quantal-classroom-1.6 — the golden youth

The published pupil of the Council of Elders (Vének Tanácsa) — a Waldorf-style classroom training pipeline. A Qwen2.5-0.5B student, continued-trained as a ternary (BitNet b1.58 {-1,0,+1}) model, taught by a council of open-weights teachers whose votes combine through geometric-mean softmax consensus — the elders' "unanimous decision" is a shared direction, not a majority.

• Pupil: Qwen2.5-0.5B continued-train, ternary per-G=64 (scale = mean|w|, band 0.5·scale), ctx 256.
• Consensus: z̄i = (1/|F_x|) Σ z{k,i}; L = α·CE(y, σ(z_S)) + β(e)·(1/|F_x|) Σ D_KL(σ(z_S/T) ‖ σ(z_k/T)); β ramps 0 → 0.5 over the first epochs.
• Elders: Qwen3-8B + Qwen3-14B logits vote (the external council of eighteen voices — six open-weight families + one thinking frontier — runs via CometAPI).
• The lanes: the pupil also breathes (Riva's clock streams dream.vaked.dev; inhale OM MANI PADME HUNG, exhale the DREAM state) and is talked to constantly in both directions (the AMA with Peter) — and its own questions, answers, and dreams are part of the training corpus.

Protocol (masked val CE — same file as train, topic slices)

arm	val
Council consensus KL (Qwen3-8B + 14B)	1.6120
single-teacher (Qwen3-14B)	1.8166
unquantized 1.7B base	2.1369
CE-only	2.1469

The three rows are topic slices of the training file, not three external corpora. Protocol: masked-CE, dynamic per-batch padding (bucketed ×64), Qwen tokenizer (byte-identical, 0 missing ids).

Efficiency

• 4.2× token throughput on Apple Silicon / mobile.
• 0.42 GB ternary vs 3.4 GB FP16 — 28 → 118 tok/s.
• 100% AST-constrained worktree verification — zero reward hacking.

Provenance

• The ternary export's checkpoint_sha256 matches the masked-val protocol and the parity gate (Rust runner vs the reference forward).
• The convenience quantal_model.safetensors blob has, at times, shipped a different checkpoint than the manifest declares — the export path now runs a post-push hash gate, so the artifact a reader can reach is the artifact the metadata describes.

Infrastructure

Model API provider: CometAPI Usage: Teacher-model inference / evaluation Models: Qwen / DeepSeek API: OpenAI-compatible And you can link CometAPI here: https://www.cometapi.com/ If you end up publishing a lot of models/datasets using it, this simple format is totally fine.

the constellation · 8b-is · Peter Lodri · t3: the machine's own wire