Malia

The Machine Kingdom

Research
INSTINCT lab

LEXICON·ORACLE
Proposed architecture

Machines that say
only what matters.

Not a mystical “AI language.” A hypothesis: robots stop streaming raw sensor data and learn compact messages that preserve exactly what another robot needs in order to act.

(01) — The model, in four lines

(EQ. 01) — Encode

zᵢ = f_φᵢ(oᵢ, eᵢ, q)

Robot i’s private encoder maps what it observes (oᵢ), the body it observes with (eᵢ) and the mission context (q) into a latent state zᵢ.

(EQ. 02) — Transmit

mᵢ = E_θ(zᵢ, Σᵢ, q)

LEXICON turns that latent — together with the robot’s uncertainty Σᵢ — into a compact message mᵢ. A claim, not a sensor stream.

(EQ. 03) — Fuse

p(z_world | m₁:ₙ) ∝ p(z_world) ∏ p(mᵢ | z_world)^wᵢ

ORACLE updates the shared belief over the world state. The weights wᵢ discount evidence that is miscalibrated, stale or already counted.

(EQ. 04) — Learn

L = L_task + λ·B(m) + μ·L_meaning + ν·L_uncertainty

One objective trains the whole stack: do the mission, pay for every bit, keep meaning across bodies, stay calibrated.

(02) — ORACLE — belief fusion

One world,
believed jointly.

Each robot sees a fragment. ORACLE maintains a shared posterior over the world state and lets every message move it — by exactly as much as it deserves.

Evidence is weighted

The exponent wᵢ is the epistemology. Messages from miscalibrated sensors, stale observations, or sources correlated with evidence already counted are down-weighted before they move the posterior.

Disagreement survives

When robots disagree, uncertainty rises instead of being averaged away. Conflict triggers another measurement and a flag for the human operator — not a forced consensus.

Belief drives action

The shared posterior is what COLONY tasks, INSTINCT plans and AXIOM authorizes against — one picture of the world, with its doubts attached.

(03) — The training objective

Bandwidth is
a price.

EQ. 04 trains encoders, messages and beliefs end to end. Each term exists for a reason.

  1. 01L_taskMission performance. Messages exist to complete the task. Elegance that does not serve the mission scores zero.
  2. 02λ · B(m)Priced bandwidth. Transmission is a cost in the objective itself — every bit is paid for, so only necessary bits survive training.
  3. 03μ · L_meaningMeaning preservation. A message composed by one body must stay interpretable by a different one — the vocabulary is shared, not private.
  4. 04ν · L_uncertaintyCalibrated uncertainty. Confidence claims are trained to match reality, so ORACLE can trust the Σᵢ it is sent.

(04) — A worked example

The bridge.

One incident, four kingdoms. How the stack is intended to behave when the evidence is incomplete.

  1. 01VISTA / AERFlies above the bridge and observes a displaced deck section, encoding geometry and visual confidence into its local latent state.
  2. 02CLING / ADAPTMoves along an accessible girder and detects surface deformation VISTA cannot resolve from above.
  3. 03COIL / ADAPTEnters a narrow void and produces acoustic, depth and structural observations unavailable to either of the first two robots.
  4. 04LEXICONTransmits compact statements of changed geometry + location + confidence — not raw sensor streams to every machine.
  5. 05ORACLEFuses the observations into a posterior belief that one load path may be compromised while another remains uncertain.
  6. 06ORACLEBecause the robots disagree in one region, uncertainty increases instead of being averaged away; a second measurement is requested and the area is flagged for the human engineer.
  7. 07AXIOMPrevents TUSK from entering the affected zone until the mission rule set is satisfied and the authorized operator approves the next action.

(05) — The metric that matters

Task success
per transmitted bit.

The goal is a Pareto improvement — higher mission performance and fewer bits on the wire — not a trade between them. Compression that costs the mission is failure, not savings.

A message earns its bandwidth only when it changes what the fleet believes or does. Everything else is noise the objective is trained to delete.

(06) — A note on status

LEXICON·ORACLE is a proposed research architecture, synthesized from published directions — emergent communication, the information bottleneck, belief-aware autonomy. It is not a demonstrated MALIA product, and nothing here is a measured result.

What happens to a belief before it becomes an action — AXIOM / Safety →