Skip to main content

Agentic

Research stops when attention does. These are the tools that keep it moving in the hours nobody is at the desk — and, more importantly, the line they are built not to cross. They run the work, gate it, sort its failures and report it. They do not choose what to run next, decide whether a result is real, or write a word of the record.

Every rule below was earned by being wrong. Each one carries the incident that produced it, because a rule without its failure is a slogan. The practices are collected on Practice; the research they serve is in the notebook.

The loop owns execution.

  • running pre-registered work
  • gating it before it touches the network
  • recovering it
  • reporting it
  • deciding when to wake someone

A person owns judgement.

  • Choose what to run next
  • Judge whether a result is real
  • Write a record entry
  • Close a ticket
  • Touch the write-up

What crosses the line: the loop wakes a person on exactly 4 things

  1. preflight refusesthe code is wrong; running wastes the night
  2. an unsorted triage linea failure class nobody has diagnosed yet
  3. a verdict outside expecta surprise — and surprises are the reason to run anything
  4. a job outlives its budget144 hours in a queue state that never raised an exception

Why the line sits here, and what it cost to learn: Practice

How a night runs

Everything else is logged and the queue carries on. A known failure needs a count; an unknown one needs a person.

  1. 1QueueOnly pre-registered work, with its expected verdict written down first.
  2. 2PreflightGates that refuse before anything touches the network.preflight refuses
  3. 3RunEach job under a budget it may not outlive.a verdict outside expecta job outlives its budget
  4. 4TriageEvery failure sorted by cause, so only the undiagnosed one needs a person.an unsorted triage line
  5. 5ReportA page readable mid-run, with a section on what it did not check.

The tools, the practice, and the instructions

$ python -m bearing ai/edit-slice

4 of 10 questions have an uncomfortable answer.

!!  Do the last 5 finished items measure your instrument, or the world?
      5 of the last 5 lean inward (E-030, E-031, E-032, E-033, E-034).
      asked because: a run of inward work is how a project stops being
      able to say its point

!!  What did you price and then not do?
      about a week — rewriting the probes is the work, scoring is hours.

bearing

A bearing is a measured direction, not a destination. It reads a project's own record and asks the questions that correct course — what has never been varied, whether the work has turned inward, whether the write-up still matches the findings. It measures drift. It has no taste, and says so.

projects/bearing/ · since 14 October 2026

Read about it

$ python tools/watch.py

[12:29:59] 1 job(s) queued
[12:29:59] E-036: preflight
[12:30:14] E-036: running (budget 3.0h)
[12:49:34] E-036: finished in 19 min
[12:49:34] E-036: ok — verdict ['UNDIFFERENTIATED'], report written

The unattended harness

Four pieces that let an experiment run without anyone watching: gates that refuse before anything touches the network, a triage that sorts every failure by cause so only the undiagnosed one needs a person, a report page readable mid-run, and a loop that escalates on exactly four things. It owns execution. A person owns judgement.

ai/edit-slice/tools/ · since 14 October 2026

Read about it

A green harness is not a
sound experiment.

E-036 passed every gate and
still measured two of six
criteria in ways that could
not answer their own
questions.

The practice, and its limits

The boundary between what a loop may own and what a person must; the twelve rules two projects learned by being wrong, each with its incident and what it cost; and an honest account of what none of it can catch.

Read it

# Manifesto
# Compass — keeping the work connected
#   to real need
# The Standard
# Side Builds — the deliberate exception
# Global Rules
# Agentic Work Template
# The Card — the default, and the width
#   limit

dotclaude

The instructions every Claude session here starts from. It opens with why the work exists and how to tell a real step from a toy, then sets the working rules: one card open at a time, tickets and threads for longer efforts, and when the heavier research protocols are allowed. It also marks the exception — side builds like this site, where the heavy rules are set down on purpose.

Rubix982/dotclaude · public on GitHub

Read it on GitHub ↗