agentic-praxis-grimoire

APG10 Karpathy Guidelines Evaluation and Selective Integration Exit

Phase identity

• Phase: APG10
• Date: 2026-07-19
• Exit sequence: 00015
• Disposition: Complete — targeted APG refinements validated
• Terminal result: targeted-apg-refinements-validated
• Controlling decision: ADR 0007, Accepted

Authority and objective

The human APG10 assignment accepts APG9 and authorizes a bounded external-source provenance review, current-owner comparison, frozen scenario evaluation, at most one correction per affected current skill, durable evidence, one private development commit and push, and linked managed reports.

The phase may not add a seventh skill by default, promote a skill, change public or reference repositories, alter global integration or RepoMap, restore Superpowers, add a dependency or generalized evaluator, publish a release, or start APG11.

APG9 acceptance

APG9 is accepted as Complete — v0.1 epic closed and v0.2 roadmap accepted. Public v0.1.0 and its squashed history, public-backed user-global integration, Superpowers retirement, bounded RepoMap smoke, APG8 acceptance, APG-TEST0, ADR 0006, the APG10 through APG14 sequence, the public wrapper limitation, and six provisional skills remain unchanged.

Source provenance and reuse result

The captured experimental file is byte-identical to multica-ai/andrej-karpathy-skills at immutable commit 64723a49ea6117894304eb491f0d32a60570bf45. The leaf and README declare MIT, but the exact tree contains no full license text, copyright notice, or notice file. The linked public observation was not retrievable in full during the phase.

Copied and adapted use remain blocked. The adopted correction is independently written synthesized expression from current APG ownership and frozen scenario evidence. No source heading, slogan, example, heuristic, template, sequence, or recognizable structure entered APG, and no guessed third-party notice was added.

Frozen baseline result

Twelve scenarios were fixed before any candidate wording: two independent positives, one non-trigger, and one edge in each of three families. Three fresh clean-context workers saw only current APG owners and their frozen family.

Family	Baseline disposition
Consequential assumptions and simpler alternatives	No material gap; both positives, non-trigger, and reversible edge passed
Objective-to-change traceability and speculative scope	No material gap; both positives, necessary-consistency non-trigger, and preserved-contract edge passed
Cleanup caused by the current change	Material gap; two independent positives were materially partial while the non-trigger and project-policy edge passed

Idea-by-idea disposition

Disposition	Material ideas
adopt-targeted-synthesis	Locally owned code or test artifacts made unnecessary solely by the current authorized change, when policy and removal authority are clear
already-covered	Proportionality; consequential assumptions and alternatives; accepted-objective scope; speculative-scope avoidance; unrelated-edit exclusion; observable completion; bounded planning; authority-limited progress
project-owned	Style and formatter policy; failure handling for allegedly impossible conditions
reject	Mandatory clarification for every uncertainty; universal 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 idea remains conditionally deferred on the current evidence.

Owner change and correction count

Only implementing-with-test-discipline changes. One coherent procedure correction makes locally owned code or test artifacts whose sole purpose ended because of the authorized change part of the coherent implementation slice when repository policy and removal authority are clear. The same step continues to exclude unrelated cleanup and new abstractions.

Correction counts:

• changed skill leaves: one;
• corrections to the changed leaf: one;
• second-correction needs: zero;
• unchanged current skill leaves: five;
• root universal changes: zero; and
• added skills or support directories: zero.

The root instructions, design, planning, review, debugging, and bounded- assignment procedures remain unchanged because the frozen evidence showed no gap in their ownership or behavior.

Candidate scenario and leaf review

Fresh candidate application made both current-change positives adequate. The unrelated-cleanup non-trigger and project-owned lifecycle edge did not regress. No over-trigger, mandatory reversible-uncertainty stop, universal cleanup rule, workflow chain, fixed cadence, fixed test order, size rule, policy import, or second correction was introduced.

A separate fresh non-author review accepted the complete resulting leaf with no blocker, material finding, minor finding, or correction. It verified valid frontmatter, preserved triggers and project parameters, proportional evidence, fresh-result completion, source-independent expression, and unchanged provisional maturity.

Artifacts

• ADR 0007;
• one public APG10 evaluation;
• reconciled README, bootstrap, provenance, roadmap, skill-index, ADR-index, and status-index owners;
• one corrected current skill leaf;
• five publication-excluded APG10 evidence records; and
• this exit.

Public artifacts do not link to publication-excluded evidence. Root instructions, executable tools, tests, dependencies, public release content, and historical exits remain unchanged.

Validation

Fresh validation from the complete resulting pre-review state established:

• exact captured and immutable public upstream blob identity;
• complete source-license and derivation classification;
• no identical eight-word source sequence in changed public Markdown;
• 22 of 22 Bats behavior families;
• 28 of 28 Python integration families;
• Bash syntax and help for all three commands;
• six canonical skills and six exact checked-in projections;
• six provisional catalog entries;
• one changed skill leaf, one correction, no seventh skill, and no support directory;
• candidate and baseline leaf hashes;
• ADR sequence 0001 through 0007;
• Markdown headings, fences, local links, and public confidentiality across 107 Markdown files;
• clean unchanged public, reference, public-backed global checkout, upstream evidence clone, and RepoMap tracked states;
• unchanged user-global link targets; and
• git diff --check; and
• git diff --cached --check.

Independent finished-diff review

A fresh reviewer that did not author candidate wording independently accepted the complete substantive staged state at raw staged-patch SHA-256 6347ede3f02a0f68e3a206ce034d872f6bd8a9f2a0e3bd8db3d3dfc17e8a70e0 with only this record integration required as follow-up. The reviewer reported no blocker, material, minor, or optional finding and independently verified:

• exact captured and immutable upstream blob identity and the conservative reuse classification;
• the frozen-baseline chronology represented by the durable evidence;
• one coherent correction, its before and after hashes, and no second- correction need;
• six canonical and projected skills, six provisional catalog entries, and no seventh skill or support framework;
• 22 of 22 Bats cases, 28 of 28 Python cases, command syntax and help, 107 tracked Markdown files, local links, confidentiality, and sequence counters;
• clean public, reference, global public checkout, and RepoMap tracked states;
• the APG11 stop boundary; and
• both staged and unstaged Git whitespace checks.

The same reviewer narrowly confirmed the record-only integration at raw staged- patch SHA-256 e37aa0d8d0480879d73259c30c5f775a30795fe62c6e162de87de2f9e73a1514. Only the two authorized records differed from the substantive reviewed state; all 187 local links resolved, the 16-file staged scope remained exact, no unstaged or untracked file appeared, and both Git whitespace checks passed.

Maturity and external state

APG retains exactly six skills and six projections. Every skill remains provisional; APG10 evidence is an APG13 input rather than promotion.

Public v0.1.0, the reference corpus, the public-backed global checkout and link chain, RepoMap, and Superpowers retirement remain unchanged. No operation targeted public release state, global integration mutation, RepoMap lifecycle, or Superpowers installation or restoration.

Limitations

• Scenarios are deliberate textual applications, not automatic-trigger or controlled comparative evidence.
• The source’s complete license and notice chain remains unresolved, so copied or adapted future use remains blocked.
• No executable fixture implements the frozen cleanup scenarios.
• APG10 does not validate distribution, publication, or stable maturity.
• Terminal delivery still requires one commit, push parity, and linked managed reports.

Next authorization

APG11 is the next accepted roadmap phase but has not started. It requires a separate human assignment. APG10 authorizes no successor work automatically.

External disposition requested

External review is requested to accept APG10 as Complete — targeted APG refinements validated; accept ADR 0007 and the recorded source dispositions; accept one implementation-discipline correction; retain exactly six provisional skills; preserve public, reference, global integration, RepoMap, and Superpowers state; and authorize no APG11 work automatically.

This site is open source. Improve this page.