Skip to content

The strongest model, at maximum effort, renaming a variable

Capability, reasoning depth, context size and verbosity collapse into one habit — everything on — because nothing ever named them separately.

→ watched the strongest model, at maximum effort, rename a variable

cost when it happens
medium
how often
constant
area
cost

The moment

A one-line rename. The session is on the most capable model available, reasoning effort at its ceiling, context loaded with half the repository, and it writes three paragraphs explaining the rename.

Why it happens

Nothing chose that configuration; it was left over from the last hard task. Capability, reasoning depth, context size and output verbosity are four separate decisions, and when nothing names them they collapse into one habit — usually "everything on". The routing documents this tool was distilled from fused effort into the model name and routed by the worker's job title, with performance figures nobody had measured.

Why a better model does not fix it

This is the one moment where a better model makes the problem worse: the stronger the default, the more expensive the habit. The failure is that four independent dials were never named, so no one can turn one of them down without turning all of them down.

What it costs

Money, in a way that shows up on an invoice and nowhere else — no artefact records which task the spend belonged to. Latency, on exactly the changes that should be instant. And depth mistaken for verbosity: a worker at maximum effort writes more, which reads like diligence and is mostly output.

What Majordomus does

A profile is a small file that sets each axis independently: capability class, reasoning effort, output verbosity, presentation, which context to load, what verification is required, and how often to checkpoint. Four ship with the tool — routine, implementation, debugging, deep-work — and a task names one at start. The generated instruction file tells the worker which profile is in force, so the rename runs under routine. Escalation, when a profile allows it, is a recorded event after a stated number of blocked attempts, not a mood.

Before and after

before   (whatever the last task left set)          rename SessionStore -> SessionRepo

after    majordomus start "rename SessionStore" --scope lib/session --profile routine
         majordomus check --explain
           profile routine  capability fast  effort low  verbosity terse
           context task, current_state

How to verify it

majordomus check --explain prints the effective axes for the active task. Start the same task under two profiles and the printed axes differ; the instruction file the worker reads states which profile is in force.

What it does not do

This is the honest part. Profiles are projected into the worker's instructions and validated as configuration. Whether the worker honours them is not observable from outside the worker, and Majordomus never selects or invokes a model. Every profile-related row in the claims matrix is marked advisory for exactly that reason. Measuring what a session actually consumed is on the roadmap and is not claimed today.

What this looks like

Concrete situations, one per audience. Each is declared in the moment's front matter, so the before and the after are data rather than prose a page could drift from.

  1. Left over from the last hard task solo-builder

    before
    Yesterday''s refactor needed the ceiling; today''s rename inherits it, and so does everything after it.
    after
    The rename runs under the `routine` profile: fast capability class, low effort, terse output, task and current state only.
  2. Fifty probes at the ceiling research-team

    before
    A sweep of fifty small experiments runs at maximum capability and effort because that is what the window was set to.
    after
    The sweep names a profile; the axes are set once, deliberately, and the expensive setting is the exception that had a reason.
  3. A budget with no dial to turn enterprise

    before
    Spend rises and the only lever anyone can describe is "use it less".
    after
    Capability class, effort, verbosity and context are four named fields in a profile a task is started under, and the profile in force is stated to the worker.

How you would know

The observable symptoms this moment declares. They are the questionnaire on the index and the input of majordomus why diagnose; nothing else defines them.

  • Nobody chose the model or the effort for the last task; it was whatever the previous task left set. never-chose-the-model
  • A trivial change came back with several paragraphs of explanation nobody needed. three-paragraphs-for-a-rename
  • There is no agreed cheap mode for routine work, so everything runs at the expensive one. no-cheap-mode

Where this lives in the tool

Everything below is read out of this moment's own front matter and resolved against the repository. A name here that did not exist would fail validation.

the commands that answer it

the capabilities of the executable that answer it

what it supervises — derived from the claims below

the claims that back this page, and the evidence behind each

the rules that govern it

  • majordomus.justified-escalation
  • majordomus.depth-is-not-verbosity
  • majordomus.minimum-sufficient-context
  • majordomus.profile-requirements

the use cases that show the way out

If this one is familiar, so is the next

What this moment names, what names it, and what shares its area, audience or tags. The second and third are derived; only the first is written down.

All 38, and how they connect to the tool →