agentic-praxis-grimoire

ADR 0007: Experimental Karpathy Guidelines Disposition

Status

Accepted

Date

2026-07-19

Acceptance authority

The human maintainer’s APG10 assignment accepts APG9 and authorizes this bounded source evaluation, one evidence-supported correction to an existing owner, private development commit and push, and managed reports. It does not authorize a seventh skill, maturity promotion, public or reference mutation, global integration change, Superpowers restoration, release publication, or APG11.

Context

APG9 closed the v0.1 epic, accepted ADR 0006, retained six provisional skills, and identified three hypotheses from an experimental coding-guideline source. APG10 must establish provenance, test current APG behavior before changing an owner, and disposition every material source idea without treating source packaging or familiarity as adoption evidence.

The captured source is byte-identical to multica-ai/andrej-karpathy-skills at immutable commit 64723a49ea6117894304eb491f0d32a60570bf45. Its leaf and README declare MIT, but that revision contains no full license text, copyright notice, or notice file. The source links an Andrej Karpathy public observation, but its complete body and reuse terms were not independently available during APG10. Copied or adapted use therefore remains blocked.

Decision

Evaluation method

APG10 froze twelve public-safe scenarios before any candidate wording: two positive cases, one non-trigger, and one edge for each of three families. Fresh concealed-source workers applied only current APG instructions and owners. A material gap required two independent materially partial positives, one high- risk weakness, or repeated real-project omission evidence.

Families A and B had no material gap. Current design behavior surfaced consequential assumptions and compared credible alternatives while keeping straightforward and low-risk reversible work proportional. Current planning, implementation, and review behavior traced accepted work semantically, rejected speculative frameworks and adjacent refactors, and retained necessary contract, consistency, documentation, and evidence surfaces.

Family C met the threshold. In two independent cases, current implementation text permitted locally owned implementation or test artifacts made unnecessary solely by the authorized change to remain while behavior and tests still passed. The same current text correctly excluded unrelated pre-existing cleanup and preserved generated, migration, compatibility, and external lifecycle policy.

Idea dispositions

Source idea Disposition
Proportional caution for non-trivial work already-covered
Consequential assumptions, interpretations, and tradeoffs already-covered
Comparison of a materially simpler viable option already-covered
Mandatory clarification for every uncertainty reject
No behavior beyond the accepted objective already-covered
Avoid speculative features, configurability, and abstraction already-covered
Prohibit every single-use abstraction reject
Omit handling for scenarios considered impossible project-owned
Fixed line-count rewrite heuristic reject
Intuition-based seniority test for complexity reject
Avoid unrelated adjacent edits and refactors already-covered
Always preserve existing local style project-owned
Always report unrelated dead code reject
Resolve locally owned artifacts made unnecessary solely by the current change adopt-targeted-synthesis
Literal line-by-line mapping to user wording reject
Observable success and resulting-state verification already-covered
Universal tests-before-and-after recipe reject
Checked plan for every multi-step task already-covered
Independent looping until success criteria pass already-covered

The already-covered plan and autonomy ideas retain current trigger, authority, and stop boundaries. Their source’s universal forms are not imported. No idea requires conditional deferral on the current evidence.

Adopted correction

APG10 makes one coherent correction to implementing-with-test-discipline. Its narrow-implementation step now includes locally owned code or test artifacts whose only purpose ended because of the authorized change, but only when repository policy and removal authority are clear. Unrelated cleanup and speculative abstraction remain outside scope.

Fresh application to the frozen Family C cases made both positives adequate, preserved the non-trigger and edge, and introduced no universal cleanup mandate, new authority, fixed workflow, test-order rule, or mandatory stop for low-risk reversible progress. A separate non-author leaf review accepted the resulting leaf without finding or second-correction need.

Derivation and notice

The correction is independently written APG-native synthesized expression based on current ownership, the frozen scenario gap, and repository boundaries. APG10 copies or adapts no source heading, slogan, example, heuristic, template, sequence, or recognizable structure. The source materially influenced the evaluation, so public provenance records that influence and the unresolved reuse boundary. No third-party notice is added because no copied or adapted material entered APG and the source supplies no complete notice payload to preserve.

Existing ownership and skill count

No root instruction, design, planning, review, debugging, or bounded-assignment procedure changes. Their frozen cases either passed or did not own the demonstrated gap. The implementation leaf remains the existing owner. APG retains exactly six direct-child skills and six matching discovery projections; all six remain provisional.

Alternatives considered

Add the source as a seventh skill

Rejected. Its themes overlap current owners, its absolute rules conflict with proportionality or project-owned policy, and its reuse boundary is incomplete.

Add cross-cutting root rules for assumptions or traceability

Rejected. Families A and B passed under current owners. Root restatement would duplicate trigger-specific procedure without a demonstrated universal gap.

Make no skill change

Rejected after baseline. Family C met the two-independent-scenario material-gap threshold, and one existing owner could correct it without over-triggering.

Add a general cleanup rule or deterministic validator

Rejected. Cleanup requires ownership and policy judgment, and the demonstrated gap is narrower than a repository-wide mandate or stable mechanical invariant.

Copy or adapt the external wording under its MIT declaration

Rejected. A label and README assertion are not a complete license and notice payload, and independently written synthesis is sufficient.

Consequences

Rollback

If later evidence shows over-triggering or policy drift, revert the one implementation-procedure sentence, restore the baseline leaf hash, rerun the Family C scenarios and complete repository gates, and supersede this decision without deleting its evidence. The six projections require no separate content rollback because they resolve to the canonical leaves.

Non-goals

APG10 does not promote a skill, create a support directory or evaluator, change tests or dependencies, define universal style or error-handling policy, modify public or reference repositories, alter global integration or RepoMap, restore Superpowers, publish v0.2, or begin APG11.

Deferred decisions