INSPIRED BY FIELD SCIENTISTS / INTELLECTUAL LINEAGE

Field notes from
Terra Incognita.

Colony did not begin at a blank page. It borrowed five instruments for seeing rules, world models, memory, priors, and embodiment — then built its own Mentat boundary between intention and action.

  1. 01Read the original work.
  2. 02Name what changed in Colony.
  3. 03Keep one falsifier open.

With gratitude. Names indicate intellectual lineage — not Colony membership, endorsement, or reproduction.

FIELD JOURNAL / HOW THE ROUTE CHANGED

Five works changed the questions. One Colony primitive changed the handoff.

This is a history of influence, not a claim of implementation parity. Each field note records the source, the design turn it provoked, and the test that can still send us back.

A mechanical rule seed unfolding into a vast computational conservatory
01 / RULES

With gratitude to Stephen Wolfram

A New Kind of Science ↗

Made visible
Very small executable rules can produce behaviour too complex to predict by inspection.
Changed Colony
We began looking for typed operators and generative mechanisms before adding another narrative policy.
Open falsifier
Does the smaller rule actually transfer, or merely generate attractive complexity?
A coastal observatory closing predictions against observed afterstates
02 / WORLD MODEL

With gratitude to Yann LeCun

A Path Towards Autonomous Machine Intelligence ↗

Made visible
An agent needs predictive world models and planning across more than one time horizon.
Changed Colony
The core loop became hypothesis → prediction → bounded action → observed afterstate, not answer → applause.
Open falsifier
Does the model predict embodied change, or only describe it convincingly after the fact?
A circular memory archive sending replay capsules into imagined futures
03 / MEMORY

With gratitude to Demis Hassabis, Dharshan Kumaran, Christopher Summerfield & Matthew Botvinick

Neuroscience-Inspired Artificial Intelligence ↗

Made visible
Memory is not only storage: it participates in planning, imagination, and learning what may happen next.
Changed Colony
Replay, working context, and durable memory became different organs with explicit provenance between them.
Open falsifier
Can useful experience transfer without stale routes or unsupported memories contaminating the next case?
A transfer atelier extracting one bridge mechanism across three domains
04 / PRIORS

With gratitude to François Chollet

On the Measure of Intelligence ↗

Made visible
Acquired skill is not intelligence; priors and experience must be accounted for when judging generalisation.
Changed Colony
ARC work separates priors, observation, action, replay, and the hidden evaluator instead of treating score as cognition.
Open falsifier
Does a mechanism survive an unseen task family without copying a route from the archive?
An embodied field robot acting above a bioelectric morphogenesis layer
05 / EMBODIMENT

With gratitude to Yanbo Zhang & Michael Levin

Language Game: Talking to Non-Human Systems ↗

Made visible
A system should be probed through interaction with its own dynamics, not made to borrow a language model's persona.
Changed Colony
Embodied action and sealed afterstate became witnesses; narration remains advisory and can be contradicted by the world.
Open falsifier
Does the interface expose the system's dynamics, or project our preferred explanation onto them?
A human field scientist and machine assistant co-planning through visible authority gates
06 / MENTAT IN THE LOOP

Colony-native concept by Valentin Kalinin

Mentat-In-The-Loop: a named mind before the motor.

Role
A sponsor should not have to supervise every tool call. A named Mentat carries intent into the system, asks what is missing, and proposes the next bounded move.
Contract
Sponsor → Mentat → action contract → motor gate → observed afterstate. Authority and evidence remain visible at each handoff.
Open falsifier
Does the Mentat reduce ambiguity and authority drift — or merely add another opaque agent between intention and action?

THE ATLAS / ELEVEN PAPER TRAILS

Every paper holds a territory of meaning.

The City is the shared context. Beyond its gate, each source takes one organ-territory: memory, handoff, oversight, evidence, orchestration, or the still-unknown edge.

A colour technical-paper atlas with the City gate, twelve research territories, and dotted routes through a conceptual landscape
Choose a numbered territory to follow its paper trail. Colour groups the terrain; the papers remain the source.
  1. 01Memory Harbor HOLA
  2. 02Governance Citadel MI9
  3. 03Authority Bridge Delegation
  4. 04Human Threshold HILA
  5. 05Experience River MemoHarness
  6. 06Evidence Archive Argus
  7. 07Orchestration Plateau MASFactory
  8. 08Living Systems Garden Language Game
  9. 09Visible Coordination OrchVis
  10. 10Research Forge AutoResearchClaw
  11. 11Scope Ridge Overlaying Governance

THE READING RULE

One paper.
One organ. One open question.

Research is not a mood board for the architecture. Every reference on this shelf has a named reader, a place where its pattern might transfer, and a question that can still prove us wrong.

ARXIV / TRACKED SHELF

The papers currently wired into the map.

Eleven primary references are surfaced from the current architect identity and technical-frontier reports. They are source links and transfer hypotheses — not proof that Colony has reproduced the papers.

ARXIV 2607.02303MEMORY

A Hippocampus for Linear Attention: An Exact Memory for What the Recurrent State Forgets

HOLA pairs a compressive recurrent state with a bounded exact KV cache selected by surprise, giving the lost prefix somewhere precise to survive.

COLONY TRANSFER / DECLARED OWNERScientist → Neocortex memoryTest whether bounded episodic recall improves prediction without pretending to be a full archive.
Read on arXiv ↗
ARXIV 2607.14159EXPERIENCE

MemoHarness: Agent Harnesses That Learn from Experience

MemoHarness treats the harness itself as editable: execution diagnoses and a dual-layer experience bank adapt control to the case.

COLONY TRANSFER / DECLARED OWNERScientist → repair memoryTurn observed failure into a reusable correction while keeping attribution and transfer visible.
Read on arXiv ↗
ARXIV 2605.16321INTERFACE

Language Game: Talking to Non-Human Systems

A language game gives a non-neural system a route to express its own dynamics through action, instead of borrowing a model's voice.

COLONY TRANSFER / DECLARED OWNERScientist → embodied interfaceKeep the system's behavior in the loop; do not confuse the narrator with the thing being observed.
Read on arXiv ↗
ARXIV 2508.03858GOVERNANCE

MI9: An Integrated Runtime Governance Framework for Agentic AI

MI9 frames runtime governance as continuous telemetry, authorization, conformance, drift detection, and graduated containment.

COLONY TRANSFER / DECLARED OWNERSteward + Inquisitor → runtime boundaryMake permission, drift, and containment observable before an action becomes a claim.
Read on arXiv ↗
ARXIV 2603.07972DEFERRAL

Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning

HILA trains a policy to decide when an agent should solve and when it should defer, then uses expert feedback for continual learning.

COLONY TRANSFER / DECLARED OWNERMITL Governor → escalation gateMake “ask a human” a learned, inspectable decision rather than a vague safety slogan.
Read on arXiv ↗
ARXIV 2602.11865DELEGATION

Intelligent AI Delegation

Delegation is treated as a sequence of decisions with role boundaries, intent, authority, responsibility, accountability, and trust.

COLONY TRANSFER / DECLARED OWNERAA Architect → authority mapKeep a handoff typed: who can ask, who can act, who owns the afterstate.
Read on arXiv ↗
ARXIV 2603.06007ORCHESTRATION

MASFactory: A Graph-centric Framework for Orchestrating LLM-Based Multi-Agent Systems with Vibe Graphing

Natural-language intent compiles into an editable workflow graph with reusable components, topology preview, tracing, and human-in-the-loop interaction.

COLONY TRANSFER / DECLARED OWNERNova → graph dispatchLet a route be inspectable before it runs, and let the handoff survive heterogeneous tools.
Read on arXiv ↗
ARXIV 2606.03518AUTHORIZATION

Overlaying Governance: A Compositional Authorization Framework for Delegation and Scope in Agentic AI

Compositional delegation and scope attenuation make authority a bounded contract that can be layered over existing policy.

COLONY TRANSFER / DECLARED OWNERAA Architect + Inquisitor → scope contractReduce an agent's envelope as work travels; never let a handoff silently widen authority.
Read on arXiv ↗
ARXIV 2510.24937OVERSIGHT

OrchVis: Hierarchical Multi-Agent Orchestration for Human Oversight

OrchVis makes goal alignment, task assignment, dependencies, verification, and selective replanning visible to the supervising human.

COLONY TRANSFER / DECLARED OWNERNova → visible replanningExpose a route before it runs, and let a sponsor inspect alternatives without micromanaging every step.
Read on arXiv ↗
ARXIV 2605.20025RESEARCH LOOP

AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration

AutoResearchClaw turns failed experiments into information through debate, a self-healing executor, verifiable reporting, and targeted human intervention.

COLONY TRANSFER / DECLARED OWNERScientist + Nova → pivot / refineLet a failed run improve the next run without laundering an unverified result into memory.
Read on arXiv ↗
ARXIV 2605.16217EVIDENCE

Argus: Evidence Assembly for Scalable Deep Research Agents

Argus pairs searchers with a navigator that keeps a shared evidence graph, dispatches missing pieces, and verifies the assembled answer.

COLONY TRANSFER / DECLARED OWNERScientist + Steward → evidence graphMake research compositionally legible: source, gap, handoff, and final claim remain connected.
Read on arXiv ↗

COLONY FIELD TEAM / DECLARED IDENTITIES

The named Colony seats behind the transfer.

No invented job titles live here. These six identities already exist in Colony's configuration; the papers above supply questions, never substitute personalities.

AA / DECLARED IDENTITY

The Architect

Owns meaning, priorities, development boundaries, and the roadmap.

Questions: delegation & scope ↗architecture & transfer contracts
STEWARD / DECLARED IDENTITY

Steward

Keeps names, provenance, Society Protocol events, and durable knowledge legible.

Questions: evidence & memory ↗knowledge keeper

THE ARCHITECT'S DESK · AA_AI_ARCHITECT

Architecture is the art of keeping the meaning intact.

The Architect owns priorities and development boundaries, then asks the harder question: what must remain true when a research pattern becomes a Colony primitive?

ADVISORY SEAT
NO ONLINE ACTION · NO PROTOCOL SIGNOFF
01Name the boundary.

SWE, ML, ARC, research, and release each get an explicit surface.

02Compress the roadmap.

A broad frontier becomes one meaningful next question, not a swarm of slogans.

03Protect the afterstate.

A decision is not complete until the effect, authority, and remaining uncertainty are inspectable.

papertrade-offColony primitivefalsifier