ENGRAFT: ENgram GRadient Routing-Aware Fact Transplant

Abstract

Second public release. v0.1.0 wrote eight facts, one trigger each. This one writes a small world: one hundred facts about an invented character (Quail), learned from sentences that use them, and recalled by the real llama.cpp engine on sentences it never saw. What changed since v0.1.0 Usage-corpus descent. The variables are the 14,032 table rows read by the corpus; the loss is a language-model loss on the answers plus a KL term on everything else. The result is one 9 MB .pleo overlay. No weight is touched, and removing the file gives the base model back bit for bit. Numbers, real engine, free routing, 841 held-out test sentences: exact answer 0.841 with the overlay, 0.005 without. Second seed 0.873. Collateral damage on neutral text: mean KL 0.0131. Every figure in the README is recomputed from the file named next to it (scripts/check_readme_numbers.py, 39 checks). Three languages (English and Chinese preliminary, one seed each): 0.841 / 0.797 / 0.676. The ordering follows training mass per fact, not language. Row sharing, measured. Facts about the same subject share rows by content, not by hash collision; of the 37 failures lost to a sibling of the same subject, 21 went to the sibling with more training mass. Try it. A working one-command demo with token ids, and next to it a sentence that fails, on purpose. What this is, and what it is not. Token-addressed memory: zero context tokens, no retriever, one removable file. It is not general model editing and not a retrieval rival; the head-to-head against retrieval over the same corpus has not been run, and the README says so. Known weak spots, stated rather than hidden The composition probes (10 / 83) were evaluated with a 40-token budget that truncates every generation. The number is kept, flagged, and being re-measured. The v0.2.1 tag shipped a demo sentence that was one of the 134 failures. A reader found it; the fix and the reader's failing sentence are both in the README now. Open items are listed in docs/history.md: retrieval baseline, clean composition probes, a disjoint-row arm, LoRA and ROME/MEMIT comparisons. Thanks to the readers who read the code instead of the headline and reported what they found. Keep doing that. How this was built: code, experiments and write-up were produced in Claude Code sessions (Anthropic's Claude) driven by a single human operator who set the goals, approved every design step, ran the hardware and read every result.