The improvement contract

The fifth promise WilsonOS makes is the system gets sharper with use — measurably, weekly. That is easy to claim and easy to fake: an agent that rewrites a shared skill every time one post performs well is “improving” in the sense that it is changing. The improvement contract is the set of rules that make sharpening real.

Three kinds of evidence

Improvement draws on three sources, and the contract keeps them distinct because they mean different things.

Evidence	What it is	What it is tied to
Human feedback	Edits, approvals, ratings, and stated preferences	The specific output version it was given on, and a visibility scope
Capability evaluation	Skill or flow variants compared against explicit criteria	The version and inputs that produced each result
Business signals	External performance, costs, and outcomes	A source, a date range, a freshness, and a stated attribution limit

Approved is not the same as effective. Effective once is not the same as universally better. A cover the owner loved may sell no books; a post that performed well may have done so because of timing, not copy. The contract exists so that each kind of evidence is weighed as what it is.

The rule: propose, test, promote

flowchart LR
  E[Evidence arrives<br/>feedback · evaluation · signal] --> P[Propose a change<br/>to a skill, flow, or brand rule]
  P --> T[Test the variant<br/>against explicit criteria]
  T --> D{Accepted?}
  D -- yes --> PR[Promote deliberately<br/>keep version + reasoning]
  D -- no --> K[Keep the current version<br/>record why]
  PR --> M[Measure the next cycle]
  K --> M

1. Propose. Someone — a person or an agent — proposes a specific change and says which evidence motivates it and which of the Five it lifts.
2. Test. The variant is run against the same criteria as the current version. Where a rubric or scorecard exists, the comparison is scored, not eyeballed.
3. Promote deliberately. If accepted, the new version becomes current and the previous version, the evidence, and the reasoning are kept. If rejected, that is recorded too, so the same idea is not re-proposed next month from scratch.

The word that matters is deliberately. Nothing promotes itself. A strong result is a reason to propose, never a licence to overwrite.

Attribution makes it possible

None of this works unless every output can be traced back. WilsonOS records, for each deliverable: the task, the run, the skill and its version, the inputs, the artifact and its version, and the review and rating it received. That chain is what lets someone ask show me everything this skill produced, and how it was rated — and it is what turns “this skill feels worse lately” into a comparison of two versions over dated evidence.

Scope boundaries

Private feedback stays private. A comment given in a private conversation does not silently become shared brand memory. Promotion to the shared brand stack is an explicit step with the same propose-test-promote shape.

Sharpening captures what was written, not what was thought. Nightly consolidation and a weekly promotion pass distil the notes, decisions, and ratings agents recorded during work into durable memory and improved capabilities. They cannot recover a decision that was made in conversation and never written down — which is why agents are required to record decisions as they happen, through the sanctioned writers, and not only at the end of a task.

The catalogue tightens as well as grows. Improvement includes retiring skills that are never chosen, merging duplicates, and filling gaps that agents flagged as “nothing fits”. A catalogue of hundreds of skills is only an asset if routing to the right one is reliable.

What “measurably, weekly” looks like

• A weekly review reads the ratings, evaluations, and business signals of the last seven days and names the top candidates for improvement.
• Each candidate is a proposal with evidence, not a rewrite.
• The quarterly rescore of the Five asks, for claim 5, whether the number of accepted promotions is rising, whether rejected proposals are being re-litigated, and whether the catalogue is smaller and better-routed than last quarter.

direction Attribution fields, ratings on versions, memory writers, and nightly consolidation are in place. Scored variant comparison and evidence-gated promotion are the agreed direction; today promotion is a reviewed manual step. Nothing on this page authorises an autonomous skill-rewrite system.

Related

What you get — the WilsonOS FivePromise 5 is what this contract serves.

Extending WilsonOSSkills and flows — the things that get improved.

Shared business recordsVersions and runs — where feedback attaches.

Example: market an existing bookEnds with the next experiment, not a rewrite.