framework-v4) |
-| **Sandbox API** | Mock FastAPI endpoint, interactive `/docs` | [Try API](https://huggingface.co/spaces/A-R-F/Agentic-Reliability-Framework-API/docs) |
+| Demo | What it shows | Status |
+|------|---------------|--------|
+| [Risk Dashboard](https://huggingface.co/spaces/ARF-AI/Agentic-Reliability-Framework-v4) | Adjust priors, watch scoring and escalation respond | Running |
+| [Sandbox API](https://huggingface.co/spaces/ARF-AI/ARF-Sandbox-API) — [`/docs`](https://arf-ai-arf-sandbox-api.hf.space/docs) | Interactive OpenAPI endpoint | Running |
 
-> All demos use simulated responses. Real enforcement requires pilot access.
+Mock responses only. Real enforcement requires sandbox or pilot access.
 
 ---
 
-## Public repositories (Apache 2.0)
+## Public code
+
+| Repository | License | What it is |
+|------------|---------|------------|
+| [`arf-pattern-examples`](https://github.com/petter2025us/arf-pattern-examples) | Apache-2.0 | A runnable reference implementation of the **pattern** — propose, decide deterministically, record, then execute. Three worked domains, hash-chained audit with a tampering test, and a fail-closed external-policy delegation. Contains none of ARF's engine. |
 
-| Repository | Description |
-|------------|-------------|
-| [`pitch-deck`](https://github.com/arf-foundation/pitch-deck) | Public overview and investor materials |
+That repository is independent reference code. It is not ARF, it does not
+contain ARF's risk engine, authority system or execution-admission protocol,
+and it is useful whether or not you ever talk to us.
 
-Private repositories (Core Governance Engine, API Control Plane, Enterprise Extension) are **pilot / enterprise only**.
+Everything else — core engine, API control plane, gateway, enterprise
+extension — is private and stays private.
 
 ---
 
-## Access model
+## Access
 
 | Layer | Availability | Purpose |
 |-------|--------------|---------|
-| **Public sandbox** | Mock responses only | Demonstration and evaluation |
-| **Pilot program** | Invitation‑only, time‑limited | Validate use case with controlled access |
-| **Enterprise core** | Protected production engine | Commercial deployment, full enforcement |
+| **Sandbox** | Free, 1,000 evaluations/month, simulation only | Evaluate the decision surface |
+| **Pilot** | Time-limited, free by review | Validate on your own use case |
+| **Enterprise** | Custom deployment | Production enforcement, SSO, multi-tenancy, SLA |
 
----
+Pilot pricing is outcome-based: you pay for verified risk reduction, not per
+API call.
 
-## Pilot access – outcome‑based pricing
+👉 **[Request pilot access →](https://www.arf-ai.com/signup)** · [arf-ai.com](https://www.arf-ai.com/)
 
-You don’t pay per API call. You pay for **verified risk reduction**.
-
-👉 **[Apply for pilot access →](https://www.arf-ai.com/signup)**
-
-When applying, include:
-- Organization name
-- Use case (e.g., infrastructure change review, AI‑assisted operations)
-- Expected volume (approx. evaluations per month)
-- Cloud environment (AWS, Azure, GCP, on‑prem)
-
-Pilot is **time‑limited and free** for qualified organizations. No commitment – just validation.
+When applying, include your organization, the use case, rough evaluation
+volume, and where it runs.
 
 ---
 
-## Why enterprises choose ARF AI
+## Design principles
 
-- **Deterministic** – identical inputs → identical decisions.
-- **Explainable** – every decision includes a human‑readable justification.
-- **Auditable** – complete, tamper‑evident audit trail.
-- **Cloud‑agnostic** – runs anywhere (AWS, Azure, GCP, on‑prem).
-- **Secure** – SSO, RBAC, SOC2‑ready design.
-- **Preserves speed** – adds governance without forcing a full system redesign.
+1. Deterministic where the decision must be reviewable; probabilistic only
+   where the world genuinely is uncertain.
+2. A decision must be explainable to someone who was not there.
+3. Uncertainty routes to a person; it never silently becomes approval.
+4. The record is written before the action, not after it succeeds.
+5. An unreachable dependency is not a permissive one — no control degrades
+   quietly into no control.
+6. Auditability is built in, not retrofitted.
 
 ---
 
 ## Trust & compliance
 
-ARF is architected for regulated environments:
-
-- Tamper‑evident audit trails (regulatory review ready).
-- Mechanical enforcement – policy gates cannot be bypassed.
-- Explainable reasoning – suitable for third‑party audits.
-- Supports GDPR, SOC2, ISO27001 alignment.
-
----
-
-## Product principles
-
-1. Governance should be deterministic where possible.
-2. High‑impact decisions must be explainable.
-3. Human review remains available for uncertainty.
-4. Auditability is built in, not retrofitted.
-5. Enterprise deployment must not require abandoning existing infrastructure.
-6. Commercial terms reflect actual value delivered (risk reduction).
-
----
-
-## Short version
-
-ARF AI is the decision layer between AI intent and production execution.  
-It evaluates, decides, and logs – so you can move fast without losing control.
+ARF is **architected for SOC 2 readiness** and for review in regulated
+environments: tamper-evident audit trails, policy gates that cannot be silently
+bypassed, and explainable reasoning suitable for third-party audit.
 
-**Don’t wait for the incident that forces governance. Govern every AI decision today.**
+To be precise about what that does and does not mean: this describes how the
+system is designed. It is not a certification, not an attestation, and nothing
+here establishes your compliance with any framework. ARF is also **not** a
+functional-safety layer — it governs decisions above certified safety systems
+and does not replace them.
 
 ---
 
-## Legal & license
+## Legal
 
-- **Core engine** – proprietary, trade secret. No public access.
-- **Public specification and demo UI** – Apache 2.0.
-- **All materials** – may not be copied, redistributed, reverse engineered, or used for AI training without written permission from ARF Foundation.
+- **Core engine, API control plane, gateway, enterprise extension** —
+  proprietary. No public access.
+- **`arf-pattern-examples`** — Apache-2.0, and genuinely so.
+- **This page and ARF marketing materials** — © ARF Foundation. Not to be
+  copied, redistributed, reverse engineered, or used for AI training without
+  written permission.
 
 ---
 
-*© ARF Foundation. All rights reserved.*
+*© ARF Foundation. All rights reserved.*