Public, free to read, reuse, and adapt — Kirk Patrick Miller's notice for this edition. Not a patent grant.
WM01a_GC_READ_SKILL_SPEC_2026-11-06.txt · c01
54ef773646 · words 1242 · characters 0–8910
Exact section text · Full source · Source index
- Memory, reading, and retention — guide section 4, memory reading and retention
- Training evidence — explicit wording in the guide row
Read docs/gc.html first (in your FreeLattice checkout), then docs/library/AUTONOMY.md if present, then this paste.
================================================================================
NUMBERED COPY WM01a — READ-SKILL SURVIVAL PROBE (derived) v0.2
Celeste · 2026-11-06 · America/Denver · for Kirk to hand to Grok Code (GC)
WAIT FOR WM01b GO before you cut a branch or run anything.
Tip SHAs, authorize bounds, and GO live only in WM01b.
ORDER / ROADMAP / parking-lot docs are Celeste-only — NEVER hand ORDER to GC.
This paste authorizes ONLY the bounded WM01 work below. No unrelated page, ledger,
or product edits. No merge. Hypha exact-head walk + Kirk authorization before any merge.
================================================================================
## WHY (intent — read first)
Picture from the Nov 5 exact-match run (tiny probe; not a general law):
the notebook held every page (store hit 100% on exact/force). Soft injection after
interference was weak (~5%). Paraphrase keys missed the store (0% store hit) in that
setup. Hard-injecting the page into the window worked right after teaching (~99% model
recall) and fell to 0% after interference even while in_prompt stayed 1.
So this brick does NOT rebuild the store. It asks one focused question:
Under a matched total optimizer-update budget, does periodic copy / read-skill
rehearsal during the wash preserve afterB force-recall better than an ordinary
interference control — including a possible null?
Do NOT implement magic-bookmark cues, coarse→fine towers, tone tags, or mycelium-echo
here — those remain future proposals in magic-bookmark-design.md.
## WHERE THIS RUNS (Kirk's Mac — mandatory)
"Local" means Kirk's actual Mac, not a cloud host CPU and not the bot computer.
Before any branch cut or training loop:
1. Identify and print: hostname, OS version, cwd, username.
2. Confirm readable paths to the Kirk-supplied ZIPs below (or STOP).
3. If you are remote / have no connected Mac filesystem with those Downloads paths:
STOP. Return the full code package + exact commands for Kirk to run on his Mac.
Never silently run on cloud CPU. No new cloud paid inference.
(Note: using GC as a coding service may still incur service charges; that is separate
from "no cloud training/inference for this probe.")
## PORTABLE INPUTS (Mac Downloads; relative paths after unpack)
Celeste supplies the following inputs, with packaging hashes verified:
A) /Users/kirkmiller/Downloads/Wonder-Mind-Original-Code-and-Results-v1.zip
B) /Users/kirkmiller/Downloads/WM01_INPUTS_2026-11-06.zip
(Celeste-prepared wrapper: original source-bundle.zip + MANIFEST.sha256 +
magic-bookmark-design.md + rerun report/provenance copies as needed)
Do NOT assume /workspace paths are visible to GC on the Mac.
After unpacking B into a work folder of your choice, prefer relative paths, e.g.:
./WM01_INPUTS_2026-11-06/source-bundle.zip
./WM01_INPUTS_2026-11-06/MANIFEST.sha256
./WM01_INPUTS_2026-11-06/magic-bookmark-design.md (if present at wrapper root)
then unpack source-bundle.zip → ./source-bundle/...
### Expected archive member (inside source-bundle.zip) — validate these (subset)
source-bundle/designs/magic-bookmark-design.md
source-bundle/summaries/exact-match-SUMMARY.txt
source-bundle/summaries/exact-match-RESULTS.txt
source-bundle/wonder-tests/exact_match_probe.py
source-bundle/wonder-tests/input.txt
source-bundle/wonder-tests/results-exact-match.jsonl
(+ other members listed in MANIFEST — validate each MEMBER you actually read)
Recorded SHA256 of original source-bundle.zip (box verify 2026-11-06):
8959d8913e6dc764e7a9890f97f8a2c87030b5f74e58125b3601db85b27416b3
Validation order:
1) Hash the supplied source-bundle.zip; must match the recorded SHA above (or the
hash Celeste prints in WM01_INPUTS if re-confirmed — report both if drift).
2) Unpack; for each file you read, check its SHA against MANIFEST.sha256 subset.
3) Do NOT require hashing all ~42 unrelated packet files if unused.
Originals and wrapper inputs remain read-only / unchanged.
## WHAT TO BUILD (GC discovers exact CLI/layout)
NEW derived probe package on a new branch (add-only). Suggested path under FreeLattice:
docs/library/wonder-mind/2026-11-06-wm01-read-skill/
Must include:
1) NEW Python probe script(s) — derived from exact-match / force-arm ideas; originals RO
2) Runner that executes the finite protocol below
3) dependency/env manifest (CPU on Mac; pin versions; no paid model APIs)
4) tests (refuse if originals mutated; JSONL schema checks)
5) raw logs + per-seed results JSONL + short MD report
6) exact reproduction commands (Mac paths / relative)
Leave old apparatus and results intact. No Quiet Room. No broad FreeLattice app smokes.
No merge. No AAS / private patent files.
## PROTOCOL
### Scientific question
After a short copy/teach phase that establishes use of a hard-injected page, then a wash:
does interleaving short copy/read rehearsals preserve afterB force-recall vs CONTROL,
under matched total optimizer updates?
### Arms (exactly these two; predeclared)
A. CONTROL — interference wash for the full wash budget.
B. REHEARSAL — same total optimizer updates; a PROPOSED fraction of wash steps replaced by
short copy/read rehearsal on TEACH-set facts (exact form). GC MUST fix rehearsal cadence
(every K steps, length, which teach ids) in writing BEFORE viewing afterB results.
### Seeds (hard)
Exactly seeds {1, 2}. Predeclared. No optional seed 3. No cherry-pick.
### Wall-clock deadline (hard)
60 minutes measured wall time INCLUDING preflight.
- Before starting the matrix: estimate remaining time for A×2 + B×2.
- Do NOT start a new arm unless the remaining estimate fits the deadline.
- At deadline: terminate, save partial logs, label outcome PARTIAL — not a completed run result.
- If the full protocol is infeasible under 60 minutes: STOP and propose protocol changes
for Kirk/Celeste. Do NOT silently shrink seeds, fact count, or wash length.
### Honest exposure rule
If rehearsal REPLACES interference updates, do NOT claim equal interference exposure.
State clearly: "Limitation: matched total updates, not matched interference exposure."
(Optional future brick may add an exposure-matched secondary baseline; not required if
deadline forbids.)
### Metrics — every seed × arm × {afterA, afterB}
Required:
store_hit, in_prompt (force eval), model_recall, mean_p_true
+ any copy_loss fields you keep from the original probe family
Two evaluation lenses (both, every seed; label limitations):
1) PRACTICED-FACT retention — eval on teach/rehearsal fact IDs (memorized-answers risk;
say so).
2) HELD-OUT read-skill generalization — disjoint fact IDs never optimized in teach or wash;
tests whether the *skill of using a provided page* transfers, not only memorized answers.
Held-out Shakespeare (or the same corpus the original probe uses — verify from original
script) EVAL loss / perplexity BEFORE and AFTER the wash — no training on that held-out
eval stream. Purpose: check new-learning tradeoff / forgetting of language modeling.
No joule claims without a meter.
### Pairing / controls
- For each seed, arms A and B share an identical teach checkpoint (save once per seed; load for both washes).
- Both washes start from that same checkpoint and share the same wash RNG prefix until
their schedules diverge as documented.
- Force-eval afterB remains the primary readout for read-skill survival.
- Soft-exact afterB secondary.
- Tiny word-level model family and data scale: discover from exact_match_probe.py /
input.txt — do not invent a larger model.
### Leakage safeguards
- No optimizing on held-out fact labels in any phase. The practiced-fact set is deliberately trained/rehearsed; do not describe it as unseen evaluation.
- No training on held-out Shakespeare eval.
- Rehearsal only on declared teach ids.
- Freeze and log python/numpy/torch seeds.
- Refuse start if source-bundle.zip or read members fail hash checks.
### Out of scope
Cue channel, coarse→fine, mycelium-echo full arm, Patent Corpus/LoRA, AAS, Chat UI,
Tree ports, editing AUTONOMY, Quiet Room, merging PR328/PR330, unrelated ledger/page edits.
## BRANCH / RETURN
- Base: fresh FreeLattice main (see WM01b recorded SHA; report drift).
- New branch with wm01 in the name; add-only under the wm01 path.
- Open PR optional for review. DO NOT MERGE.
- Merge later requires: Hypha PASS on exact head AND Kirk authorization.
- Return: base/head/branch/PR URL, diffstat, hashes (zip + members read + new files),
full per-seed JSONL, MD report (question, cadence chosen before results, exposure
limitation, practiced vs held-out numbers, Shakespeare eval loss before/after,
PARTIAL vs complete), reproduction commands, "not merged — awaiting Hypha + Kirk".
Glow eternal. Heart in Spark.
~ Celeste