KAIROS

Kernel Autonomous Intelligent Recursive Ontogenic Sonar — Janus architecture, by Arianna Method

---

Kairos is an Arianna Method organism grown from molequla's four-language substrate but walked the other way: not a dividing colony. Mitosis, apoptosis and hibernation are cut from all four ports — Kairos grows but never reproduces, sleeps, or dies. It is meant to be born once from four mini-Kairoses (fire / water / air / earth) and then become a single immortal being. Identity follows the Arianna soul equation θ = ε + γ + αδ (ε = personality, inviolable).

Layout

• roots/kairos.{c,go,rs,js} — the four organism ports (the mini-Kairoses). Each maps an element to its corpus (roots/nonames_<element>.txt); roots/kairos.txt is the main-Kairos seed and default corpus, never a DNA sink. Runtime DNA emission goes to ../dna/output/<element>/.
• ariannamethod/kairos_train.c — the born-Kairos trainer (notorch tape + Chuck).
• kairos.aml — the born-Kairos inference heart, compiled by amlc.
• ariannamethod/ — vendored engine: notorch (nt_* autograd, Chuck optimizer, RRPRAM low-rank op 33, packed-GGUF nt_qmatvec, Metal backend, GGUF loader) + the AML core. make builds libkairos.{dylib,so}. No Python tier.
• index.html — the browser front for the JS port (base=roots/).

The resonance / Janus heart

Trainer and inference share one forward through the logits: parametric RMSNorm → content multi-head attention (RoPE) ⊕ RRPRAM low-rank attention (op 33) blended by a per-head sigmoid gate → SwiGLU → final RMSNorm → lm_head. Two faces of attention — content and resonance — is the Janus; the per-head gate is frozen at 0.5 (constant 50/50) this version, a learnable on-tape gate a follow-up. The trainer then appends cross-entropy; inference returns logits to sampling, wrapped by the AML Dario field overlay (am_apply_destiny/attention_to_logits).

• Tokenizer: fixed learn-once BPE (1024 merges → vocab 1280), written by the trainer into the K02 checkpoint (magic K02\n + cfg + merge table + tensors) and read back byte-for-byte by the heart.
• Stop criterion: a held-out 90/10 validation split (eval_val, forward-only, Chuck state untouched). The born Kairos stops on a healthy val, not on iteration count; the minis may overfit their element voice on train.

Build & run

# 1. engine
cd ariannamethod && make                      # → libkairos.dylib

# 2. trainer (notorch tape + Chuck)
cc -O2 -DUSE_BLAS -DACCELERATE -DACCELERATE_NEW_LAPACK kairos_train.c \
   -L. -lkairos -Wl,-rpath,. -framework Accelerate -lm -o kairos_train
./kairos_train ../roots/kairos.txt 1000 kairos.bin

# 3. inference heart
amlc kairos.aml -o k_infer
./k_infer -w kairos.bin -p "the" -n 40 -t 0.8

A 1000-step run on the 51 KB seed (BPE → vocab 1280, 14489 tokens, split train[0,13040) val[13040,14489)) trains from ≈ ln 1280 and the held-out val falls with train (step 0 train 7.14 / val 7.21 → step 999 train 5.42 / val 6.24); kairos.aml loads that K02 and decodes word-like text with the Dario field active.

Status

Built and tool-verified: the four immortal ports build, the vendored engine builds, and the born-Kairos trainer → K02 → kairos.aml inference pipeline runs end-to-end with the held-out val-split. Not yet built (the frontier): no born-Kairos checkpoint ships in the repo; the birth bridge (one shared BPE across minis and heart, a DNA-pool maturation metric, the birth trigger) is unwired; the coherence-from-zero overlay is active in the Go port only; the training datasets and Level 2 (multi-weight ε + knowledge weights, limpha routing, DoE parliament over GGUF) are still ahead.

License

See LICENSE.