research-ops

The state layer of a research pipeline. Four artifacts — cursor, push log, register, TODO — plus one habit: say what changed and what it cost.

This skill produces no research content. No cards, no matrices, no proposals, no verdicts. It keeps the bookkeeping that makes those possible, and it is the only skill allowed to write to the state files.

What it owns

The rule that makes the layer work: each fact lives in exactly one file, and other files carry a pointer, never a copy. Two copies of a cursor diverge within a week; a pointer cannot.

→ references/state-files.md for fields, update protocol, and the hash-check that proves read-only sources are untouched.

Delivery format

Every run ends with three lines, in this order:

files:   <what changed, one line>
next:    <the single next action, not a plan>
new:     <items newly registered, with IDs; "none" is a valid answer>

And when the run defends a claim it also carries three lines of evidence — what is being defended / what the evidence is / what remains uncovered — with speculation excluded and unverifiable items booked into the register. This is not ceremony: it is the only place where "I checked" becomes auditable.

→ references/delivery-format.md for the ledger that backs the second triple, and the rule for what may not appear in either.

Disagreement is part of the delivery

A run that only reports what it did is a run that cannot be argued with.

1. Every delivery carries a final section that names what the executor disagrees with — the instruction as given, the pipeline itself, or the last ruling — with the reason, in the same place every time.
2. "None" is a valid answer and must be written out. An absent section and a section saying "none" mean different things, and only one of them is trustworthy.
3. This section is an artefact of its own, not a fourth delivery line: it is not folded into files / next / new, because a disagreement about the instruction is not a fact about the files.

The failure this prevents: a pipeline where the only recorded objections are the ones that were already resolved, so every past decision looks unanimous in hindsight.

Stopping

Five conditions stop work; everything else continues and gets registered:

Hit one → mark [BLOCKER], stop that thread, continue the rest. Nothing else is a blocker; calling a preference a blocker is how a pipeline loses a day.

→ references/blocker-levels.md — including why B1–B5, P-1–P-4, and the M-type tags must never be converted into each other.

Registering

Two layers, two permissions:

→ references/ledger-rules.md for the boundary cases, including the one that breaks most pipelines: a same-day ruling that starts as a registration.

Does not trigger

Three prompts that look adjacent and must go elsewhere. These are the regression fixtures for this skill's description; if a change makes any of them route here, the change is wrong.

The last one matters most: research-ops is not a note-taking skill, and a pipeline that treats it as one ends up with state files nobody trusts.

Routing