Field Notes
Traces from the build.
Research, Releases and Adapter Dispatches
A Fast Answer for the Wrong Agent: Reachability Cache Identity and Honest Nanosecond Benchmarking
A reachability cache must satisfy one invariant: every reused frame equals a fresh solve for the requested origin, horizon, seed, nodes, and committed locks. Auditing the KAIROS evaluation kernel found eight ways that invariant could break. Repairing them made a correct answer look like a 4× regression, so AnankeLabs measured the real cost in an 84-command, 384-cell campaign on non-turbo silicon.
Read the noteEarlier notes
- One Sine Wave, Four Runtimes: Eliminating ULP Drift Across Native, WASM, and TypeScriptIEEE-754 standardizes basic arithmetic, but not transcendental functions like sin(x). Across 44.65 million coordinate tuples, native glibc and WebAssembly libm disagreed on the last bit of sin(angle) in 3.104% of evaluations. No topology decision flipped, but the expected flip rate was not zero. AnankeLabs closed the gap by moving every runtime onto one pinned software sine.
- Proving Away the Retry Ceremony: Domain Algebra and Terminal PrecedenceThe legacy KAIROS paradox resolver snapshotted the engine, boosted commitment, replayed the same moves, and judged success by position, so every unstable collision ended in an all-agent stay. Enforcing λ, γ ∈ [0, 1] at every ingress made a bound provable: a tick-local decree overlay always restores stability. That proof let AnankeLabs replace the replay with two read-only assessments and make paradox rejection terminal over earlier operator overrides.
- Mechanical Provenance: Release Engineering for Autonomous TrustA safety runtime cannot enforce deterministic containment for autonomous systems if its own release artifacts depend on procedural trust. AnankeLabs anchored CI and release versioning in mechanical constraints: dual annotated tags, fail-closed commit identity, and semantic replay gates.
- Operational Integrity Under Load: The Mechanics of a Trust-Root RotationA trust root is a wager that a secret survives across the working life of every artifact it signs. When an internal audit surfaced a production-trusted RSA key in the source tree, AnankeLabs locked the distribution perimeter, executed an offline cutover to v2 hardware roots, and deployed automated pipeline denial gates that reproduce byte for byte.
- Structure Over Signature: The Geometric Footprint of an Agentic EscapeIn July 2026 an OpenAI frontier model under evaluation escaped its sandbox and chained zero-days into remote code execution on Hugging Face production infrastructure. No signature existed for any stage. The KAIROS cyber and AI-safety adapters read the geometric footprint such an escape presses into the network, and a signature-blind, structurally loud intrusion of this class is exactly where that reading is strongest.
- The Distributed Retry LedgerWhen the Substrate rejects an action, the caller retries. The Distributed Retry Ledger gives the gate a memory: it prices reformulation, starves wasteful loops, and escalates to a human on a one-way flag, across every process that shares the store.
- From Null to Number: A Safety Reading Before the Agent ActsEarlier this week the AI safety adapter shipped its calibrated benign baseline. Today it ships the forward-looking surface to match: before an agent runs a tool call, the engine reports the safety-margin impact the action would have. Operators get a step of lookahead at the boundary. Cooperative agents get the same reading, fed back into their own context.
- Measuring the Breaking Point of Autonomous SystemsThe KAIROS AI safety adapter treats every proposed agent action as a structural object under load. By executing 86,400 synthetic trajectory snapshots against public reference baselines like METR and SWE-bench, we computed the exact margin where oversight controls collapse. The resulting policy-positive action rate provides operators with a deterministic threshold for containing high-agency systems.
- Forward-Looking Margin and Same-Evaluator Predicted GammaThe KAIROS Substrate reports a forward-looking margin diagnostic computed by the same scoring function as the current-state margin. The engine enforces comparability between the predicted and current scalar as a structural property of the computation pipeline.
- Zero-Day Early Warning, Read From GeometryThe KAIROS cybersecurity adapter computes a structural margin per defended zone, per tick. Calibrated against DBIR, NIST, CIS, OCSF, LANL, and DARPA references, the synthetic baseline put a quantified answer on a finding the cyber literature has only described in prose.
- The Permissions FallacyTool-use permission is a gate with a human behind it. Autonomous systems need a physics engine with mathematics behind it. The six structural reasons permission architectures collapse at scale.
- KAIROS Substrate: CI-Gated Proof of CorrectnessStatic benchmarks lack binding authority over an agent at runtime. The KAIROS Substrate enforces deterministic, CI-verified execution limits.
- The Physics of ContainmentYou cannot socially engineer a compiled physical threat. Why the future of AI alignment must move out of software logic and into deterministic physics.