Metabolic Intelligence

mei-arg-01 Why this is the future

Teaching diagram Settled law + pack labels

Read top to bottom. Energy binds by settled law, not by a slogan. The transformer corridor can keep improving and still plateau as a deployment future, because uncapped import stops buying the next useful answer. Metabolic Intelligence obtains the superior answer under the metabolic envelope. Not a cheaper answer. Mixture of Limits chooses the gear. Two plants, edge and resource-optimized central, remove the hyperscale water and power prerequisite by design. That is not a liter count and not a board measured_j. Klere is the product home. Research owns the class.

Static teaching figure. Study section “Why this is the future” and Appendix F. Soft-ref: board_synth_claimed=false.

mei-diag-01 Budget-in → best-answer-out (dual scale)

Teaching diagram Unmetered

Look at both ends of the same class. On the left, the plant is a coin cell. On the right, the plant is a grid-following import cap. The middle box is the rule: the envelope B is the physical condition of the superior answer. Dedicated water, power, and hyperscale cooling plants are not required by design at these envelopes. That is a design consequence, not a liters-saved meter.

Manim clip plus static SVG. Soft-ref: board_synth_claimed=false. Study: §1–§2 and Appendix C.

mei-diag-02 J(m|q) obtain router

Calculation + looked-up Estimated

Watch the cascade left to right. Lookup and Formula close first when grammar covers the need. Solver settles when physics needs it. Model opens last. Guess is residual. The rule under Mixture of Limits is m* = argmin J(m|q). Ledger Wh/answer bands on this path are estimates, not board measured_j.

Lookup → Formula / Calculate → Solver → Model LAST → Guess
m* = argmin_m J(m|q) = E_m + λ · ½ · S_q² · var_m(t)

Standing obtain rule. Study §4. Price the obtain zone, not only which LLM answers.

mei-tab-01 Digital enzyme vs int8 MLP (instruction counts)

Emulator-measured Pack MCU

Compare recognition cost under a matched Cortex-M4 decision setting. Enzymes with early exit take fewer instructions than a dense int8 MLP. On SmellNet sequential learning (50 foods, five groups, no revisit), enzymes keep accuracy across groups while the MLP collapses. Instruction counts are not a wattmeter. Do not upgrade them to board joules.

Chicken-spoilage Anwar numbers stay internal (no licence). Study §5.

mei-tab-02 Tag metabolic schedule (~10 µW class)

Datasheet + schedule model

The heater is not always the dominant cost. On the metabolic schedule with T/RH wake and a 4 h cap, sleep current dominates. Total average power sits near 10 µW in the pack model. Capacity years are a model from CR2032 datasheet capacity, not shelf-measured years.

From pack tag_budget_2026_results.json. Study §6.

mei-chart-01 Energy vs speed (pack chart)

Curves / measured-kernel model Estimated

Speed-matching and Flash-first are test-time levers under a live budget. They are not a new pretrain. Wh/answer stays Estimated until Metered.

mei-chart-02 Grid day mechanism (pack chart)

Plant model Mechanism test

Cluster power tracks a simulated grid day under several policies. Use this with OpenADR/DSX protocol evidence. Do not treat the plant model as a field meter.

mei-chart-03 Grid service scenario A (notice, closed loop)

Real protocol + plant model

Protocols (OpenADR 3.1 VTN, DSX Flex MQTT) are real. Campus watts are the plant model. With node sleep available, the closed scenario set shows zero time over cap in the pack headline.

mei-chart-04 Stress day (pack chart)

Plant model

Stress day shows how levers stack when the grid asks harder. Still a plant model. Pair with scenario A and the D/E sleep contrast below.

mei-chart-05 Scenario D vs E: 90–95% cut, sleep on/off

Real protocol + plant model

Same deep cut. Left (D) keeps node sleep available. Right (E) removes sleep. The contrast teaches why sleep is a first-class lever, not an afterthought.

Pack assets under datacenter/service/tests/. Do not sell “100% served through 90% cut” as a universal field result.

mei-chart-06 Vera Rubin vs GB300 (field peer curve)

External field / pack curve Estimated composite

This chart orients campus energy-per-answer talk against published peer hardware curves. It is an envelope actuator context chart, not an OpenIE board meter.

mei-chart-07 Consumer regime detection (synthesis)

Synthesis / schedule model

Consumer speed tiers feed the metabolic gateway. Regime detection tells which lever to pull first under a live budget. This is software-path evidence, not a GPU wattmeter.

mei-diag-03 Grid lever order (least service lost per MW)

Real protocol demos Plant model

Pull levers left to right. Pause first. Then slow speed. Then route to a smaller model. Then lower effort. Then battery. Shed last. AEMA has no published protocol spec yet (1 Oct 2026 radar). Do not sell AEMA latency numbers.

OpenADR 3.1 + DSX Flex–shaped service. Manim motion for this figure is backlog. Study §8.

mei-map-01 Hardware dual-map

Operational map Unmetered

The class runs today on GPUs and MCUs. It points at neuromorphic, analog / in-memory, photonic, and reversible / adiabatic substrates next. Map status is not a meter reading.

Teaching map from study Appendix B. Soft-ref path: board_synth_claimed=false.

Measurement note

No board package joules are claimed here. Evidence classes: real protocol / emulator / datasheet / curves / plant model. Read the study claim ledger (Appendix E) for proven pack vs roadmap vs estimate. Pack third-party notices: PACK_THIRD_PARTY_NOTICES.md.