Three independent gradings place a self-improving system
Type: kb/types/note.md · Tags: foundations, self-improving-systems
Membership in the self-improving system category is deliberately cheap. The definition asks only for operative change to the system's own organization, responsive to evidence bearing on an improvement objective — a plain gradient learner satisfies it trivially, with no self-representation and no evaluator anywhere in its mechanism. So the classification carries little design information on its own. The information is in where a system sits along three gradings, each answering a different question, each movable without the others.
- Retention form — how is improvement held? Operative, cumulative, or addressable, since reflection buys addressability. Free end: any retention that steers. Costly end: retention the loop can read, criticize, selectively revise, and transfer.
- Coverage — what is represented? Claimed per representational form and operation depth, since reflective coverage is graded across representational forms. Zero end: no behavior-bearing form is represented, as with the plain gradient learner. Free nonzero end: some behavior-bearing form represented somewhere. Costly end: coverage of every form the behavior spans, and of the mappings between them.
- Closure — what does the methodology settle? A methodology governs its own extension only as far as it settles the meta-decisions it raises. Free end: decisions improvised as they arise. Costly end: each meta-decision a change raises has a governed answer.
Autonomy — who performs a pathway — is a fourth, independent thing worth reporting for any given system, but it does not need its own placement scheme here. The base definition already grades autonomy pathway by pathway against a declared boundary. The human-boundary note shows why counting established humans as internal reflective components would make reflection cheap; it does not make autonomy derivable from reflection. So autonomy is reported with that boundary-relative grading when needed, not tracked as a fourth axis alongside these three.
The gradings move independently
A system can widen one axis without touching the others: widen coverage while keeping retention at the operative floor; settle more meta-decisions without changing what is represented; move retention from weights to artifacts without changing coverage. The axes are analytically distinct — a reading on one does not entail a reading on the others, though they may correlate at the extremes — which is what makes the gradings usable as a placement scheme: all three must be reported.
Each grading pairs a free attribution with a costly one
The pattern repeats across all three axes, and it is the reason the gradings rather than the membership carry the information. The floor is nearly universal where the axis has a positive floor: it retains, it has answers; coverage also names the explicit zero case before asking whether anything behavior-bearing is represented. The costly half names what can be trusted or built on: addressable, covered per form, settled. A report of the floor or free half is close to no information; assessment should ask for the costly half by name.
This is also why the definition keeps every axis out of membership: the reflective/non-reflective distinction is stated pathway-relative rather than as a membership test, autonomy is graded against a declared boundary rather than folded into reflection, and the gate architecture is scoped to the proposal-selection subtype. The axes are where systems differ, and folding one into the category would spend the vocabulary's discriminating power on a boundary that was never the interesting one.
Open Questions
- Whether the three axes are exhaustive, or further independent gradings (e.g. over search reach) await naming.
Relevant Notes:
- Self-improving system — defined-in: the base category the gradings place, and the autonomy grading this note explains why it isn't a fourth placement axis here
- Reflection buys addressability — grounds: the retention-form axis and its costly end
- Reflective coverage is graded across representational forms — grounds: the coverage axis
- A methodology governs its own extension only as far as it settles the meta-decisions it raises — grounds: the closure axis
- Admitting a human into the boundary trades reflectivity for autonomy — grounds: why human-inclusive reflection makes reflection cheap while leaving autonomy boundary-relative
- Gödel machines are a proof-governed case of reflective self-modification — exemplifies: an extreme where warrant and closure appear to tie together