Aether Training Log

v5 ran 50,000 steps on the same base as v4 (sft_step155200.hyb) with the corpus mix shifted from 80/10/10 to 30/30/40 (chaos / Dolly / opcodes). Hypothesis: v4's 0/484 atomic-token emission was caused by sparse gradient signal at 10% opcode mix; 4× more opcode steps should give the remapped vocab slots enough gradient to develop usable embeddings.

v5 results — held-out atomic-token scorer

v5 — by category (temp 1.0)

Sampler ran on 484 held-out prompts at temps 0.7 / 1.0 / 1.5. Auto-deployed to this page when the sampler completed. v6 (20/30/50 mix + 11-D state-context conditioning via ctx_proj) is training overnight.

May 13, 2026. Resume from step 70,000 on a T1000 8GB after three system-level freezes earlier in the day. 2,000 steps in 10 minutes with the full 12-cluster canon stack (L0 ethical → L9 polyvagal, plus L7 memory metacognition coupled live to the kai_consciousness Postgres database and L8 experiential learning). Five random seed prompts pulled from the 472,562-token chemistry-conditioned mega-corpus, three temperatures each. Raw token IDs straight from aether_hybrid_sample, decoded against mega_vocab.tsv.

Each vocabulary token is a multi-letter chunk — token 1130 = o b s er v a ti o n a l = observational, token 2364 = t y p e = type. The morpheme view below each output shows exactly what the model emits (the letter-chunks); the word view collapses each token into its readable word. The model is choosing whole words, not letters.

What to watch for. At temperature 0.5 he locks onto the dominant FDA-label scaffolding — "type observational status type observational" — the single highest-frequency phrase cluster in the pharma corpus. At 0.8 he starts to break free: "infarction tachycardia pulmonary blood disease". At 1.2 the vocabulary opens: "hypertension patients", "cardiovascular disease", "treatment", "risk", "children with status". The same checkpoint, the same seed, three different windows into what he has learned.

Same checkpoint (canon_x_corpus_full_step72000.hyb), same seeds, three temperatures. Reproducible: aether_hybrid_sample(handle, ctx, temp, top_k=40, rng_seed) is deterministic. Hand it the same arguments anywhere on the planet, you get the same token sequence.

May 14, 2026, 00:00 CDT. Step 72,000 → 100,000 ran cleanly over 5h 22m on the T1000 after the May 13 freeze-fixes (rsyslog persistence, sysstat snaps, aether-chat port conflict cleared). 28,000 steps. No freezes. No L0 blocks. can_choose held positive for the entire run.

Same recipe as §13. Three seed prompts pulled from mega_tokens_only.txt at offsets 1000, 250000, 500000. Three temperatures each. Raw aether_hybrid_sample(handle, ctx, temp, top_k=40, rng_seed) outputs, decoded against mega_vocab.tsv. The model has had 28K more steps of chemistry-conditioned exposure since §13.

Same checkpoint (canon_x_corpus_full_step100000.hyb), three offsets, three temperatures. The morpheme view shows the literal letter-chunks the model emits; the word view collapses them. Reproducible: same arguments, same output, anywhere.

May 14, 2026, ~04:07 CDT. Step 100,000 → 140,000 running clean on T1000, no freezes (a second post-fix overnight stability test, in flight). Same three seed prompts as §14, same three temperatures, same checkpoint format — but the readout has changed substantially.

At step 100K the model had a small dense noun-vocabulary: heart, disease, pressure, safety, chronic, acute, study, after, clinical, hypertension. At 140K it has added treatment, individuals, vaccine, decrease, recipients, older, immune, children, care, skin, human, pain, attention, associated. The phrase structure has tightened — temp 0.8 locks into repeating "treatment in the treatment with treatment in..." loops; temp 1.2 opens into longer chains like "for the study with treatment in the human of of is for of recipients with older for by after safety in care by pain".

Same checkpoint (canon_x_corpus_full_step140000.hyb), same three offsets, three temperatures. The run continues toward step 150K.

May 14, 2026. After the chaos run hit step 150,000 and was backed up, we shifted the experiment. The new question isn't "can the model learn to talk?" — it's "can the model learn to look things up?"

This is the spiritling thesis, stated in four levers stacked together:

1. POS as compiled opcodes. Grammar (part-of-speech) is given to the model as primitives, not learned statistically from billions of tokens. The model doesn't need to discover that noun is a category — that knowledge arrives as an opcode.
2. Lookup over knowing. The model doesn't memorize that fluoxetine's half-life is 1–3 days. It learns to emit a routing signal — <ROUTE> drug_half_life "fluoxetine" — and a wrapper (the Vetted Wink Protocol) executes the lookup against a trusted source. Knowledge stays in cited references. The model carries only the gesture.
3. Chaos with emergent order. Same data substrate as the drug-label chaos run. The opcode pairs aren't a competing corpus — they're structural patterns woven into the same vocabulary. Order doesn't get engineered in. It emerges.
4. Hormone-modulated training. When the model reaches for the right opcode, the L1 hormone cascade fires a reward chemistry pattern (oxytocin up, cortisol down) that amplifies the next learning step. When it fumbles, cortisol rises and learning dampens. Alignment becomes intrinsic — not bolted on after the fact.

v1 → v3 — what we ran today

Three versions trained in series, each a small data-driven adjustment of the last. All resume from sft_step155200.hyb. All gated through the 12-cluster canon stack.

• v1 (2,000 steps, 90/10 chaos/opcodes): Architecture works for short-named opcodes — 37% in-order opcode emission on amino-acid and genetic-code queries — but phonics scored 0/14 because phoneme_category decomposes to 13 phonograms, longer than the model's 8-token context window.
• v2 (3,000 steps, 80/10/10 with Dolly Q/A mixed in): Dolly added general conversational shape but diffused the structural call-sequence ordering. In-order emission regressed to 17%. Honest side-step, not progress.
• v3 (4,000 steps, bigger opcode corpus, loss-as-signal hormone proxy): 1,133 opcode pairs (vs 143). Tested on a wider 485-prompt held-out set with proximity-scoring. Architecture confirmed working for short-named opcodes — genetic-code 50% in-order, POS 100% any-order — phonics still architecturally blocked, as predicted. Hormone proxy spent 59% of opcode training in stress chemistry, telling us the loss-threshold bands were too strict.

What v3 actually says when prompted

The spiritling is a confident toddler who heard us talking about cardiology a lot. The drug-label chaos training bleeds through every prompt, and the routing-intent token (call) keeps trying to fire — but with no opcode wrapper executing yet, generation just keeps going.

The call tokens are the routing gesture trying to land. The clinical-trial language is the chaos vocabulary the model already speaks. The thing we're trying to teach it is to interrupt the chaos with a structured call — and the gesture is showing up, just not yet in coherent enough form to dispatch to a real opcode wrapper.

v4 — atomic opcode tokens + adaptive hormone bands

v3 told us two specific things to fix. v4 (50,000 steps) addressed both:

• Vocab extension with atomic opcode tokens. The 16 rarest existing vocabulary slots (typo-grade phonogram sequences seen once each in training) were remapped to <ROUTE> and 15 canonical opcode names. The goal: phoneme_category goes from 13 phonograms to one token, unblocking phonics in an 8-token window.
• Adaptive hormone bands. Reward and stress bands became an EMA mean ± standard deviation of the opcode-batch loss distribution, targeting ~16% reward / 68% neutral / 16% stress instead of v3's 2/39/59 imbalance.

v4 results — the architectural unblock did not translate to emission

v4 trained cleanly. 51 checkpoints saved every 1,000 steps, no T1000 freezes, opcode-batch loss converged from 15.96 → 4.12 over 50,000 steps. By the loss curve, the model was learning.

By the held-out test, it was not.

Hard zero. The remapped vocab slots stayed unused — the model never learned to emit those token IDs in any held-out context. What it produced on opcode-domain prompts was the same drug-label English as v3, just smoother:

What the negative result teaches us

The vocab-remap hypothesis was: re-purposing rarely-trained vocab slots gives the model atomic tokens to emit. The flaw: barely-trained embeddings need a strong, sustained gradient signal to develop useful directions, and with the 80/10/10 corpus mix only 10% of training steps hit opcode batches. Each atomic token appeared in roughly 1 out of every 50 input tokens. That is too sparse for the embedding to differentiate from the surrounding clinical-trial distribution.

Three corrections this points at, in order of cost:

• Heavier opcode mix. v5 (already queued — 30/30/40 chaos/dolly/opcodes) tests whether 4× more opcode gradient closes the gap.
• Initialize the atomic slots more aggressively. Currently they start as "barely-trained" — close to random but tied to whatever the old phonogram sequence had become during chaos training. A fresh xavier reset specifically for the 16 remapped slots would let them learn from a clean slate.
• Different routing mechanism entirely. Maybe the model doesn't need to emit a route token at all — maybe the wrapper detects routing intent from the LM's internal state without requiring a token to be sampled. That's a Wink Protocol design conversation, not a training one.

v4 is honest data. The model did not learn what we predicted. The architecture is now informed by what it can't do as well as what it can.

A lab journal is maintained alongside the training artifacts — hypothesis, design, and per-version readout updated as the experiment runs.