0-bro
Menu

Research ledger

The claim, the command, and the limit.

This page separates the model we intend to test from the evidence already collected. The architecture section is a specification, not a result. Each evidence card names the exact script, default setup, result, and the thing it does not establish.

Revision 6 // design under test

00The current model, stage by stage

Mind-1B is not six speculative subsystems switched on at once. It is one recurrent spine with a promotion ladder. The first trainable reference is deliberately plain: a bilateral GDN2 core in bf16, fast and slow state/head banks, and a deterministic four-slot workspace. Each later row changes one controlled variable.

Promotion order; a stage cannot borrow a claim from a later stage
StageComponent enteringRequired comparisonStatus
M3a18-block bilateral GDN2, fast/slow clocks, deterministic four-slot workspace, bf16Single-stream GDN2, 3:1 GDN2/attention hybrid, and Transformer++ at matched parameters and active computenext run
M3bProgressive ternary QAT with shadow weightsThe winning M3a checkpoint before quantizationblocked by M3a
M3cCausal workspace auditScrambled, delayed, bypassed, and capacity-matched workspace controlsblocked by M3a
M4aPrecision Workspace: belief mean and diagonal uncertainty per slot; predictive compute budgetDeterministic workspace at equal parameters and active computespecified, untrained
M5aCausal event-boundary head plus bounded entorhinal/DG/CA3/CA1-style episodic pathFixed chunks, turn-only boundaries, and flat retrieval at the same byte budgetspecified, untrained
M5bScheduled replay with frozen base weights; plastic sleep and REM-like integration only as later ablationsNo replay, matched online replay, and replay-only controlsspecified, untrained
M6Scale only the components that passedPublished checkpoint, matched control suite, quality, latency, memory, calibration, and failure tableblocked by M3–M5

What “bilateral” means

It does not mean two independent models or a left-brain/right-brain personality split. A shared GDN2 spine maintains two timescales: a fast precision stream for local ordered updates and a slower context stream for integration. Both can propose a write; neither owns the workspace. A matched single-stream model decides whether the extra structure earns its parameters.

What the Precision Workspace adds

M3a slots are ordinary deterministic vectors. M4a may give each slot a belief mean and diagonal variance, then update it with a bounded Kalman-style gain. Reliable evidence can revise an uncertain slot; noisy evidence should be discounted. The module ships only if proper scoring, calibration, state tracking, and language loss improve—or writes/compute fall at matched loss. It is inspired by Bayesian filtering in Kalman Linear Attention; it does not replace GDN2 as the sequence mixer.

What event memory and sleep add

M5a closes an episode only after a causal boundary detector sees the event's final token. A bounded record carries a temporal key, compressed summary, terminal belief and uncertainty, source span, and confidence. The design lesson comes from event-segmented hierarchical memory. M5b then tests scheduled replay into a protected candidate memory before permitting base-weight plasticity. Wake serving never silently rewrites the deployed base model.

Limit. Revision 6 is an executable research specification, not a trained checkpoint. GDN2 itself has published evidence; the bilateral clocks, workspace, event path, and sleep lifecycle do not yet have language-model results in this repository.

Read the full Revision 6 paper / Gated DeltaNet-2 paper / GDN2 reference code / workspace interpretability study

M0 // character-model toy

01Ternary QAT

0.072 gap
final loss gap at 4,000 steps: fp32 finishes at 0.015 and ternary QAT at 0.086, with 35% zero weights.

Run: python3 mind-1b/proto/toy_ternary_lm.py --steps 4000

Limit. This trains on the text of its own source file. It establishes that this quantization recipe can converge in a tiny deterministic setting; it does not establish language-model parity at useful scale.

Read the script on GitHub

M3 scout // operation-count check

02Surprise gate

7.7%
of dense MACs on a generated 512-dimensional, 2,000-step stream at redundancy 0.95 and threshold 0. The relative RMS error is 4.14e-6, not mathematically zero.

Run: python3 mind-1b/proto/surprise_gating.py

Default result sweep
RedundancyThresholdMAC ratioRelative RMS error
0.850.0016.7%0.000007
0.850.1015.5%0.014681
0.850.4012.2%0.115351
0.950.007.7%0.000004
0.950.107.1%0.016392
0.950.405.2%0.121315

Limit. This counts operations for an exact delta-coded linear matvec on a generated stream. It does not measure language-model quality, kernel latency, or training savings.

Read the script on GitHub / Read the original lab note

retained scout // synthetic routing rule

03Phase-routing curiosity

1.90/ 8
active regions per token in the default eight-domain demo: 23.75% of dense region compute, with zero trainable router parameters and 12 settling steps.

Run: python3 mind-1b/proto/phase_binding.py

Default routing output
Input kindActive regionsRelative error
Simple, one domain1.650.257
Composite, two domains2.370.466
Dense all-region baseline8.00 (4.2× compute)0.914

Limit. The script seeds each region's prototype and expert matrix from the synthetic task's ground-truth centers and rules. It tests a routing rule, not learned specialization or a fair trained-router comparison. Phase routing is not in M3a and earns no flagship claim; if revisited, it must beat a learned router at equal active compute.

Read the script on GitHub

M2 // synthetic continual-memory check

04Sleep cleanup

81.4%
mean sign recall after eight cycles with cleanup, versus 34.4% for the naive raw-magnitude retention control.

Run: python3 mind-1b/proto/sleep_cleanup.py

20,000 synthetic synapses; eight tasks; 400 signal synapses per task; seed 0
CycleWith cleanupNaive control
189.2%49.8%
288.5%44.1%
485.2%39.1%
682.3%35.6%
881.4%34.4%

Limit. Neither arm is a language model. The control is a naive raw-magnitude retention strategy, not an otherwise-identical model "without sleep." This result earns a larger experiment; it does not solve catastrophic forgetting.

Read the script on GitHub / Read the original lab note

Configuration calculator

05Memory budget

242 MB
packed non-embedding ternary core: 969.02M weights at two bits each. This is one component of the 1.035B planning configuration.

Run: python3 mind-1b/calc/params.py and python3 mind-1b/calc/memory_budget.py

Current default configuration
MeasureValueScope
Total parameters1.035BTied embeddings, 18 GDN2 blocks, workspace, router estimate, and MTP transforms
Non-embedding parameters969.02MArithmetic planning count
Training estimate14.83 GB/GPUQAT planning envelope, batch 8, sequence 2,048, checkpointing; excludes runtime overhead
Inference minimum0.37 GBPacked core plus fp16 embeddings; excludes unmeasured GDN2 state and future episodic memory

Limit. These are arithmetic planning estimates, not measured device memory or throughput. The calculator currently sets recurrent-state and hippocampal-memory bytes to zero and warns about both, so 0.37 GB is a floor—not a deployment budget.

Parameter calculator / Memory calculator / Design specification

What earns the next claim

06The gates still open

The next material claim needs a small bf16 GDN2 model that reproduces recurrence correctly, followed by the 100M M3a control ladder: single-stream GDN2, bilateral GDN2, a 3:1 GDN2/attention hybrid, and Transformer++. The first H200 reservation is a ten-hour iteration window for environment validation and smoke runs, not a promise to finish pretraining. Held-out language quality, peak memory, tokens per second, stability, matched active compute, and failed ablations all ship together.

Phase routing, sparse MoE, MatFormer widths, multi-token prediction, custom kernels, and additional brain analogies remain outside the reference model until a simpler stage exposes a measured need. That is how the roadmap stays a living status board instead of a list of features.