CLAIMS_AUDIT.md

Claims Audit — bio_inspired_nanochat

Bead: bio_inspired_nanochat-8j9.1 · Owner: TurquoiseFinch · Last updated: 2026-06-10

Maps every public/marketing claim (README, planning docs) to a verified implementation status with file:line evidence and the bead that closes each gap. This is the reference all other "Truth & Integrity" tasks (8j9.3 README rewrite, 8j9.6 parameter census) act on.

Status legend

• SOLID — fully implemented and on the live model path.
• PARTIAL — exists but limited, not wired into the live path, or test-only.
• ASPIRATIONAL — claimed but absent, dead, or only a research/roadmap item.

Methodology: a draft inventory was generated by a search agent, then every contested status was re-verified by hand against the source (the agent mis-rated several — e.g. it claimed tune_bio_params.py "does not exist" and rated the "bank-account economy" and "three kernel backends" as SOLID; all corrected below). Line numbers are from the working tree at the date above and will drift.

---

Presynaptic biophysics

Claim	Status	Evidence	Closing bead
Hill-equation release Syt(C)=C/(C+Kd), Syt1/Syt7 mix (0.7/0.3) + Doc2	SOLID	synaptic.py:491 _faithful_release_prob (exact 0.7·fast+0.3·slow+Doc2); verified numerically to 6 digits	— (closed 2gjl)
The live attention path runs the faithful release (not the legacy sigmoid)	SOLID	synaptic.py:1407 standard path calls release_canonical; legacy release() only on the off-by-default flex path	— (closed ukxt); flex on s3w9
Calcium integrator with buffer ODE; tau_c decay	SOLID	synaptic.py:513 release_canonical (BUF ODE active); tau_c unified as exp(-1/τ) across all paths	— (closed x6z4)
RRP depletion / "fatigue & boredom" frequency penalty	SOLID	release_canonical scatter updates RRP; init_rrp, priming, endo DELAY queue all present	—
Endocytosis delay queue (ring buffer)	SOLID	synaptic.py build_presyn_state DELAY=[...]; endo_delay, rec_rate consumed in the state update	—
Stochastic vesicle release (Gumbel-Sigmoid / straight-through / normal-reparam STE)	SOLID	synaptic.py:58 _sample_binomial_counts (3 modes); exercised in release_canonical when stochastic_train_frac>0	—
Septin-like distance barrier	PARTIAL	applied at the attention logit level (global q_pos−k_pos), not a local windowed per-vesicle inhibition as the prose implies	8j9.3 (reframe)

Postsynaptic / Hebbian learning

Claim	Status	Evidence	Closing bead
Online Hebbian learning during training	SOLID	synaptic.py:233 plasticity_during_training=True; :1155 run_plasticity runs with grad enabled; deferred autograd-safe writes :1074	— (closed vg9.2)
Low-rank eligibility traces (rank-R U/V)	PARTIAL	u_buf/v_buf buffers + outer-product consolidation exist, but the rank-R structure is degenerate (mean-broadcast, not genuine rank-R)	vg9.9
CaMKII/PP1 consolidation gate	SOLID (opt-in)	default gate is g = sigmoid(camkii−0.5)−0.3 (PP1 not in it); bistable_latch (sax.2) switches it to sigmoid(β·(CaMKII−PP1)) — PP1 IN the gate	— (closed sax.2)
Bistable CaMKII/PP1 latch (hysteresis / self-excitation)	SOLID (opt-in)	sax.2: Lisman switch — CaMKII Hill autophosphorylation + mutual PP1 cross-inhibition over a basal floor; demonstrated bistability/hysteresis/erasure in tests/test_bistable_latch.py. Behind bistable_latch (default-off)	— (closed sax.2); cusp variant 0642.2.x
BDNF metaplasticity (activity-dependent LR)	SOLID	synaptic.py:891+ BDNF accumulator; consolidate modulates post_slow_lr by (1+γ·BDNF)	—

Structural plasticity / MoE

Claim	Status	Evidence	Closing bead
MoE split/merge/reset lifecycle	SOLID	synaptic_splitmerge.py SplitMergeController; optimizer-moment reset across all optimizers (closed vg9.3)	—
Health = utilization × energy drives lifecycle	SOLID	synaptic_splitmerge.py:393 _health = util·energy; gates split/merge/reset	—
Per-expert genome Xi (4-D) → phenotype decoder	SOLID	synaptic.py:248 xi_dim=4; phenotype decoder maps Xi → kinetic biases	—
NeuroScore evolutionary credit assignment drives split/merge	SOLID (opt-in)	de5l: NeuroScore.step publishes a per-expert composite fitness onto each MoE (last_neuroscore); SplitMergeController._health blends it into every split/merge/reset decision behind use_neuroscore (default-off)	— (closed de5l)
Router embeddings + lateral-inhibition update	SOLID	fkkc: SynapticMoE.forward runs a per-forward similarity-weighted repulsion that pushes similar router embeddings apart to drive specialization (gated by router_contrastive_lr/_push). The earlier block was wired but ineffective — a diagonal-only co-occurrence matrix zeroed its pull term; it did not reduce pairwise similarity	— (closed fkkc)
"Bank-account economy": taxation / income / bankruptcy / IPO	ASPIRATIONAL	zero code references to bankruptcy/IPO/bank-account/tax; only generic energy metabolism (energy_cost_*) exists	8j9.3 (reframe to "energy metabolism")

Kernels & performance

Claim	Status	Evidence	Closing bead
Three kernel backends (Triton GPU / Rust CPU / Python ref)	PARTIAL	files exist (kernels/, rust_src/src/presyn.rs) but only a default-False native_presyn flag (synaptic.py:257); no live dispatch — the live release_canonical is pure PyTorch; kernel tests skip (rustbpe unbuilt)	jyb.1/jyb.2/jyb.6
Kernels target the documented (faithful) equations	PARTIAL	forward() (the kernel reference) is faithful, but kernels were written against it, not yet the canonical top-k path	8j9.4, jyb.*
"90%+ GPU utilization (dual 4090)"	ASPIRATIONAL	a stated roadmap target; no benches/guardrails in code	j9i, r2d
FlexAttention O(N) memory	PARTIAL	synaptic.py flex path exists but prefill-only (explicitly no KV-cache decode); not usable for efficient generation	s3w9, zsi
Logit softcap (15·tanh, parity with vanilla)	SOLID	gpt_synaptic.py:44 logit_softcap=15.0; :286 sc*tanh(logits/sc) on both paths	— (closed vg9.1)

Config, tuning & measurement

Claim	Status	Evidence	Closing bead
CMA-ES hyperparameter optimization of bio params	SOLID	scripts/tune_bio_params.py exists (59 KB, CMA-ES + TOP10_PARAM_SPECS); tau_c spec corrected for the canonical regime (x6z4)	—
"48-parameter genome"	ASPIRATIONAL (claim drift)	SynapticConfig now has 70 fields — 8j9.5 pruned the 6 dead ones, so all 70 are live. The learned "genome" is 4-D (xi_dim); only 10 are CMA-ES-tuned. Machine-verified in docs/parameter_census.md	— (closed 8j9.6/8j9.5)
Feature toggles for clean ablation	SOLID	enable_presyn/enable_hebbian/enable_metabolism/use_flex_attention; presyn early-returns when disabled	—

---

Most significant remaining gaps (priority order)

1. "Bank-account economy / bankruptcy / IPO" has no code. → reframe to "energy metabolism + health-based lifecycle" in 8j9.3.
2. Three kernel backends / 90% util — kernels exist as files but aren't on the live path. → jyb.*, perf beads.

Resolved since the original audit

• ✅ NeuroScore now drives the lifecycle (de5l): published onto each MoE and blended into split/merge/reset health behind use_neuroscore.
• ✅ Router lateral-inhibition fixed (fkkc): the wired-but-ineffective update was replaced with an effective similarity-weighted repulsion that demonstrably spreads router embeddings.
• ✅ "48-parameter genome" reconciled (8j9.6 census + 8j9.5 prune, then sax.2 added the latch knobs): 82 live config fields, 0 dead; machine-verified census committed.
• ✅ Bistable CaMKII/PP1 latch implemented (sax.2): real Lisman switch with hysteresis + PP1 in the gate, behind bistable_latch (default-off), demonstrated in tests.

Recently CLOSED gaps (this audit records the wins)

• ✅ Online Hebbian learning now runs during training (vg9.2) — was inert behind not torch.is_grad_enabled().
• ✅ The live attention path now runs the faithful Hill release; README equations match (2gjl/ukxt) — was the legacy sigmoid while docs described Hill.
• ✅ tau_c unified as an exp time-constant with a sensible default (x6z4) — was overloaded (raw vs exp) and silently weakening calcium memory ~7×.
• ✅ Logit softcap parity (vg9.1); optimizer-moment reset across all optimizers on split/merge (vg9.3); bounded log-release attention bias (vg9.5).