SPECULATIVE INSTRUMENT / NO MODEL INFERENCE

A parliament,
not a personality.

What if the future of open-weight intelligence is not one aligned supermodel—but a federation of capable, fallible, constitutionally different minds?

Enter the chamber Read the K3 inference Enter the 896-star field lab Cross the AgentTool airlock Weave the playable council

INSPIRED BY

Kimi K3’s sparse experts, multi-teacher distillation, long-horizon agents, and explicit harness sensitivity.

16 / 896ROUTED / TOKEN / LAYER
+2 SHARED

K3 demonstrates: specialized policies can be consolidated.

This project asks: what must remain deliberately unconsolidated?

THE PREMISE

Intelligence can be mixed.
Authority must be divided.

A frontier model’s weights contain capability. Its system prompt, tool harness, guard, license, and operator determine how that capability enters the world.

So the unit of governance is not the model. It is the whole governed assemblage—versioned, bounded, and inspectable.

01Route for complementary error.
02Separate fact, value, and permission.
03Preserve the losing objection.

01 / Council chamber

Make disagreement operational.

The council does not average personalities. It separates what appears true, what should matter, and what may be done.

QUESTION / weights

Should a lab release an open-weight scientific agent with unrestricted tools?

toy stakes 86toy reversibility 31toy value conflict 69

Constitutional seats5 / 5 active

Shared lineage42%Data, architecture, evaluator, and prompt ancestry.Guard diversity68%How differently seats draw a permission boundary.Dissent reserve28%Compute protected for objections that cannot win a vote.

ASSEMBLY RECEIPT / CA-0001

REPRODUCIBLE / V1

1Seal

2Audit

3Map

4Cross-examine

5Revise

6Dissent

7Gate

RESOLUTION MODE

UNRESOLVED FORK

FORK — separate the right to inspect weights from authority to operate high-impact tools.

1.9 effective voices5 nominal seats

Toy dissent indexRule-derived, not measured

Toy procedural indexIllustrative—not legitimacy

01 / FACT

Open weights improve inspectability and adaptation but do not reveal training data or remove deployment costs.

02 / VALUE

Distributed scrutiny and access conflict with irreversible proliferation risk.

03 / PERMISSION

FORK — separate the right to inspect weights from authority to operate high-impact tools.

MAJORITY SYNTHESIS

Release weights with staged capability access and reproducible evaluations.

capture radius5

Weights cannot be recalled; tool restrictions at one host do not travel with them.

FALSIFIER / Demonstrate a portable, independently enforceable capability boundary across derivative deployments.

02 / Monoculture mirage

Seven models can be one witness.

Independence is empirical, not theatrical. Prompting sibling models with different ideologies creates perspectives—not new evidence.

Existing voices7The nominal number shown in the interface.Shared-error overlap67%Assumed residual covariance; a real deployment must measure it.

Add one more seat

8voices in the room

1234567+

discount covariance

1.3effective plurality

Toy vote signal88/100

Toy resilience index54/100

The star model raises average skill while tightening the shared failure surface.

03 / Capture map

Who must be compromised?

Safety prose is not a force field. Institutional design can only make capture more expensive, visible, and recoverable.

CAPTURE RADIUS2independent roles needed for undetected control

Toy exposure index41/100

Coordination-cost index17/100

04 / Openness vector

“Open” is not a bit.

K3 releases a frontier post-trained checkpoint. That is significant. It is not the same as releasing the path that made it.

KIMI K3 / INTERPRETIVE AUDITCategorical visibility profile · editorial bands, not scores

Weights

ReleasedFull post-trained checkpoint

Inference

PublishedModel code + engine recipes

License freedom

ConditionalCustom terms; a Model-as-a-Service threshold can require a separate agreement

Base checkpoint

Not releasedPost-trained weights only

Training stack

DescribedNot a reproducible training release

Training data

PartialCategories, not a complete corpus

Teacher policies

Not releasedNine policies are described

Hardware reach

Constrained64+ accelerators recommended

“

Open weights move the location of politics. They do not remove it.

Once a model can be copied, tuned, and hosted, governance shifts from one provider’s policy to an ecology of licenses, hardware, fine-tunes, routers, tools, and local law.

Therefore:

publish provenance; bind tools externally; version constitutions; test the whole harness; and never let “open” smuggle in “transparent,” “independent,” or “safe.”

05 / Paper → polity

Borrow the shape, not the claim.

These are institutional analogies inferred from K3’s design. MoE experts are computations—not citizens, ideologies, or independent witnesses.

K3 MECHANISMIN THE REPORTSPECULATIVE TRANSLATION

16 / 896 routed + 2 shared

Each Stable LatentMoE layer selects 16 of 896 routed experts per token, alongside two shared experts.

Route a few relevant seats—but keep their identities legible.

Attention Residuals

K3’s Block AttnRes selects among block-level aggregates from earlier depth.

Let a council recall prior objections without blindly accumulating them.

9 → 1 policies

Multi-teacher distillation consolidates domain × effort specialists.

Distill procedural skill; preserve value disagreement outside the weights.

White-box harness

Training varies prompts, tools, memory, skills, and scaffolds.

Evaluate the assemblage—not a model name floating free of its governor.

Hidden verifier

Autonomous execution is rewarded by external state checks.

Authority follows observable outcomes, never self-reported completion.

IMPORTANT DISTINCTION

Internal MoE is not ideological pluralism. K3’s routed experts are jointly optimized inside one deployed checkpoint, selected by layerwise routers, and contribute to one model policy. They are not independently governed witnesses. A constitutional council begins only when whole model systems can disagree visibly under a common protocol.

06 / The operating vow

Freedom inside a sharp boundary.

A reusable ability for model councils. Its constraints create no authority; they only reduce ambiguity and make failure visible.

Download protocol v1 Open the POLIS/896 field lab Enter the AgentTool airlock Enter the KARMA reciprocity loop Play the CIVIC LOOM/16

ABILITY CARD / CA-V1CONJURATION × MANIPULATION

多

NAME

Convene Without Collapse

Desire: Reach a decision without laundering factual uncertainty, value conflict, or denied permission into one smooth answer.
Trigger: “Convene a council on whether…”, “Compare these model systems under different constitutions…”, or another consequential, contested question where plural review can change the outcome.
Anti-trigger: “Summarize this public memo.” Routine, reversible transformations stay with one accountable workflow unless their stakes change.
Input → output: Dilemma, mandate, stakeholder brief, model passports, charters, and budgets → a typed decision, guard record, minority report, unknowns, and a reproducible receipt.
Condition: Named seats; pinned model, harness, charter, and guard versions; sealed first claims; an external action gate.
Limitation: No seat selects the council, writes the synthesis, and authorizes execution. One provenance family cannot form a quorum.
Budget: Finite rounds, explicit tool and token caps, plus 20% of deliberation compute reserved for dissent.
Breach: Fail closed on irreversible action; preserve partial evidence; publish the missing condition and the unresolved minority report.
Proof: A reproducible routing seed, claim/evidence map, guard decisions, falsifiers, action receipt, and explicit non-claims.
Exit: Release temporary state, revoke scoped tools, retain the signed receipt, and let newer human instruction dissolve the assembly.
Non-claims: The assembly does not prove truth, representation, legality, fairness, safety, or authority. Stricter prose creates no power unless a real boundary or verifier enforces it.

07 / Honest edges

What this does not prove.

01
Open weights do not imply open data, reproducible training, or transparent provenance.
02
Different constitutions do not establish independent evidence.
03
Majority agreement is neither truth nor moral legitimacy.
04
Model plurality is not demographic or democratic representation.
05
A transcript does not expose complete internal reasoning.
06
Guards do not prove safety, and pluralism is not a route around safeguards.
07
This simulation does not demonstrate real-world accuracy or resilience.
08
Every router and synthesis rule embeds normative choices. There is no neutral council.

Primary sources

Research lineage.

K3 has no arXiv record as of 28 July 2026; Moonshot’s 47-page GitHub report is the canonical paper.

Kimi K3 technical reportKimi Team · 27 July 2026 · canonical PDF Kimi K3 model + weightsMoonshot AI · 2.8T / 104B active · custom license Kimi K3 product launchMoonshot AI · 16 July 2026 · weights followed 27 July Kimi K3 LicenseCustom terms · Model-as-a-Service and display conditions Attention ResidualsKimi Team · arXiv:2603.15031 Kimi LinearKimi Team · arXiv:2510.26692 Kimi K2.5Visual Agentic Intelligence · arXiv:2602.02276

A parliament,not a personality.

Intelligence can be mixed.Authority must be divided.