Axion Dynamix AI OS In build

Architecture

A task is perceived once and processed concurrently by seven regions — perception (task class, risk terms, vagueness), hippocampus (hybrid memory recall scored by cosine relevance, with provenance), amygdala (salience and risk posture from the limbic state plus explicit destructive/credential terms), prefrontal (decomposition into checkable steps, verified against active goal contracts), cerebellum (procedural recall), basal ganglia (action selection among answer · plan-and-execute · delegate · ask · refuse) and anterior cingulate (conflict among decision-bearing regions). Signals compete for a global workspace; the cortex integrates the winning coalition into a Decision.

Confidence = salience-weighted region confidence × agreement × a calibration shrink learned per task class. A critic then reviews the draft and may only lower confidence or demand clarification — an invariant covered by tests over the full product space of inputs. Every decision carries temporal context (wall time, uptime, tempo, time since last feedback) and objective context (which approved goal it serves and how closely it approaches a forbidden clause).

Feedback closes the loop: the prospective prediction is resolved, memories used are credited or discounted, affect updates, an uncertainty gate and a Bayesian change-point detector adapt, and two calibration tracks (per task class, and for the critic itself) are appended. Grounding: global workspace theory, appraisal theory, metacognitive monitoring; the self-model follows the component list of arXiv:2607.12254, which defines self-awareness functionally and disclaims subjective experience — so do we.

Universal router

A loopback gateway (127.0.0.1:47392) accepting POST /v1/chat/completions (OpenAI) and POST /v1/messages (Anthropic), forwarding to any configured OpenAI-compatible, Anthropic or Ollama upstream. Point a tool at it by setting its base URL; keys are referenced by environment-variable name in router.json and never serialised.

Capability detection: declared in config → conservative model-family matrix → a one-time live probe whose result is cached. Supplementation: models without tool calling receive the axion-json-v1 fenced protocol and their output is parsed back into standard tool_calls (malformed blocks are returned as tool errors, never guessed); image parts are described by a configured vision provider for text-only models, and if none is configured the request is refused with HTTP 422 rather than silently stripped; relevant memories are injected only above a cosine-relevance floor; an optional cortex pre-pass adds plan, risk and confidence for multi-step directives; limbic modulation sets temperature/top-k only when the client left them unset.

Streaming: native SSE pass-through whenever possible; when tool supplementation forces buffering the response is still delivered as SSE and the x-axion-stream: buffered header says so. Every response carries x-axion-caps and x-axion-notes so you can see exactly what was supplemented.

Memory & privacy

Three tiers — episodic, semantic, procedural — in a SQLite store under your home directory, with embeddings, FTS5 keyword index and a provenance graph. Retrieval orders candidates by reciprocal-rank fusion of semantic and keyword ranks plus one-hop provenance expansion, then gates each on cosine relevance, so a lone unrelated record is never injected merely because it ranked first. Writes pass an affect-derived salience gate; consolidation and insight passes run on schedule. Export and purge are single commands; nothing is transmitted to Axion Dynamix.

Licensing & tiers

An entitlement is the store's signed answer to "is this account's subscription active for this product, and at which tier" — fetched with your account session, cached on-device for seven days, and re-checked at most daily. There are no licence files, issuer keys or device bindings to manage; revoking a session in the dashboard locks that device at its next check. Subscriptions are confirmed server-side against PayPal before any licence is minted; the licence lease never outlives the paid period plus a seven-day grace window. Access states: active → renewing (24 h before expiry) → grace (7 days, banner) → locked (paywall, subscription management, data export and diagnostics remain available; model calls, tools and automation pause). Changing tier creates a new subscription; nothing unlocks until the new tier's licence is installed. Device binding is optional. 36 automated tests cover forgery, expiry, downgrade attempts and tier gates.

Security model

All local services bind to loopback and reject non-loopback Host headers. Credentials live in a vault with explicit grant/deny/revoke and an audit log; the router reads provider keys by variable name only. Signed entitlements, signed skill packages (planned), and a post-quantum vault using ML-KEM-1024 and ML-DSA-87 (NIST FIPS 203/204) releasing secrets only to the verified active session, approved automations and scheduled tasks (planned). What we deliberately do not do: TLS interception, process injection, or capturing traffic from applications you have not pointed at the router.

Evidence & claims

Shipped core: cortex (31 tests), router (20), entitlements (24), paywall (12), account (6), CLI (3), plus the pre-existing substrate suites — 183 passing on the live tree. The consciousness-indicator evaluation follows arXiv:2308.08708 and reports each indicator as present, partial or absent. Benchmarks of a bare model versus the same model behind Axion (recall, coherence, calibration, latency overhead) are published with confidence intervals when measured — never before. Our citation audit corrected five mis-titled references in earlier material and is public on request.