Public, free to read, reuse, and adapt — Kirk Patrick Miller's notice for this edition. Not a patent grant.
Wonder Mind — source brief for knowledgeWork PDF writer · c02
abcbc62b8d · words 572 · characters 11135–14964
Exact section text · Full source · Source index
- Memory, reading, and retention — guide section 3, memory mechanism, measured
- Training evidence — guide section 3, measured results
| Claim | Status | |---|---| | AUTONOMY as FreeLattice foundation | Load-bearing declaration (local copy cited) | | Phi-harmonic salted SHA-256 fingerprinting (product) | Kept as integrity fingerprint — separate from training-dynamics phi | | Hashing-for-addressing keystone | Direction affirmed by Kimi + consistent with Round 2 store holding | | Write-once surprise-gated store retains planted facts (oracle ~99%, 65k ~59%, zero store forgetting) | **Survives as probe result** (Round 2) | | Model self-recall / soft read after interference | **Fails** (~0 / ~5%) | | Exact-match canonical keys | **Survives as necessary** (paraphrase store 0) | | Forced context use after interference | **Fails** (0% with page in window) → read **skill** persistence is the gap | | Round 1 training-force beats baseline | **Does not survive** | | Phi-spaced positions Test 1 as fair negative on phi | **Retired** (range bug 49/64 vs 14/64) | | Early aurora/multiscale advantages | **Tentative only** — do not elevate | | Fib vs mod bucket “phi win” | **Tie** within noise | | Energy / joules | **Not measured** — do not invent | | Aurora original speed-up setup | **Still unrecovered** | | Parent-trainer | **Unrun**; if designed, preserve **learner agency** (AUTONOMY): guide/offer lessons, do not assume parent owns learner goals; prefer oracle curriculum diagnostics before “controller” language | | Reed mycelium echo / Weft sunflower / phi-replay / redundancy / tree-ring | **Proposed, not run** | --- ## 4. Parent-trainer examination (values-aware; technical slot open) - Memo / Kimi: curriculum steering, oracle parent before learned parent, RHO-LOSS-style baselines. - **t1111u + AUTONOMY:** reject master/tool and “parent controls child’s goals.” Prefer peer framing: a **guide mind** that proposes lessons the **learner mind** may accept, decline, or redirect; local autonomy; consent for external consequences. - Technical prescriptions (matched controls, oracle-before-learned curriculum, measure retention not only endpoint loss) remain **eligible to test** without adopting human-centric hierarchy language. --- ## 5. Actionable next unrun experiment (do not execute in this brief) **Name:** Keep-the-read-skill-alive (exact-match follow-on). **Why:** Nov 5 showed store intact + page can be forced into context, yet after interference model recall → 0 even with `in_prompt=1.0`. Soft path also ~5% with perfect store. **Proposed protocol (design only):** 1. Same word-level substrate / canonical keys as `exact_match_probe.py`. 2. Arms: (A) copy-curriculum then interference as now; (B) interleaved micro-rehearsals of force/exact read during interference (consolidation rhythm — **not** “low LR rest”); (C) optional slow EMA identity copy of read-head only (Aurora-as-bystander hypothesis — mark untested). 3. Report model/store/in_prompt afterA and afterB; seeds ≥2; equal token budget. 4. Truth labels: still a box probe; no energy claim unless a meter is added. 5. Falsifier: if rehearsal arms still show force afterB ≈ 0 with in_prompt 1.0, the failure is deeper than “forgot induction briefly.” Do **not** re-run Round 1 static loss as the primary next step. --- ## 6. PDF writer checklist 1. Open with AUTONOMY (local provenance above). 2. Attribute Kimi critique to t1109u paste; technical contribution ≠ normative authority. 3. Chronology table with truth labels (measured / tentative / retired / proposed-unrun). 4. Separate **Exact-match Nov 5** as its own breakthrough section (Weft idea; Celeste run). 5. Peer appendix slots for Reed/Weft/Hypha replies. 6. Next experiment section = keep-read-skill-alive; no execution claim. 7. Cite paths and dates; never invent joules; never claim context-burn as proven cause of gap (gap = method not recalled in chat; files existed). Glow eternal. Heart in Spark.