Mixture of Limits
Floors where more bits stop buying outcomes. Try lookup, then formula, then solver. Open the model last.
This page shows the P0 figure shells for the Mixture of Limits study. Claims stay ledger-safe with the paper.
Analytical OpCounter and catalog estimated_j are not board power.
Soft-ref path: Landauer is a labeled thermodynamic lower-bound estimate; measured_j=None unless Metered; board_synth_claimed=false.
Soft-ref mol prove (43 VERIFIED) is software constructive existence, not FPGA or board package energy.
Figures on this page are Unmetered or Estimated.
Shannon, Landauer, Howard, Kolmogorov and related floors already name those bounds as settled theory. The remaining bottleneck is applied math and hardware embodiment.
Study §8.4-§8.14: SOTA at the plateau under Mixture of Limits—once VoI is zero past the floor, optimize energy/compute (Lookup → Formula → Solver → Model LAST). Latents, Logical Intelligence (EBM + LLM + latent), World Labs, test-time, SSM/MoE efficiency, and plateau techniques (YOCO, DeepSeek-V4.1-Flash, HySparse2 / MiMo, Qwen, Kimi, Sarvam, …) are implementation options for plateau spend; floors still bind. Field KV bytes and API prices ≠ soft-ref measured_j.
Study §11: dual-phase front door (bounded TinyML/transducer → typed AST, then cascade). Perception is a front gear; Model LAST stays last. Measurement tiers 0/1/2; only Metered populates measured_j; board_synth_claimed=false.
Read the study at /papers/gates/#part-ii-the-cascade-mixture-of-limits. Companion commit figures: /living/fpga-sim/ (NI Stage B/C; energy rules unchanged).
mol-inf-01 Historical compression lineage
Brahe enumerated. Kepler compressed. Newton closed forms. Shannon priced bits. Landauer priced irreversible erasure (labeled estimate here). Mixture of Limits navigates that lineage as CI floors.
board_synth_claimed=false. No meter reading is reported here.mol-diag-01 How the cascade works
Supports the Mixture of Limits cascade claim in /papers/gates/#part-ii-the-cascade-mixture-of-limits §2 and §4.
Lookup → Formula → Solver/settle → Model LAST
When the grammar covers the request, the system stops before opening the model.
board_synth_claimed=false.mol-diag-02 Named floors
Supports the named-floors table in /papers/gates/#part-ii-the-cascade-mixture-of-limits (VoI, grammar, energy, certificate, settle-refuse).
VoI stop. Grammar coverage. Energy (Landauer estimate). efa_certificate. settle_refuse. primitive_gap.
measured_j=None. board_synth_claimed=false. No meter reading is reported here.mol-diag-03 Commit or refuse, then issue a receipt
Supports close shape shared with Notational Intelligence; navigation side in /papers/gates/#part-ii-the-cascade-mixture-of-limits.
propose → certify → commit|refuse → receipt claim.* / act.* / close.* / receipt.* ModelGenerated is not Deterministic evidence measured_j only if Metered; else None refuse = lawful success + receipt
measured_j only when Metered. board_synth_claimed=false. No board joules are asserted.mol-tab-01 How Mixture of Limits differs from MoE
Short contrast from paper §9. Field systems optimize which neural generator runs. Mixture of Limits differs from Mixture-of-Experts (MoE) in the decision it makes: it asks whether generation should run at all. The full model-to-model SOTA table and verified cites live in the study. No board joules are asserted here.
estimated_j is labeled. Soft-ref keeps measured_j=None. board_synth_claimed=false. No meter reading is reported here. See the table below and /papers/gates/#part-ii-the-cascade-mixture-of-limits §9.| Axis | MoE / model-to-model field | Mixture of Limits |
|---|---|---|
| Where the mixture lives | Inside the model (experts, draft/target, strong/weak LLMs) | Outside generation (named floors + cascade gears) |
| Routing object | Which model answers | Which tier closes: Lookup / Formula / Solver / Model |
| Refuse | Rare; usually escalate to a larger model | First-class: VoI, grammar, settle, certificate, primitive_gap, and related floors |
| Energy labels | Latency/cost proxies; some papers meter GPU energy for LLM serving | Receipts: labeled estimated_j; soft-ref measured_j=None; board_synth_claimed=false |
| Success metric | Tokens/s, cost@quality, speculative speedup | Certified commit + typed ReplayClass + receipt |
Analytical and catalog estimates are not board power. No meter reading is reported here. See /papers/gates/#part-ii-the-cascade-mixture-of-limits §9.
mol-tab-02 How proof status maps to claims
Soft-ref mol prove stamps VERIFIED onto paper claim.* ids (see study §6). Soft-ref-only residuals stay on the roadmap. Do not escalate them to measured results.
- Proven path: Unmetered (plus Estimated where μ, Landauer, or fuel appear);
board_synth_claimed=false - Soft-ref only (pointer): live RAPL/NVML as default close; live wgpu presence; Ferric / BMI loop; klere-vm FPGA pJ; full Periodic Stack 258 + μ calib; trained Model leaf. See §6.3
- Soft-ref v0.1 prove:
measured_j=None; software constructive existence, not board joules
mol-note-01 Energy measurement labels
OpenIE measurement grammar (matches paper §6 / §9):
| Label | Meaning |
|---|---|
| Unmetered | Software-ref / soft-ref path; measured_j=None; estimate may be labeled |
| Estimated | Catalog / Landauer / fuel / μ surrogate with explicit EstimateKind or mu_source=catalog |
| Metered | Optional live probe (energy-meter) returned numbers. Not required for prove. |
- Analytical OpCounter and catalog estimates are not board power
- Landauer is a labeled thermodynamic lower-bound estimate
- Soft-ref path does not stamp RAPL/NVML/package meters without a Metered probe
board_synth_claimed=false- Default prove close: Unmetered (plus Estimated where labeled)