Workbench

isomorph

Most hard problems have already been solved somewhere else, in a different skin. isomorph finds that place and makes its rules work on my system.

Stage
Decide
Since
Jun 2026
Leaves
Optionally, a twin in VISION.md
“Isomorph isn't a bug finder. I would say it's more of an architect.”
Me, Jun 2026

Two skins, one skeleton

Two systems are isomorphic when their parts relate to each other the same way, even though they look nothing alike. Constant carries one coding conversation between Codex and Claude Code. Strip the software away and the shape is older than computers: one mind that has to survive, more than one body it can live in, and a risky transfer in between.

Medicine and science fiction have both thought hard about that shape. Software that does it is new.

Constant

the harnessnative resumethe conversationthe transcoderCodex sessionClaude session

Re-sleeving

the operating theatrewaking upthe stack (a mind)the surgeonsleeve Asleeve B

Hover or tap any part. Its twin lights up on the other side.

The same skeleton in two skins. Each part on the left has exactly one partner on the right, and the arrows run the same way.

Why borrow from a mature domain

A young system only knows the failures it has already had. A mature domain has had generations of them, and it wrote down what it learned as rules. Surgeons don't destroy the donor. Factories catch a defect at the station that made it. Emergency rooms treat by severity, not by order of arrival.

If the structure really matches, those rules carry over. Each one becomes an invariant my system should hold, and I can check the code against it today.

A spec I didn't have
The twin's laws become invariants, each scored satisfied, partial or violated against the real code.
Blindspots
Laws with nothing to map onto point at problems nobody had thought to look for.
A reframe
Sometimes one structural change satisfies several laws at once. That is the best outcome.

Eight moves, used as a palette

A run is a back-and-forth. I throw out a rough direction, the agent sharpens it, I push back, and we build the world together until I can stand inside it. The test for that part is a line from my own brief: “i want to feel it.” Small problems skip moves. Forcing all eight onto a small problem is its own failure.

  1. Find the twin
    1. Frame

      Name the parts, what flows between them, and what must stay true.

    2. Climb

      Find the altitude where something else starts to rhyme.

    3. Latch

      Float one to three mature domains. The first guess is a direction.

    4. Sharpen

      Refine to the precise twin, then test it.

      GuardStructure, not surface. If you can't name the matching relations, there is no twin.

  2. Live in it
    1. Build the world

      Map every part one to one, in the twin's own words, until it can be felt.

  3. Mine it
    1. Go generative

      List the twin's laws: what must hold or it dies, forks or wakes up wrong.

    2. Laws to invariants

      Turn each law into a rule for the real system and score it honestly.

    3. Read the gaps

      Laws with nothing to map onto are blindspots. So are abilities the twin has and the system lacks.

Finding the right altitude

Describe a system too precisely and nothing else in the world looks like it. Describe it too loosely and everything does, which is just as useless. The climb stops at the altitude where a few real twins come into view.

  1. Too abstract

    “Something that moves data from A to B.”

    Twins in sight: Everything rhymes, so nothing useful carries over.

  2. The right altitude

    “One continuous memory, moved between bodies that each run it their own way.”

    Twins in sight: Transplant surgery. Re-sleeving. A play recast halfway through a run.

  3. Too concrete

    “A Rust CLI that rewrites JSONL session files between two formats.”

    Twins in sight: Nothing else rhymes with it.

Constant described at three altitudes. Only the middle one finds twins with laws worth inheriting.

The guard: structure, not surface

My first guess for Constant was biohacking. It sounds close, but an analogy that only shares the look will hand over the wrong rules. So every candidate has to pass one test: can you name the relations that match? Biohacking has one body and no transfer. Re-sleeving has the same parts, constrained the same way.

The first guess

Biohacking

one body+++

It sounds right, but biohacking improves one body. Nothing moves anywhere.Surface match. Dropped.

The sharpened twin

Re-sleeving

body Abody Bmind

One mind, more than one body, and a transfer between them. That is Constant's shape.Structural match. Kept.

The same first guess, drawn as relations.

Run one: Constant, through re-sleeving

Constant moves a live coding conversation between Codex, Claude Code and OpenCode. With the twin agreed, the next step was to build the world in its own words, then mine it for laws.

ConstantRe-sleeving
  • the conversationthe stack, the mind that must survive
  • a model sessiona sleeve, one body
  • weights and trainingthe sleeve's DNA
  • the transcoderthe surgeon
  • native resumethe new body waking up
  • the harnessthe operating theatre
  • Don't kill the donor

    Invariant: Never destroy the source session.

    Found: Switching back overwrote the original session in place, with a lossy copy of a copy.

    Violated
  • Check compatibility before insertion

    Invariant: The target must accept the data's native format before the transfer starts.

    Found: Formats were matched only after the target rejected them.

    Violated
  • No power cut mid-operation

    Invariant: Tear the old session down gracefully.

    Found: Processes were force-killed.

    Violated
  • Know what transfers and what belongs to the body

    Invariant: Carry the memory, leave the tool reflexes behind.

    Found: Tool calls were already dropped on purpose. The loss is deliberate.

    Satisfied

Blindspots the frame exposed. The new body can reject memories of tools it doesn't have. The “mind” had been defined too narrowly: the chat, without the state of the world around it. A transfer is also an injection surface, because someone can forge memories. And switches weren't atomic, so the donor could be gone before the recipient was confirmed alive.

The reframe

There was no real stack. Every switch copied one body into another and hoped. The twin's answer was to keep one canonical record of the conversation and treat every model session as a disposable projection of it. That one change satisfied four laws at once.

It is how Constant works today: the neutral snapshot is the durable record, and native sessions can be reprinted from it. Constant

Run two: the Fantopy Orchestrator, through a supply chain

The Orchestrator turned GitHub issues into draft pull requests through a line of agents: triage, grading, a fixer and a reviewer. My rough guess was a factory. Sharpened, the twin became a multi-stage supply chain run on lean manufacturing and the theory of constraints.

  1. Issue filedRaw material at the dock

    No reject path. A factory sends bad material back. Triage could only push work forward.

  2. TriageGoods-in inspection
  3. Difficulty gradeMaterial grading
  4. Fixer agentAssembly station

    Perishable stock. The repo keeps moving while a fix is in flight, so a fix built on an old base spoils.

  5. Reviewer agentQA gate

    Rework loop with no age limit. Reviewer and fixer could hand work back and forth forever.

  6. Pull requestFinished goods shipped
The line as the twin sees it. Marked stations carry a blindspot the factory frame exposed.

Factories also know two things the orchestrator didn't measure. The slowest station governs throughput, so effort should go there and not to the most visible station. And scrap analysis: where on the line do jobs die? That is the signal for the weakest station, and nobody was collecting it.

The twin named perishable stock on 5 June. Three days later the orchestrator hit exactly that: a worker ran its first fix before refreshing its base, so an issue that depended on newly merged work couldn't find its file.

When one twin isn't enough

Real buildings have wings. The orchestrator's happy path is a factory, but a job that fails verification behaves like a patient. So it got a second twin, an emergency room, with the border between them written down. That way a factory law never gets applied to a patient.

Happy path

The factory wing

Material flows through stations and ships. Factory laws apply here: throughput, work-in-progress caps, quality at the source.

fails verificationdischarged back
Failure path

The hospital wing

A sick job is triaged, treated and discharged. Emergency-room laws apply here: severity first, bed capacity, bounce-back as a warning sign.

The emergency-room twin turned five planned features into one control loop for severity and capacity. Two parts shipped. The twin read a job that keeps bouncing back as a warning sign, and I turned that into convergence-gated retries: keep going while each attempt fails on something new, stop early when it stalls. And dispatch began treating the most severe issue first, not the oldest issue number. That second idea wasn't in the original plan. The twin surfaced it.

When two twins agree

The two runs used unrelated domains on unrelated systems. Both still flagged an unguarded intake. When two mature fields point at the same failure independently, I treat it as real.

Re-sleeving

Forged memories poured into a new body.

Supply chain

Unvetted material accepted at the dock.

Same failure

An unguarded intake

When it stays quiet

isomorph steps aside, and says why, when:

  • The problem is concrete and local. “Why is this function slow” wants a fix, not a cosmology.
  • No mature twin exists. If the structure is genuinely new, there is nobody to inherit from. Reason from first principles instead.
  • The only analogy is skin-deep. A twin that shares looks but not structure hands over the wrong laws. Better to say so.

What each move stands on

None of these names come up during a run, because a textbook sentence breaks the world we are building. They are there so every move has something solid under it.

FrameSystems thinking, Donella Meadows
A grammar for structure: stocks, flows, feedback, leverage points.
ClimbAbstraction laddering
Control over the altitude where a real twin becomes visible.
LatchDesign by analogy, biomimicry, TRIZ
The habit of assuming the problem is already solved somewhere else.
SharpenStructure-mapping, Dedre Gentner (1983)
Good analogies carry relations, not looks. First principles checks the rest.
Go generativeThe outside view, Kahneman and Flyvbjerg
Inherit a mature class's failure record instead of rediscovering it.
Read the gapsPremortem (Klein), FMEA, inversion (Munger)
A systematic search for how it fails and what nobody looked at.
Stay quietCynefin, Dave Snowden
Knowing when a problem is too novel for analogy.

How a run is graded

  1. 1A structural twin: the matching relations can be named, not just the feeling.
  2. 2Several of the twin's laws, each mapped to a concrete invariant and scored satisfied, partial or violated.
  3. 3Blindspots that weren't visible from inside the original frame.
  4. 4Ideally, an architectural reframe that the twin implies.

A metaphor that produces no invariants and no blindspots is decoration. That run failed, and it gets redone.