The tag-readme trace read as a self-improving loop

Type: kb/types/note.md

The observed trace discharges causal connection. Read against self-improving system, the same change also discharges that definition, which asks for two things the reflective obligations do not: a change to the system's own organization, and a response to evidence about an objective it could have failed. Commonplace improves in the proposal-selection shape — draft a candidate, review it with a real chance of rejection, merge selectively — so that loop is the grid to lay over the trace. This note does the full mapping; the classification draws the partial-autonomy conclusion from it.

Requirement	In the trace	Runs in
Change to the system itself	The edit landed on kb/types/tag-readme.md, a self-representing artifact rather than ordinary content.	—
Search	Splits: a maintainer noticed the index type was doing two jobs and that the learning-theory head had outgrown its completeness claim — human, non-reflective. Formulating that into ADR 026's specific candidate ran through an agent retrieving stale indexes are worse than no indexes as self-representation and drafting the split around it — reflective, autonomous.	Split
Improvement objective	The bar the change could have missed: a marked head must not mislead a thorough reader, per stale indexes are worse than no indexes. ADR 026 makes it testable — complete is a mark the validator can falsify.	Human frames it; code checks it
Evaluation	Tests and the validator mechanically check that the marks are consistent; the judgment that the split was the right shape was the maintainer's.	Split
Operative retention	The three consumers — enforcement, routing, advice — keep acting on the change after merge.	Code (after human merge)

What the mapping settles, and what it does not

Naming the objective is the step the reflective reading never had to take, and it is what makes the loop improvement-directed rather than merely change-directed. A change loop with no objective can only report that something moved; this one can report that a specific bar — do not mislead a thorough reader — is being held, and can fail against it.

The honesty check lives here too: the validator passing means the marks are consistent, not that the type split actually made the KB better. Evaluating the objective is mechanical; evaluating whether the objective was worth adopting is not, and that judgment stays the maintainer's. The pass is a result; the improvement is still a claim.

The autonomy the mapping exposes

Two of the four rows split rather than land wholly on one side, and the autonomous halves cluster wherever a step routes through self-representation:

• Search splits. Noticing the strain and choosing the target was the maintainer's — non-reflective, human-inclusive. Formulating it into ADR 026's specific candidate ran through an agent retrieving and reading stale indexes are worse than no indexes as self-representation — reflective, autonomous.
• Evaluation splits the same way. The objective-check half is autonomous: because ADR 026 codified the objective into an enforced complete mark, the validator falsifies it on every run — validating an ordinary note now fails if a marked README sharing its tags is missing a member (test_note_target_also_validates_marked_tag_readmes). No human is consulted for that check. The shape-judgment half stays human: judging that the type split was the right shape is a step no validator can pass on.
• Retention is autonomous once the human merge lands. The constraint rejects future violations by itself, and an agent skips a search by itself; the merge installed the change, but its continued operation is the system's.

So the autonomy reaches wherever a step turns reflective — the agent formulating the candidate, the codified objective-check, the retention that enforces it — while noticing the strain and judging the fix's shape stay human throughout. Humans keep exactly the gates and the framing work where no adequate automatic check or retrievable claim reaches, since warranted autonomy is bounded by oracle domain.

---

Relevant Notes:

• Commonplace as a partially autonomous, reflective self-improving system — part-of: the classification this reading supports
• The tag-readme change as an observed causal-connection trace — part-of: the trace this reading interprets
• Where change candidates come from in Commonplace — part-of: surveys the wider set of candidate-forming mechanisms the agent's role in Search is one instance of
• Admitting a human into the boundary trades reflectivity for autonomy — grounds: why the agent-formulated half of Search counts as reflective and thereby autonomous
• Self-improving system — rationale: the definition this trace is read against
• A proposal-selection improvement loop requires search, evaluation, and operative retention — rationale: the search/evaluation/retention grid used here
• warranted autonomy is bounded by oracle domain — rationale: why the human gates sit where the oracle runs out
• stale indexes are worse than no indexes — rationale: the improvement objective the loop aims at