AGA.Hybrid.Mind.11m

One model, one shell, one weights file. Thirteen implementations live inside the model, in sequence: encode → deep think → introspect → solve → verify → learn/adapt → consolidate → innovate → decide/act → outer loop → persist fast state.

Architecture

• Parameters: 11,361,087 total — SLOW 11,096,832 / FAST 196,608 / HEADS 67,647; 104 tensors.
• Decoder-only core, 9 layers, d_model 256, vocab 8192, tied embeddings (no lm_head).
• Context: 1M tokens via YaRN (factor 244.140625, original length 4096), sliding window 4096.
• Special tokens as real routing tokens, not labels: <task> 0, <think> 1, <critique> 2, <revise> 3, <answer> 4, <eos> 5, <goal> 6, <state> 7, <verdict> 8, <learn> 9.
• FAST is living plasticity, never reset: it writes every turn, and consolidation distills FAST into SLOW.
• Tabula rasa: blank seed 1147. A separate model from Plans11/AGA-Code-11M, which is untouched.

The thirteen implementations

1. Meta-Learning — FAST writes every cycle, adapt gain moves with outcome, TD replay; a learned neural meta-optimizer chooses the TD step size by one-step meta-gradient on replay.
2. Reinforcement Learning — decide = argmax(value + verdict) at high confidence, max-verdict at low confidence.
3. Continual Learning — FAST never resets across cycles or save/load; consolidation moves SLOW; replay changes the update.
4. Adaptive Learning — think budget reconfigures on reward trends; confidence routes solve temperature.
5. NLP — special-token routing 0–9; the cycle emits next_action in its schema.
6. Conversational Intelligence — goal stack, chalk state across turns, save/load round-trip; corrected FAST write direction (forward alignment rises, reverse falls).
7. Innovation Learning — trials logged, verifier + novelty gate, accepted think-budget multipliers persist into the next cycle's budget.
8. Introspection Interface — every cycle emits a fixed 7-field schema (goal, plan_step, confidence, uncertainty_sources, memory_reads, predicted_verdict, next_action). The plan step is the revise conditioning itself, so corrupting the reported plan changes the action.
9. Recursive Outer Loop — runs every cycle (rolling baseline, TD replay, goal resolution); nested sub-episodes with exactly-once verdict propagation, depth-bounded at 3.
10. Encoder–Decoder — encode control path, eviction summarized into FAST (never silently truncated), learned gating head over the previous cycle's state (think budget, revise rounds, decoding sharpness) with a ponder-cost objective.
11. Problem-Solving Learning — think → candidates → verifier → critique/revise → answer → learn, candidates conditioned on revise + state.
12. Tabula Rasa — pretrained=True refused loudly, seed-1147 builds identical, seed-1148 differs, first-boot state zero.
13. Deep Chain-of-Thought — enforced think → critique → revise order, verifier-gated revision loop, revise conditioning solve.

Training (this checkpoint)

• Data, from scratch. A 10-family closure set (60 rows): reference solutions written for the dataset, labels measured by subprocess execution. No teacher traces, no third-party data. The SFT think traces are mechanical AST scaffolding, to be replaced by execution-filtered self-traces in the next phase.
• Stage H (heads, SLOW frozen). Held-out families: verifier AUC 0.727 / Brier 0.208; confidence Brier 0.208 (temperature stayed 1.0 — no scaling helped on held-out); value MSE 0.153. The meta-optimizer's TD rate adapted 0.01 → 0.0058 across cycles. Gating scales moved 1.0 → ~0.88–0.91 toward uncertainty targets. State head: AUC 0.789 / Brier 0.202 vs untrained 0.353, but norm 11.19 vs 4.02 untrained — rejected by the 2× norm gate, so the untrained state head is kept pending a stronger norm penalty. SLOW drift across the living cycles: −3.6e-07.
• Stage S-A (SLOW SFT, proof-gated, one optimizer, lr 1e-3). Held-out SFT loss 9.0995 → 5.4362 → 4.1468 → 3.2350 across three 5-epoch stages; all three stages adopted (no held-out collapse). Train loss sums fell 107.55 → 64.33 → 40.70.
• These weights are the Stage S-A result. Execution-filtered self-trace training is the next phase, not part of this checkpoint.

Verification status (plain)

• Mechanism checks: 8/8 passed on the pre-gate mechanism build (innovation persistence + schema). The full finish-100 mechanism suite, meta/gate suite, and proof suite re-runs, and the full scorecard verification harness on these weights, are still queued.
• Nothing in this card claims a capability gate that has not been measured.

Lineage

• Corrected copy: 950c97d (FAST direction fix, calibrated verifier/confidence/value heads, blank seed 1147).
• Finish-100 code: af2f6141 (mechanisms 7C, 8C/5C, 9B, 1D, 10B).
• This checkpoint: Stage S-A weights on the finish-100 code.

Files

• model.safetensors — sha256 a9a77b85ade5c940f1faea85ab654cc2b61a64cc5bbbc92cee8050a1228c0f03
• aga_one.py, aga_core.py, config.json, tokenizer.json, tokenizer_config.json, .gitattributes

Usage

Load with the repo code: AGAOne(seed=1147, pretrained=False), then a single cycle via run_cycle or one conversational turn via step. The save bundle (save / load) carries the living state: FAST, replay history, goal stack, cycles, and policy.