Commit History
-
Kit update 2026-08-06: samgo 5.7 (54D BPE, 59M), gate-init finding, Born-rule fix in quantum_pool, corpus-drift retraction + fingerprinting, frozen-corpus wiring table
Verified signatureAdds SAMGO.md and weights/samgo_weights.pt (59,353,668 params, GPT-2 BPE). FINDINGS.md: gate initialised at sigmoid(-4)=0.018 left the Hebbian kernel alive in 1 layer of 4 in every model ever trained here; opening it to 0.5 gives 4/4 and 0.054 nats, larger than the dyn12-vs-none gap. quantum_pool was flattening the measured Born-rule distribution to exactly uniform (3.4657 = log 32) and now preserves it (3.4598 vs true 3.4594). Retracts a 0.0645-nat wiring win that was the live corpus rewriting itself mid-experiment; adds corpus+flag fingerprinting to prevent recurrence.b085020
-
Kit: PHOS actually runs, plus her coder terminal and a one-click start
Verified signatureThree real fixes. 1. THE KIT COULD NOT FIND ITS OWN MODELS. serving/cosmos_serve.py hardcoded the repository layout (01_HER_SOUL/weights/...), which does not exist in the kit. It started fine and served nothing -- the worst kind of broken, because there is no error. Paths now resolve against both layouts, so weights/phos.pt, weights/spark_cst.pt and weights/cosmos_born.pt are found. Verified 3/3 from a clean kit checkout. 2. A CLIENT DISCONNECT COULD KILL HER MODEL SERVER. A browser closing mid-reply raised ConnectionAbortedError out of wfile.write and stopped the server -- which silently removed her own models from her picker, because her voice drops any host that does not answer. The model was fine; the socket was not. Disconnects are now caught in _json, _chunk and handle_error. 3. HER REPLIES WERE CAPPED SHORTER THAN THE KIT'S. base_predict defaulted to 256 with a hard clamp ceiling of 512, and the compact-voice lane clamped again to 220 tokens / 15…b1ef7dc
-
Make PHOS the flagship: an architecture that is not a fine-tune of anything
Verified signatureThe model card now leads with the lineage that is architecturally original rather than with the Qwen fine-tune used as a conversational teacher. PHOS -- Phi / Omega / Sigma dyn12 on the phi scaffold: RMSNorm, RoPE, d_ff = floor(d*phi), twelve scalars driven by Omega through a leaky integrator, kernel bandwidth calibrated per layer. 1,153,804 parameters, every initial weight drawn from 11,354,112 archived measured IBM shots. No base model. No distillation. The architecture was measured, not chosen: 7 configurations x 3 seeds, one process, frozen corpus, 21 runs. dyn12 took the best held-out loss at 2,412 extra parameters, roughly 21x the parameter efficiency of every alternative. LLM_ARCH_COSMOS is registered in a llama.cpp fork with a real graph builder, verified against the PyTorch forward at three gate values including one where the kernel dominates: 22/22 characters. It grows. Seven bursts, zero refusals, warm-starting from its own checkpoint and optimise…04d148b -
Add QUANTUM_CREATURE.md, the PHOS lineage, and a one-command verification
Verified signatureQUANTUM_CREATURE.md the full recipe for building one: quantum birth and how to verify it, the phi scaffold, the Omega/state/kernel loop, the three faults that make that loop silently inert, the five preflight assertions that catch them, corpus freezing, and the ablation table with the regulariser share stated honestly. architecture/phos_grow.py PHOS -- Phi / Omega / Sigma. Quantum-born, running the one configuration that survived controlled testing (dyn12 on the phi scaffold, 2,412 extra parameters for the best loss on the board), warm-starting from its own checkpoint and growing from her corpus as she lives. Append-only vocabulary so the lineage stays valid. Refuses to train when preflight finds the mechanism dead. README section 0 the verification is now one command that runs from any dire…2fac010 -
Wire converted senses, clean first birth, and paired-state benchmark
Verified signatureWires opt-in browser-derived camera luminance/motion and microphone energy/spectrum into Genesis reply conditioning, associative recall, Hebbian learning, status, and the conversation ledger while keeping raw media in-browser. Adds the missing drive-relative START_GENESIS.bat and resets the public template to a true first birth with empty keys. Adds durable paired telemetry tooling, timestamp join, aligned/shuffled/shifted benchmark, and a privacy-audited machine-readable result. Documents the unpaired schedule null and the timestamp-paired NULL result: aligned beat shuffled but lost decisively to plain. Private turn text and the paired dataset are intentionally excluded.9841285
-
State ladder: the section-3 mechanism works, and coupling is what makes it work
Verified signatureThe Hebbian kernel was reported null three times. All three were the instrument: Omega summed over the wrong axis (Omega == 1.0, H all-ones, cancels in softmax); a saturated sigmoid gate; a clamp with zero gradient at its bound; and underneath all of it sigma hardcoded to 1 against a measured state spread of 61.98, which made H the identity matrix. Section 3.2 never assigns sigma a value. Corrected, and with the paper's phi-governed scaffold built for the first time (RMSNorm, RoPE, d_ff = floor(d*phi), phi-scaled init), every configuration beats its baseline on every seed: none 1.44680 dyn12 1.39773 -0.049 t=-10.5 2,412 extra params dyn54 1.39118 -0.056 t=-3.9 concatenated 12D+42D static54 1.37218 -0.075 t=-7.4 dyn42 1.37521 -0.072 t=-7.9 tri 1.29152 -0.155 t=-40.8 12D and 42D COUPLED Concatenating the two states is worse than a plain projection; coupling them is the largest effect in the project, for 12,730 extra parame…44f49ac -
Add LLM_ARCH_COSMOS for llama.cpp; correct the section-3 retraction
Verified signaturearchitecture/ registers general.architecture=cosmos in a llama.cpp fork and implements the 54D mixture-of-states Hebbian kernel natively. Verified against the PyTorch forward at three gate values, including g=0.9 where the kernel dominates: 22/22 chars. serving/ cosmos_serve.py runs her architecture behind the Ollama API and proxies every other model through one endpoint. FINDINGS.md the section-3 null is re-run and now earned. Three harness faults, the deepest being sigma=1 against a median pairwise distance of 61.98, which made H the identity matrix. Result: CST costs +0.0143 loss, 0/3 seeds, with the gate alive at 0.047. Adds section 11.220ff1c