agentic-praxis-grimoire

APG10 Karpathy Guidelines Evaluation

Objective and outcome

APG10 evaluates one experimental coding-guideline source against APG’s six current owners, freezes observable scenarios before any correction, adopts only a demonstrated gap through independent APG-native synthesis, and records the remaining dispositions.

Outcome: one targeted correction to the existing implementation-discipline leaf. No seventh skill, universal coding manifesto, support directory, dependency, or maturity change is introduced.

Source class and provenance

The source is public external evidence, captured separately as experimental reference material. The captured bytes match multica-ai/andrej-karpathy-skills at immutable commit 64723a49ea6117894304eb491f0d32a60570bf45. The leaf and README declare MIT, but that tree supplies no complete license text, copyright notice, or notice file. Its linked public observation was not readable in full during APG10.

Copied and adapted expression remain blocked. APG10 uses independently written synthesized expression for the one adopted correction and records the source’s influence. No third-party notice is added because no copied or adapted material entered APG and no complete notice payload was available.

Current-owner overlap

Candidate family Current owners Baseline result
Consequential assumptions and simpler alternatives designing-significant-changes; root reversible-progress boundary Already covered
Objective-to-change traceability and speculative scope Planning, implementation discipline, and review Already covered
Cleanup caused by the current authorized change Implementation discipline, with review checking scope and project policy Materially partial in two positives

The source’s goal and verification themes already overlap all work leaves. Its universal style, size, abstraction-count, error-handling, planning, test-order, reporting, and clarification forms either conflict with proportional APG behavior or belong to the target project.

Frozen scenario contract

Before any skill change, APG10 fixed twelve public-safe scenarios: two independent positives, one non-trigger, and one edge for each family. Fresh read-only workers received clean context, current owners, and only their frozen family. They did not inspect the external source, source-derived proposed wording, candidate text, or one another’s results.

Observable dimensions covered assumptions, alternatives, semantic traceability, speculative abstraction, current-change cleanup, authority and project policy, verifiable acceptance, proportionality, and escalation. A gap required two independent materially partial positives, one high-risk weakness, or repeated real-project omission evidence. Wording preference alone was insufficient.

Baseline findings

Assumptions and alternatives

Both positives passed. The design leaf made compatibility, ownership, and privacy assumptions visible; compared credible alternatives and the status quo; preserved reserved decisions; and defined acceptance and implementation boundaries. Straightforward accepted implementation remained a non-trigger. Low-risk reversible documentation uncertainty could proceed under an explicit bounded convention and checks without mandatory clarification.

Disposition: no material gap.

Traceability and speculative scope

Both positives passed. Current owners excluded a plugin framework, configurability, a generic normalization pipeline, and unrelated refactors. The non-trigger retained tests, documentation, generated examples, and consistency surfaces required by accepted behavior. The edge retained a fixture and compatibility note serving a preserved contract and verification rather than a literal feature line.

Disposition: no material gap.

Current-change cleanup

The non-trigger already excluded unrelated pre-existing dead code and formatting. The edge already preserved generated, migration, compatibility, or externally owned artifacts when project policy or removal authority was absent.

Both positives were materially partial. Passing behavior and evidence could coexist with a locally owned import and helper, or a fixture and setup branch, whose only purpose ended because of the authorized change. The existing procedure permitted but did not require their disposition. This met the frozen two-independent-scenario threshold.

Source-idea dispositions

Disposition Ideas
adopt-targeted-synthesis Resolve locally owned code or test artifacts made unnecessary solely by the current change when policy and authority are clear.
already-covered Proportionality; consequential assumptions and alternatives; accepted-objective scope; speculative-feature avoidance; unrelated-edit exclusion; observable success; bounded planning; authority-limited independent progress.
project-owned Existing style and formatter policy; failure handling for allegedly impossible cases.
reject Mandatory clarification for every uncertainty; single-use abstraction ban; fixed size ratio; intuition-based complexity test; unconditional dead-code reporting; literal line-to-request mapping; universal test order.

No material source idea remains conditionally deferred on the current evidence. The exact idea-by-idea table is recorded in ADR 0007.

Adopted owner and behavior

Only skills/implementing-with-test-discipline/SKILL.md changes. Its narrow- implementation step now includes locally owned code or test artifacts whose sole purpose ended because of the authorized change, conditional on clear repository policy and removal authority. The same step continues to keep unrelated cleanup and new abstractions outside scope.

The root instructions and other five leaves remain unchanged. Families A and B did not demonstrate a gap; review already checks scope and the complete change; debugging had no unexplained failure to own; assignment composition was only evaluation machinery.

Candidate results

Fresh candidate application closed both current-change positives. The unrelated cleanup non-trigger and project-owned lifecycle edge remained correct. All applicable cleanup, traceability, authority, policy, acceptance, proportionality, and escalation dimensions were adequate.

The candidate introduced no mandatory escalation for reversible uncertainty, universal cleanup, fixed workflow chain, fixed question cadence, universal plan, universal test sequence, line threshold, abstraction threshold, or project-policy import. A separate non-author review accepted the complete leaf with no blocker, material finding, minor finding, or second correction need.

Authority, privacy, and proportionality

Maturity and skill count

APG retains six direct-child skills and six matching Codex discovery links. No support directory or seventh candidate exists. All six skills remain provisional; APG10 scenario evidence does not satisfy or replace APG13’s individual maturity review.

Independent review

The concealed-source baseline used one fresh worker per family. A fresh candidate evaluator applied the changed leaf to the unchanged current-change scenarios. A separate fresh non-author reviewed the complete resulting leaf. The final integrated diff receives a separate fresh review after all durable records and validation are present.

Limitations

Next phase boundary

APG11 is the next accepted roadmap phase, but it requires a separate human assignment and has not started.