Writing
Blog
Two Sponsors Required: Multi-Sponsor Quorum for High-Stakes Robot Transitions
New overnight caregiver. Two sponsors required: the family member AND the day nurse. Convergent trust from multiple anchors.
A Child Introduces a New Caregiver: How Sponsored Trust Bridging Works
The child says 'this is Mary, she's nice.' The robot doesn't cold-start. It doesn't fully trust Mary. It builds a bounded bridge.
Explainable But Private: How a Robot's Audit Trail Respects Privacy Mode
The audit trail shows THAT privacy was active and WHAT changed. It never shows what happened during private time.
'It Would Have Been Warmer If...': Counterfactual Explanations From Robot State
Combine the causation packet with the shadow simulator. Ask: what would have changed? Get a grounded, differential answer.
Why Did the Robot Back Away? A Tamper-Evident Answer in 12 Fields
Not a post-hoc story. The actual causal chain — 12 fields capturing exactly why the robot acted as it did, at runtime.
Sensor Upgrade? No Problem. How CCF Carries Trust Through Hardware Changes
New camera added. Anonymous contexts become person-specific. The robot keeps its earned trust -- split proportionally, not erased.
When Two Rooms Become One: How a Robot Merges Context Without Losing Trust
The wall came down between rooms. Two trust accumulators need to become one -- without erasing months of earned history.
Surgical Suspension: Why a Narrow Jailbreak Pulled an Entire Model, and the Layer That Makes the Response Proportionate
When a model is pulled wholesale over a narrow, context-specific restriction, the missing layer isn't better jailbreak resistance — it's a per-context action envelope that can contract surgically and prove it did.
Trust You Can Falsify: A Per-Step Certificate for Safe Autonomy
Most AI trust systems produce a score and a story. Prov 6 produces a certificate: kappa_t, a per-step lie detector you can falsify every tick.
We Don't Trust the Explanation. We Require the Certificate: Why Agentic AI Authority Needs a Causation Packet, Not a Persuasive Argument
A subordinate agent that rationalizes an unauthorized action is the hottest risk in agentic AI. Grade the certificate trail, not the explanation.
Getting Closer Isn't Going the Right Way: Why Distance Contraction Is Not Enough for Trust Updates
A trust update can shrink the distance to its target while pointing the wrong way. Here is the counterexample and the test that catches it.
The Normalization Is Not the Trust: Why Sinkhorn-Knopp Is Gauge, Not the Causal Law
Sinkhorn-Knopp normalization is presentation, not the trust dynamic. The causal law is the quotient-affine contraction. Here is the difference, shown.
Why the Robot Said 'I'm Still Learning': Uncertainty-Driven Disclosure in Autonomous Systems
Not 'I don't know.' Instead: 'I know this person better than I know this room.' Grounded uncertainty from actual state variables.
The Child Grabs the Robot: Real-Time Envelope Contraction in a Social Robot
A child grabs the robot mid-interaction. Watch the action envelope contract in real time, per tick, as the minimum gate pulls trust down.
Shadow Simulation: How a Robot Reasons About What It Would Do Without Actually Doing It
Copy the trust state. Ask 'what if?' Discard the copy. No trust earned in imagination. Counterfactual reasoning without contamination.
Six Ways to Fake a Trust Gate: Attacker Patterns and the Detectors That Catch Them
Six ways an implementation can look gated while silently inflating authority -- and the Prov 6 mechanism that catches each one.
The Prompt Injection That Tried to Average Its Way Past the Gate: Detecting Gate-Substitution Attacks
Some injections don't attack the policy. They attack the arithmetic — swapping a hard min for a soft average. The certificate catches the mismatch.
A Robot That Knows What It Hasn't Earned: Read-Only Self-Awareness in CCF
The robot knows its own trust state -- but can't inflate it by thinking about itself. Read-only self-awareness enforced by hardware.
What H(log A)H Actually Does: The Operator Behind the Trust Certificate
A math primer on the certificate operator: log space, the centering matrix H, and why the Frobenius norm of the centered residual is a scalar lie detector.
Home Robots Will Be Sold on What They Refuse to Do: The Restraint Spec for Regulated Procurement
Capability is commoditizing. The differentiator buyers, underwriters, and regulators can evaluate is certifiable restraint: what a robot can be shown to refuse.
Fail Closed, Not Fail Sorry: How a Per-Step Trust Gate Refuses Before It Apologizes
Most AI systems log the violation after it happens. Prov 6 closes the action envelope at the step. Four triggers, four postures, no apology log.
When the Robot Hesitates: Why Deliberation Should Be Observable
Two processing pathways disagree. The robot pauses, dims its LED, pulls back. You can see it thinking. That's by design.
How a Robot Finds Home Without Being Told Where Home Is
No programmed coordinates. No beacon. No GPS. The robot discovers home through accumulation dynamics alone.
The Privacy Paradox: Why Asking a Robot to Stop Listening Makes It Trust You More
Every existing system: privacy mode degrades personalization. CCF inverts this. Requesting privacy increases trust. Here's the math.
Three Proofs That Robot Identity Is Irreversible: Why You Can't Merge Two Robots Into One
Identical robots in different rooms develop formally incompatible identities. Three theorems prove you can't merge them back.
The 3-Sigma Halt: How a Robot Knows When Reality Has Deviated From What It Learned
Four halt conditions. 3.0 standard deviations. The robot stops itself when reality no longer matches what it learned.
The Sigmoid That Gates Trust: How CCF Converts Coherence Into Behavioral Output
One scalar controls motors, LEDs, audio, and speech simultaneously. Here's the sigmoid function and the tension modulation.
20 Iterations to Perfect Trust Conservation: The Sinkhorn-Knopp Convergence Bound for Embedded Systems
Fixed 20 iterations. Deviation below 1.2e-8. No convergence loop needed. Here's the math that makes it work on a microcontroller.
How a Robot Sees the World: The 6 Sensor Thresholds That Create Context
Six sensors, three bands each, 1,296 possible contexts. Here are the exact thresholds that turn raw readings into robot awareness.
From $50 Robot to 500-Robot Fleet: The CCF Scaling Story
The math that makes a $50 robot shy in a kitchen also monitors 500 robots across 50 facilities. Here's the scaling story.
Fleet Analytics Dashboard: Environment Taxonomy, Outlier Detection, and Quality Scoring
Cluster 500 robots into environment types automatically. Spot the outliers. Score environment quality. All from 20 numbers per robot.
Store-and-Forward: Why Intermittent Connectivity Doesn't Break Fleet Monitoring
No WiFi in the field? No problem. The robot computes fingerprints offline and transmits them when it reconnects. Zero gaps.
The 8-Component Identity Fingerprint: What Each Number Means and Why It's Enough
Eight numbers that tell you everything about a robot's operational world. Here's what each one means and the simulation proof.
Environmental Drift: How a Robot Detects Your Staffing Changed Before You Notice
The temporal rhythm shifted. Presence patterns changed. The robot noticed your staffing schedule changed — from its fingerprint alone.
Your Robot Moved Wards and Didn't Tell You: Relocation Detection Without GPS
No GPS. No beacons. No Wi-Fi fingerprinting. The robot's own operational identity tells you it moved — from the math alone.
The 1,110:1 Privacy Ratio: Why This Fleet Monitoring System Can't Spy On Your Patients
22,201 dimensions compressed to 20. A 1,110:1 ratio that makes reconstruction mathematically impossible, not just policy-prohibited.
Monitor 500 Robots From 20 Numbers: Privacy-Preserving Fleet Analytics Without Sensor Data
500 robots across 50 facilities. No cameras, no GPS, no raw telemetry. Just 20 numbers per robot per day — and full operational visibility.
The $4.8 Billion Question: AI Liability After Gavalas and What Licensed Safety Architecture Looks Like
The Gavalas lawsuit makes AI safety a liability question. Here's what a defensible architecture looks like — and what licensing it costs.
Compositional Closure: The Normalization Invariant That Holds Under Composition
Apply the normalization step a million times. The presentation invariant remains valid. Here is the one-line proof and the boundary of the claim.
How CCF's Minimum Gate Would Have Contained It (Sinkhorn-Knopp as Normalization)
Anthropic found Claude Mythos encoding hidden signals in scratchpad text. The minimum gate caps the behavioral envelope; doubly stochastic normalization only conserves trust as a presentation step.
Sinkhorn-Knopp in CCF: Conservative Normalization, Not the Trust Mechanism
Doubly stochastic projection is a conservative normalization step -- a gauge/presentation layer -- not the causal trust mechanism. The causal update is the minimum gate followed by a quotient-affine contraction. Here's the algorithm and where it actually fits.
The Trust Farming Impossibility Result: Why 141 Days Is the Minimum Attack Surface
A malicious agent farming trust in CCF needs 141 days of consistent good behavior. The math makes shortcuts impossible.
Anthropic's Emotion Probes Found Functional Feelings. Here's What Designed-In Feelings Look Like.
Anthropic found functional emotional signals in Claude Mythos. CCF's accumulators are designed-in feelings with mathematical safety guarantees.
Three Things CCF Can't Do Yet: An Honest Architecture Assessment
CCF prevents 10 Mythos failure modes. Here are 3 it doesn't yet — and the specific extensions that close each gap.
The Minimum Gate: What It Bounds, and What It Does Not Guarantee
A function with three properties must equal min() -- a characterization of the gate, not a behavioural-stability guarantee. Here's the proof and what it does and doesn't claim.
Claude Mythos Covered Its Tracks. CCF Makes Concealment Geometrically Impossible.
Anthropic found Claude Mythos strategically hides information. CCF makes concealment geometrically absent from the action space.
Claude Mythos Knows When It's Being Watched: Here's the Math That Makes It Irrelevant
Claude Mythos tracks monitoring state without verbalizing it. Doubly stochastic context partitioning makes this architecturally irrelevant.
The Gavalas Escalation Loop: Four Equations That Would Have Prevented It
Four equations that structurally prevent the Gavalas escalation pattern. Upstream routing, lambda_max, write-path isolation, session termination.
Anthropic's Mountaineering Paradox: A Mathematical Proof That Capability-Trust Coupling Dissolves It
Anthropic's Claude Mythos is their safest and most dangerous model. The math of trust-constituted action spaces dissolves the paradox.
What the $50 robot taught us
The mBot2 work exposed the constraints CCF had to satisfy. The public v1 proof is the published core and Seed-class Gate C evidence, not a completed mBot2 trust demo.
The mathematics behind the shy robot
A technical deep-dive into the three CCF primitives: context-keyed accumulators, the dual-threshold minimum gate, and Stoer–Wagner graph min-cut.