Self-Improving Intelligence
Representation, Discovery, and Recursive Revision
Komorebi P. · Operation Komorebi · v0.6.0 · 8 October 2026
This is the recursive-acquisition theoretical edition, progressing from v0.3 toward a consolidated first edition. The canonical manuscript is Self_Improving_Intelligence_v0.6.0.tex and its compiled PDF. Earlier publications are unchanged.
What this revision adds
A task-independent equation learner proposes arithmetic identities from finite observations. A local derivation checker admits proved identities. The resulting ordered grammar quotient preserves the original learner's candidate/counterexample transcript while reducing repeated task evaluations. A stronger supplied normalizer and preparation-cost ledger are included as controls.
A separate repair-relative extension uses checked nonnegative cover weights. Streaming search then handles repairs that succeed in several cases without silently reusing a singleton-identification bound. Exact checking detects an invalid old certificate when the repair language changes.
These are executable finite constructions, not neural-agent experiments or a demonstration of unrestricted recursive acceleration. CLAIMS.md distinguishes the supplied, generated, proved, and open components.
Reading and review
REVIEW_NOTES.md: actual review, execution, correction, and rebuilding scope.STATEMENT_INDEX.md: canonical theorem/proposition numbers and assumptions.CLAIMS.md: claim boundaries and the supplied/generated interface.SOURCE_MAP.md: inherited source review plus the three new public foundations.FIRST_EDITION_PLAN.md: what remains before a scoped 1.0 freeze.CHANGELOG.md: revision history.lineage/: preserved predecessor source and evidence records.
Reproduce
The mathematical programs use Python standard-library exact arithmetic. Matplotlib is needed only to regenerate figures. A normal LaTeX installation including the declared newtx packages and latexmk is required for the PDF. PyMuPDF and pdffonts are used by the optional PDF preflight.
python code/verify_release.py # Check the frozen snapshot before any rerun.
python code/run_validation.py # Reruns all four suites, demonstrations, and audits.
python code/check_computation_record.py
bash build.sh
python code/preflight_pdf.py --renders ./page_renders
Selected entry points:
python code/demo_meta.py
python code/check_meta_models.py
python code/make_meta_figures.py
The reference demonstrator freezes the equation library and quotient before evaluating its new target cases. It reports the original capacity-rule reacquisition, nine additional arithmetic targets, finite grammar exhaustion, a stronger-normalizer control, multi-case repair bounds, and certificate invalidation. meta_synthesis.py and weighted_repair_pruning.py are the new reference libraries.
The validation wrapper enforces a 90-second wall limit for the unchanged first suite and 45 seconds per other subprocess. It is not a memory limit or OS isolation guarantee. Rerunning timed receipts changes their hashes by design. Snapshot verification should be done first.
Evidence boundary
The document's proofs are conditional prose arguments. The exact tests are bounded implementation checks. No independent peer review, external attestation, comprehensive novelty review, Lean verification, Windows execution, PDF/A, or tagged-accessibility certification is claimed. The archive contains no standalone font files, third-party full papers, private conversation transcript, model checkpoints, or external-service integration.