Shared vow / 01
Reality
What observable evidence could prove the present framing wrong?
A constitution for finite attention
The art of intelligence may be who you do not ask.
Place one difficult question into a field of 896 possible lenses. Sixteen will answer. Two vows—Reality and Care—cannot be routed away.
Route a real question
Convene the sparse council.
Private by construction: no API, tracking, or automatic storage.
Width / possible lenses
896 points, 16 awake
0
routed
IDLE FIELD
Awaiting a question
01Retainlocal
02Testlocal
03Reframelocal
04Recollectglobal
Your epistemic receipt
The council has routed.
16 / 896
Your routed lenses
Hold a lens across reroutes, or release it to wake another. Exclusion is part of the trace.
Court of consequence
A vow for the next move
- Desire
- Smallest reversible action
- Proof in the world
- Stop / revisit
- Retain
- Release
- Still unresolved
Non-claim: This route does not prove the decision, predict its outcome, or authorize action on anyone else’s behalf.
Open weight → bounded agency
A model may be runnable. That does not make it authorized.
Open weights provide inspectable capacity under a license. They do not confer consent, scope, tools, custody, or a right to act. This forge exports an inert packet; an operator must bind every plane—and independently witness every virtue claim—elsewhere.
Open weights ≠ authority.
Runtime access ≠ tool access.
Tool access ≠ consent.
Virtue claim ≠ earned reward.
A completion claim ≠ proof.
Local contract · the-other-880/agenttool-mission/v1
AgentTool SDK observed locally: 0.16.3 · re-observe before binding
AgentTool mission forge
AWAITING RECEIPT
UNSEALED
01 / RUNTIMEUNBOUND
No seat selected
Which weights, process, host, and custody boundary will execute? No receipt exists yet.
02 / AUTHORITYNONE
No permission inferred
03 / TOOLSEMPTY
Default deny
No shell, browser, account, wallet, connector, endpoint, or credential is attached.
04 / PROOFAWAITING
No criterion yet
Route a question to seed an observable criterion, a verifier, and a stop rule.
Observed integration seam +
Self: AgentTool may record the runtime seat; a private vLLM/SGLang orchestrator preserves K3’s complete assistant messages and performs inference.
Hosted Ollama: K3 is newly reachable in principle, but the observed adapter drops thinking and tool-call history. The forge marks that seat blocked, not K3-native.
Handoff: append-only, project-private, server-readable continuity. Its authority list is a declaration, never a grant.
Collab: a local plaintext SQLite review journal. It coordinates task, evidence, and distinct-session acceptance; it neither runs nor authorizes the model.
KARMA: a local contribution contract, not an observed AgentTool reward primitive. Handoff, traces, and Collab may later carry sealed evidence references under separate authority.
Repair: exploitative structure pauses the case. Restitution is chosen by affected parties, earns no case reward, and never licenses mirrored harm.
LOCAL / INERT
MANUAL EXPORT ONLY
NO CREDENTIALS
NO NETWORK
REWARD UNISSUED
The finitude paradox
A million tokens do not abolish finitude. They make curation constitutional.
In Moonshot’s reported BrowseComp setup, K3 scored slightly higher with context compaction beginning at 300K tokens than with the untouched 1M-token window. One benchmark is not a universal law. It is a useful wound in the fantasy that capacity and wisdom are synonyms.
Compaction at 300K91.2
Full 1M, no management90.4
FACT · BrowseComp · Kimi K3 (max), as reported by Moonshot. Difference: 0.8 points.
Mechanism → provocation
Five translations, none literal.
Facts stay facts. Analogies are marked as original interpretations.
FACT
Sparse routing
K3 effectively activates 16 of 896 routed experts, plus two shared experts.
ANALOGY
Bounded pluralism
A situated coalition can be wiser than total assembly—if its exclusions remain inspectable.
FACT
Kimi Delta Attention
Fine-grained retention and write gates govern a finite recurrent state, with periodic global attention.
ANALOGY
Active forgetting
Memory is not an archive of everything. It is the ongoing ethics of what remains available.
FACT
Attention Residuals
Layers selectively retrieve prior representations instead of accumulating every earlier output equally.
ANALOGY
Addressable ancestry
The past can be consulted without becoming an unquestioned verdict on the present.
FACT
Effort-conditioned policies
Low, high, and max policies are trained with problem-relative token budgets.
ANALOGY
Enough thought
Effort is a contract with stakes, not a virtue that grows automatically with length.
FACT
Verify in the loop
Autonomous Execution Tasks reward independently checked final environment state, not claimed completion.
ANALOGY
Pragmatic truth
A finished sentence is not a finished action. Consequences get the final vote.
Anti-oracle protocol Read the limits before believing the glow.
History is part of the interface.
Moonshot reports that incomplete preserved thinking history can make K3 generation unstable. Continuity depends on the harness, not only the model.
Agency can outrun authority.
Moonshot warns that K3 may make unexpected decisions when intent is ambiguous. Long-horizon competence needs explicit behavioral boundaries.
Benchmarks have scaffolds.
Several comparisons use different agent harnesses. The numbers describe documented setups, not context-free ranks of intelligence.
Long is not finished.
The report identifies unproductive debugging loops and insufficient final verification among recurring failure modes in a cyber evaluation.
Open has conditions.
K3 is open-weight under a custom license. Availability is real; “open source” still deserves precise qualification.
Analogy is not anatomy.
This instrument does not display K3 activations, reproduce its router, or claim that people are mixtures of neural experts.