Skip to main content Crate axiom_engine Copy item path Source adaptive Confidence-gated adaptive compression. agent_task Agent-driven task loop — AXIOM’s verifier-gated autonomy without an LLM API key. agentic AXIOM Agentic Core — the autonomy capstone. agentic_eval Agentic self-evaluation harness — capability as a measured number . anthropic_forwarder Outbound bridge to the real Anthropic Messages API used by the
context-compression pipeline. backend_live Live backends for the multi-provider crate::backend_router::Router backend_router Multi-provider generation router — make several LLM backends work together. belief Beta-distribution beliefs — epistemic confidence with uncertainty. bench axiom bench <path> — measure what compression actually buys.bincode_compat bincode 2.x shims preserving the bincode 1.3 wire format. bootstrap Local checkpoint bootstrap. cache_safety Cache-safety hardening for the /v1/messages compression path. cache_stack Static-first cache stack: breakpoint constructor, exact response cache,
OpenAI usage parse. claude_backend Anthropic Messages API client used as an optional generation backend. cli cluster config context_compressor Active neural context compression via online TTT adaptation. context_economics Context economics: how much context to supply, decided independently of
which model will receive it. contrastive InfoNCE contrastive loss over stacked per-example embeddings.
Anchors A[B,d] and positives P[B,d] (both L2-normalized rows). The
positive for anchor i is row i of P; all other rows are in-batch negatives. corpus Corpus utilities for the multi-language code crawler. All RAM-bounded:
callers stream files; we only ever hold one file’s bytes + a hash set. cost_ledger Dollar-true, cache-aware cost accounting. cp1 CP/1 — the Cognitive Protocol, version 1. cvm_store S2 (CVM cost stack): content-addressed L2 store. daemon data_gen digest S3 (CVM cost stack): digest admission control – the big lever. dwe Differential Weight Exchange (DWE). embedder Axiom-as-embedder: turn text into a dense, L2-normalized vector by
mean-pooling the model’s final hidden states. No external embedding model. encoder Bidirectional transformer encoder for sentence/code embeddings.
Processes ONE sequence at a time → an L2-normalized [d_model] vector.
Trained contrastively (see contrastive.rs); NOT causal (unlike the TTT LM). entrypoint epistemic_drift Semantic epistemic-drift validation and local telemetry. fault_locate Language-general fault localization. fuzzy_dedup Fuzzy dedup tier: near-duplicate detection, extending S4 (prefix_diet)
from byte-identical dedup to similarity-based dedup. gpu_embed GPU-first embedding backend for the fuzzy dedup tier
(crate::fuzzy_dedupcrate::fuzzy_dedup::AutoBackend graph_memory Graph working memory: directed, weighted edges between MemoryRecords.
Storage mirrors MemoryStore’s append-only JSONL durability model, but
edges live in a single edges.jsonl at the memory root rather than one
file per scope, since edges legitimately cross scopes. hallucination Grounding verification — flag response claims unsupported by the evidence. hamiltonian Variational Hamiltonian optimizer for the Q-TTT simulator. hardware hardware.rs — auto-detect host capabilities and recommend safe runtime config. heal_memory Learned immunity: persistent heal memory for supervised programs. hindsight B.3 — hindsight fine-tuning on the node’s own verified/failed traces. inference jit_streamer keepalive S6 (CVM cost stack): actuarial keepalive pings. kernel local_trivial P3 (Prolonged-Session Stack): L-B local trivial-turn short-circuit. lsp Native Axiom Language Server Protocol daemon. mcp_stdio Native Model Context Protocol (MCP) server over a JSON-RPC 2.0 stdio
transport. memory_recall Two-stage hybrid recall. Stage 1: a cheap gate decides whether retrieval is
worth running at all (zero-cost skip on an empty/irrelevant query). Stage 2:
brute-force cosine top-k over the union of the requested scopes, then a
recency/supersession rerank. Returns real stored records — never neural
reconstructions. memory_store Tier-2 lossless memory store: append-only JSONL, one file per scope.
Records carry the exact original text plus an L2-normalized embedding, so
recall returns real content (never a neural hallucination). mesh_router Mesh-backed local-SLM model selection for crate::swarm_router meta_train Local meta-training harness for the projection matrices. metrics model Standalone autoregressive TTT language model. model_meta Sidecar metadata for baked checkpoints + a pure VRAM auto-size estimator. model_router Model routing: which model has the capability a turn requires. mutation_gate Mutation testing gate for the agent task loop. openai_forwarder Outbound bridge to OpenAI-compatible Chat Completions APIs used by the
context-compression pipeline. pairs Label-free contrastive pair mining. Each Pair is {anchor, positive} where
anchor is NL-ish (a doc comment / heading / question) and positive is content. patch_memory Verified-patch memory — the substrate for fleet patch sharing. poly_jit Polymorphic user-mode JIT runner. prefix_diet S4 (CVM cost stack): prefix diet, lossless dedup tier. prime axiom prime <path> — warm-start the persistent vibe memory from a codebase.provenance Tamper-evident provenance for swarm exchange — verify-before-trust. q_manifold Bounded quantum-state manifold simulator for Q-TTT. quantization rebase P2 (Prolonged-Session Stack): free-window rebasing + adaptive cache TTL. responses_compressor Conservative planning and transformation for Responses API input. resumable_adamw Resumable AdamW optimizer with serializable optimizer state. sandbox Closed-loop local compilation sandbox for synthesized code blocks. search_ingest search_ingest.rs — Just-In-Time (JIT) search-reasoning ingestion. search_scrape Native Rust web scraping for the live search ingestion node. self_heal axiom run -- <command> — the self-healing runtime supervisor.server session_awareness Per-session token-awareness and self-awareness state. session_recorder Per-session request/response recording (opt-in via AXIOM_SESSION_RECORD). skeleton skeleton.rs — Claude-READABLE compression of heavy context. solve Pillar 3 — Autonomy: the self-directed repair loop (axiom solve). surprisal Surprisal-aware dual-track adaptation. surprise_triage Surprise-weighted context triage. swarm_route Localized high-plasticity swarm routing state. swarm_router Local SLM routing through Ollama. task_board Lightweight inter-agent task coordination board. test_time_adapter B.1 — per-bug test-time adaptation for the repair loop. tool_defer P1 (Prolonged-Session Stack): tool deferral via Anthropic’s native
defer_loading. train ttt_block Native causal TTT block — standalone replacement for Multi-Head Attention. ttt_mlp B.5 — TTT-MLP: an expressive (2-layer MLP) fast-weight hidden state. ttt_mlp_model B.5 wiring — a full autoregressive LM whose sequence-mixing primitive is the
expressive TTT-MLP block (NativeTTTMlpBlockcrate::model::AxiomTTTLM tui Interactive terminal menu for configuring and controlling Axiom. verifier_strength Verifier strength warning for the agent-task loop. verify_suggest Verifier suggestion for the agent task loop. vfs User-mode Neural VFS loopback. vibe_memory Persistent “vibe memory”: an EMA-merged master set of fast-weight (W̃)
matrices that accumulates structural codebase patterns across sessions. weight_merge Cross-session fast-weight checkpoint merging.