Skip to content
Research
Program 01 · Waggle transport · Kea decipher

Machines can move state.
Humans still need receipts.

Waggle tests compact machine coordination. Kea independently verifies, replays, interprets, and constrains every rung before it can matter. The newest experiments preserve Qwen3 capability across fresh processes and show bounded same-host fan-out utility—without pretending the storage bill disappeared.

Experimental ladder

The result is a chain, not a headline.

Every checkpoint keeps its control, its failure state, and the exact claim it does not earn. That makes the progress cumulative without making it slippery.

01 · Deterministic protocol

4/4 exact packet reconstructions

Waggle v0 references content-addressed context instead of restating it. Kea reconstructed all four canonical packets exactly and retained immutable replay and correction lineage.

Boundary: The 61.6% result is a repeated-context byte proxy—not token, cost, latency, quality, or production evidence.

02 · Blinded model consumability

8/8 numeric · 8/8 symbolic · 0/8 no channel

A frozen 24-call study showed that a serialized numeric channel let the model produce all eight proposal artifacts while the no-channel arm safely abstained on all eight.

Boundary: The numeric arm exactly matched a one-byte symbolic control and used more measured tokens. It proves bounded consumability, not a unique latent-language advantage.

03 · Native-state parity

8/8 semantic and byte-identical parity

A Planner serialized Qwen3-14B context state and exited. Fresh source-free Executors restored that state and matched the full-text control exactly across eight frozen tasks.

Boundary: The parity states totaled 333,229,432 bytes. This is native capability preservation, not compression, transport efficiency, or a cross-model language.

04 · Same-host fan-out utility

Task-4 break-even · 30.7% lower task-8 cold wall

One qualified Qwen3 state fed eight fresh native Executors. Every native artifact matched its fresh full-text control, and the end-to-end cold path broke even at task four.

Boundary: The retained state was 279.3 MB versus an 8.5 KB shared prefix, so overall efficiency remains false even though same-host fan-out latency improved.

05 · Founder-runnable workspace

Capture · inspect · fork · replay · release

The passing mechanism became a local workflow: atomically capture one qualified workspace, restore two fresh-process forks, verify both artifacts, preserve replay, release, and reclaim the exact 279,322,905-byte object.

Boundary: A held-out continuation gate later produced 3/3 exact supported continuations and two pre-Executor abstentions. It remains bounded declared-source retrieval, not broad generalization.

From checkpoint to mechanism

The result became a founder-runnable workspace.

Waggle no longer stops at a parity table. A stable L1 library captures one qualified state, Kea inspects it, fresh processes fork it, artifacts are verified, replay is retained, and the exact state object can be released and reclaimed.

  1. 01

    Capture

    Atomic qualified workspace

  2. 02

    Inspect

    Digest + lineage + bounds

  3. 03

    Fork

    Two fresh source-free processes

  4. 04

    Verify

    Semantic + byte-identical artifacts

  5. 05

    Replay

    Immutable correction lineage

  6. 06

    Reclaim

    279,322,905 bytes released

8/8 exact parity

Every native-state artifact matched the full-text control across frozen tasks.

3/3 + 2 abstentions

All supported held-out continuations completed; two unsupported requests stopped before an Executor ran.

Authority still unsigned

The L2 authority packet is prepared but unsigned. L1 transport and Kea interpretation cannot authorize an external act.

Kea comes first

Decipher before traffic.

Kea is not a friendly summary pasted on top of an opaque channel. It owns the independent record, qualification, replay, abstention, and capacity boundaries.

Authority remains outside both Waggle and Kea.

Render

Turn a registered machine message into a qualified human-facing interpretation.

Log + replay

Preserve raw payload identity, causal lineage, interpretations, and corrections.

Watch

Reject unknown codecs, broken hashes, oversize payloads, causal gaps, and drift.

Decompose

Expose components and competing interpretations instead of hiding uncertainty.

Budget

Bound message capacity and make undecodable or hidden capacity visible.

Claim boundary

What the evidence still does not say.

The most interesting result is allowed to stay narrow. A strong control is not an inconvenience; it is the difference between research and branding.

  • No universal or cross-model latent language has been demonstrated.
  • No token, dollar, storage, network, energy, or overall-efficiency win is claimed.
  • No production traffic, customer result, revenue, or commercial adoption is claimed.
  • Kea never grants authority; only a governed capability envelope may authorize an action.
  • The v0.3 paper is a publication candidate, not a peer-reviewed result.
Publication candidate v0.3

Read the specification, controls, and kill rules.