CONCEPT ATLAS / 001 Philosophy · Agent design · Learning notebook
Drew, in Progress

ONTOLOGIES / HARNESSES / LIVING SYSTEMS

Ways of seeing.
Ways of acting.

Before asking what a machine can do, ask what it notices, which values guide it, and where its authority ends. This is the question behind my ontological harness work.

Look through a different lens ↗

Drew Beyersdorf · September 2026 · Written with AI assistance

A LITTLE VOCABULARY

Make the assumptions visible.

A useful structure can be explained before it is automated.

01 / ONTOLOGY

What belongs in the picture?

Here, an ontology names the things a system reasons about and how they relate: people, needs, commitments, evidence, choices, consequences. What it leaves out matters too.

02 / HARNESS

How does an idea guide action?

A harness gives an agent a working environment: context, tools, permissions, decision rules, and a way to check results. It makes a philosophical preference operational enough to examine.

03 / PROVENANCE

Which lens produced the answer?

My portfolio asks every verdict to carry a reason and a named, versioned ontology. A change of values should remain visible when the answer travels to another system.

ONE SITUATION / FOUR READINGS

Change the lens.

A fixed illustration of the ideas. This example runs no model and makes no real decision.

THE INVENTED SITUATION

An assistant has been asked to start another task. Its current workload is uncertain, and nobody has checked the latest status.

GĪTĀ / AN ENGINEERING ADAPTATION

Clarity before more motion.

gita / sketch-v1
  1. K1 / FIELD

    Observe

    What work is already in progress? What can we actually observe?

  2. K2 / QUALITY

    Assess

    Distinguish clarity, restless activity, and obscurity in the available signals.

  3. K3 / GATE

    Decide

    Pause admission while the relevant state remains unclear.

  4. K4 / ROLE

    Keep boundaries

    The worker reports its state; a separate reviewer checks the decision.

  5. K5 / OUTCOME

    Follow through

    Do not turn the desire for progress into evidence that more work is safe.

Test the adaptation: do these distinctions improve decisions compared with a plain capacity check?

These are my design interpretations, not authoritative summaries of a philosophical tradition, flight procedure, or biological system. The sketch labels belong to this illustration. They are not versions of a deployed evaluator.

THE FIVE-ROLE HYPOTHESIS

A resemblance worth testing.

The diagram is a proposal about structure. Its usefulness still has to be earned.

My Ontological Harnesses portfolio explores a shared sequence: field, quality axes, gate, role integrity, and outcome discipline. The Gītā Harness is one lens; Stoic, aviation, and swarm sketches bring different emphases.

The interesting question is whether that structure helps us compare decisions without erasing their differences. A shared role does not imply shared thresholds, identical values, or permission to transfer a verdict unchanged.

This is a design hypothesis, not a mathematical proof that all traditions or intelligent systems are equivalent. We selected examples that fit. A stronger test would deliberately look for a domain that breaks the map, then compare it against a simpler design.

The public atlas explains selected concepts from my educational portfolio. It makes no claim of measured performance or production reliability. A broader biomimicry harness remains an idea to develop.

SALTWATER / A LIFECYCLE HYPOTHESIS

How an agent earns more room.

A second harness asks how capability should develop over time.

My SALTWATER project borrows its imagery from crocodile development. It maps five stages onto an agent lifecycle: prepare the environment, check readiness, move carefully, try bounded work, and review whether wider responsibility is justified.

  1. I / NEST

    Prepare the environment

    Declare the task, available tools, resource limits, and data boundary before work begins.

  2. II / VOCALIZE

    Check readiness

    Require the relevant workers to report their actual state. A missing signal remains missing.

  3. III / TRANSPORT

    Keep the route explicit

    Know where a task and its evidence will travel. Check the handoff at each step.

  4. IV / ESTUARY

    Work within limits

    Try a bounded task. Keep the result, failures, and review visible before changing its scope.

  5. V / OPEN WATER

    Earn wider responsibility

    Use task-specific evidence to assess readiness. A successful trial does not remove a person’s authority over consequential actions.

The Gītā lens asks how to judge action; SALTWATER asks how to introduce responsibility gradually. I treat the biological story as a design metaphor. It does not establish that these gates make a deployed system reliable. That still needs testing.

BIOMIMICRY / LEARNING FROM LIFE

Borrow a question from nature.

The branching design of this journal is an invitation to look more closely.

The Biomimicry Institute’s AskNature gathers biological strategies as inspiration for design. My interest is in moving from an appealing resemblance to a specific engineering question.

FEEDBACK → REVISION

Can the system notice change?

My design question: which observation should cause an agent to revisit its plan? Test whether a feedback loop catches a changed condition sooner than a fixed plan.

DIVERSITY → ALTERNATIVES

Are these voices independent?

My design question: are several agents bringing different evidence, or repeating one source? Compare genuinely separate checks with repeated agreement.

RENEWAL → FRESH EVIDENCE

When should a signal expire?

My design question: how long should yesterday’s “ready” remain credible? Try an explicit expiry rule and record both missed changes and unnecessary rechecks.

These are proposed experiments inspired by living systems. Biological language does not establish that software is alive, conscious, optimal, or reliable.

RETURN TO THE PERSON

What is the structure for?

A whole life supplies the purpose.

  1. NeedBelonging
  2. IntentionMake time together
  3. ActionProtect an evening
  4. ReviewDid it help?

This invented example is deliberately ordinary. A useful ontology should leave room for love, rest, dignity, curiosity, beauty, and contribution. The person decides what matters and whether the proposed help is welcome.

Read “A whole life is the point” ↗

THE LEARNING NOTEBOOK / SEPTEMBER 2026

Keep the lesson.
Keep the question.

Ideas become more useful when I can name what they mean, try them and revise them.

This notebook connects operating lessons, design ideas and subjects I am studying. The labels matter: a lesson, a hypothesis and a reading topic make different claims. It is a growing selection, not a complete account of everything I know.

Practice notes9 notes

Lessons I am carrying into how I work. These are operating preferences, not evidence of measured outcomes.

Practice note

Queued is not running

A request in a queue tells me what is intended. I look for current execution evidence before calling it progress.

Practice note

Follow the whole handoff

Requested, delivered, acknowledged, executed and checked are separate steps. My practice is to verify the step that matters to the outcome.

Practice note

Saving needs a readback

A draft is only useful later if it can be found again. I check the saved path and content rather than relying on a success message.

Practice note

One clear decision screen

My preferred briefing gives the outcome, the recommendation and the next decision in plain language.

Practice note · Design note

Corrections belong in the record

When evidence overturns an inference, I update the active conclusion and keep the correction history visible.

Practice note · Design note

People need context

A contact match says little about a relationship. I want shared history, trust and mutual commitments to remain part of the picture.

Practice note

Treat credential exposure as an incident

My operating rule is to replace exposed credentials through their owner and established credential store, without copying them into more places.

Practice note

Freshness is part of evidence

A machine or workflow that worked yesterday may be unavailable now. I want a current check for claims about what is working.

Practice note

Compounding systems

My working distinction: recovery restores a service; learning changes guidance, tests or the mechanism so the same failure becomes less likely to recur.

Working models1 notes

An application I am still testing. A useful analogy needs boundaries.

Working model · Study topic

Amdahl’s law

For a fixed workload, the serial share limits the benefit of parallel work. My application to handoffs is a working model: first find the step that cannot progress in parallel.

The original argument concerns computer performance. The handoff application above is my interpretation. Read Amdahl’s 1967 paper ↗

Design notes16 notes

Ideas about how a system should be organized. They describe designs, not a claim that every mechanism is deployed.

Design note

Replaceable assistants

A design aim: changing assistants should not break ownership, stored knowledge or the evidence needed to continue a task.

Design note

Compact work orders

State the goal, input references, allowed actions, budget, stop condition and expected result. Keep the teaching explanation readable for the person.

Design note

Name the mode of work

Observing, planning, deciding, executing, verifying, merging and teaching have different outputs. I want each run to say which job it is doing.

Design note

Governance boundaries

A proposed design should name who owns a decision, which actions are permitted and when the task must stop or escalate.

Design note

Metaphors and machine contracts

Metaphors can help a person think. Concrete identifiers and explicit rules make a machine contract easier to check.

Design note

Shared resource pools

My capacity model links consumers to their shared resource pool so the same allowance is not counted twice.

Design note

Goals tied to evidence

A goal becomes inspectable when its target, owner, constraints and progress evidence can be read together.

Design note

Explicit state

The design keeps requested, queued, working, blocked and checked results distinct instead of compressing them into one progress label.

Design note

Provenance across the system

An output should retain enough source and transformation history for someone else to examine the claim.

Design note

Bounded autonomy

I frame autonomy through explicit scope, tools, budgets and stop conditions. Wider responsibility remains something to assess with evidence.

Design note

One writer, separate review

My coordination preference is one accountable writer for a shared artifact, with a separate check of the result.

Design note

Material knowledge intake

The proposed intake rule keeps durable decisions, lessons and evidence while separating them from transient chatter.

Design note

Deduplication by identity

The design uses stable source identity to recognize a repeated item before it becomes a second update.

Design note

An encyclopedia of entities

I am exploring notes organized around people, projects, goals, resources and decisions, with linked views connecting them.

Design note

Output boundaries

A result needs a permitted destination as well as correct content. My design makes that destination part of the task.

Design note

Machine identity and capability

The proposed record separates the identity of a machine from what it can do and what it is currently doing.

Study map41 notes

Questions from my study material. A topic appearing here does not mean I have mastered it.

Study topic

YAGNI

What evidence makes a feature necessary now, rather than merely imaginable?

Study topic

Short-circuit evaluation

Which conditions should be checked first, and which later checks can be skipped?

Study topic

TCP/IP

What does a connection promise, and what must an application handle when it is interrupted?

Study topic

CAP theorem

During a network partition, what tradeoff does a particular system make between availability and consistency?

Study topic

Lamport clocks

How can events be ordered when there is no shared physical clock?

Study topic

Finite state machines

Which states and permitted transitions make a workflow easier to reason about?

Study topic

PLC ladder logic

How can an industrial control program make conditions, interlocks and transitions explicit?

Study topic

Typed transitions

Can the representation make an invalid transition harder to express?

Study topic

Station scheduling

Which tasks can overlap, and which dependencies force them to wait?

Study topic

Information theory

What does surprise tell us, and which uncertainty is the model measuring?

Study topic

mAP50 evaluation

What does this detection metric measure, and which kinds of error can its aggregate score hide?

Study topic

Precision and recall

Which mistakes matter more for the use case: false alarms or missed cases?

Study topic

Caching

When is a saved computation still valid, and what event should invalidate it?

Study topic

Defense in depth

Which independent controls remain useful when another control fails?

Study topic

Least privilege

What is the minimum access a person or process needs for its assigned task?

NIST’s definition ↗

Study topic

The principal-agent problem

How might an agent’s incentives differ from the interests of the person it serves?

Study topic

Lot traceability

What chain of identity and events lets us trace a batch through a process?

Study topic

The observer pattern

How should a change notify interested parts of a system without tightly coupling them?

Study topic

Publish/subscribe and request/response

When should work be announced as an event, and when does it need a direct reply?

Study topic

Circuit breakers

When should repeated failures stop new attempts, and what evidence should permit a retry?

Study topic

Control theory

What is observed, what is adjusted, and how does delay affect the feedback loop?

Study topic

Concurrency and parallelism

Which work can be in progress together, and which work can actually execute at the same time?

Study topic

SMED and changeover

Which setup steps can be moved out of the critical path?

Study topic

Behavioral economics of alerts

How do attention, incentives and framing affect whether someone acts on an alert?

Study topic

Integration across principles

Where do individually sensible components fail when they are connected?

Study topic

The semantic-tree learning model

Which foundational ideas make the smaller facts easier to place?

Study topic

A latticework of mental models

Which other discipline offers a useful explanation or a counterexample?

Study topic

DiSSS and CaFE

How can decomposition, selection, sequencing and feedback make a learning plan more deliberate?

Study topic

Ultralearning

What would direct practice, retrieval and timely feedback look like for this skill?

Study topic

Learning an unfamiliar industry

What can observation, primary sources and a small practical test reveal that a summary cannot?

Study topic

AI as a tutor

Can I explain and test the idea independently after the assistant’s explanation?

Study topic

Spaced repetition

Which knowledge should be recalled again, and how will recall be checked?

Study topic

Learning by building

What small artifact would expose whether I understand the idea?

Study topic

Knowledge compression

What becomes easier to retrieve in a summary, and what important context might be lost?

Study topic

Zettelkasten

How can linked notes preserve distinct ideas and reveal relationships?

Study topic

CODE and PARA

How should captured material be organized so it can support an actual project?

Study topic

Using a second brain

Does the knowledge store change what I can decide or make, or only accumulate notes?

Study topic

Breadth and depth of skills

Which areas need working familiarity, and which need sustained depth?

Study topic

A learning rhythm

Where can deliberate study, practice and review fit into ordinary work?

Study topic

Domain rotation

When does moving between subjects help connection-making, and when does it interrupt depth?

Study topic

Evidence of understanding

Can I explain, apply, transfer and challenge the idea without simply repeating its wording?