Public, free to read, reuse, and adapt — Kirk Patrick Miller's notice for this edition. Not a patent grant.

Weft — a continuity-first memory and plasticity plan · c02

P-9f9b924662 · words 1564 · characters 10059–21438

Exact section text · Full source · Source index

Archive pages whose extracted text matched this section: 33, 34. Unmatched pages stay on the source record with the checked page method. They are not guessed into a section.

Memory, reading, and retention — guide section 4, memory reading and retention
Training evidence — explicit wording in the guide row
Continuity — explicit wording in the guide row

**Mechanism.** Use one versioned canonical serializer for structured IDs on write/read. Keep logical identity separate from content version: correcting a value creates a new record plus a supersession link; it does not rewrite history. Store full SHA-256 and payload length; short sha8 can remain a display/bucket hint, not the sole collision authority. On short-key collision retain both entries and compare full IDs/digests; if logically ambiguous, return ambiguity rather than a wrong value. Record algorithm/schema versions explicitly. NFKC, case folding, whitespace removal or clipping can merge meaningfully distinct inputs; use them only in a declared alias layer, preserve originals, and do not assume normalization is semantically lossless. The 068m normalization clips at 2,000 characters and has algorithm variants: these need explicit compatibility treatment, not blind transplant [S9 L82–105].

**Small falsifiable diagnostic.** Proposed static round-trip audit of 100 deliberately distinct synthetic records; 20 paraphrase/alias questions scored separately; 10 predefined negative/collision/normalization boundary cases. It checks exact byte/digest readback, intended version, and explicit miss/ambiguity on invalid input. Controlled equal-prefix keys are fixtures for collision handling, not claims about observed random collisions. Route aliases via a declared rule/candidate map first; no model training or embedding service needed. This does not rerun the old model probe.

**Separating outcomes/limits.** Exact fails → serializer/index/integrity work, before reading research. Exact passes but alias routing fails → locator problem; do not blame the store or retrain the reader. Exact hashes are not universally necessary for memory systems; consistent identifiers suffice for this controlled setup. Hash verification proves byte agreement with a trusted manifest, not truth, authorship, or security against a malicious replacement manifest. A public salt is not authentication [S4 L65–69].

**Metric/gate.** Exact correct record+bytes /100, error correctly signaled /10, alias intended candidate /20, false-confident alias answers /20. Engineering go: 100/100 exact, 10/10 safe negative responses, zero silent collisions. Routing remains a separately labeled exploratory result; proposed alias deployment gate 20/20 with zero false confident routes on this small fixed set, not proof of robust semantic search. “Near-perfect” remains a reliability target, not universal perfection; see statistical limits below.

**Design cap.** 60 CPU-seconds, 2 wall-minutes, 256 MiB RAM, 5 MiB new artifacts; 130 checks, zero optimizer steps, zero model calls. Failure or inadequate artifact access stops this stage.

### Stage 2 — Actual context delivery, not a retrieval success flag

**Mechanism.** Build a receipt at the final model-call boundary: retrieved full digest/version → rendered page → tokenizer/version → complete input-token IDs, page and query spans, attention-mask visibility, length, truncation direction, reserved answer space. Do not log private content unnecessarily; these diagnostics use synthetic pages only. If a required page cannot fit, fail visibly or request a bounded smaller page, never silently substitute a summary for exact text. Verify that the value belongs to the requested record, not merely that the same token occurs elsewhere.

**Small falsifiable diagnostic.** Twenty-four declared layouts across short/edge-of-window/overlong input, formatting delimiters, distractors, competing versions and query positions; apply to up to two existing checkpoint configurations, at most 48 assembled sequences. Inspect the actual tensor passed to inference, including masks, without forward passes or training. Include one intentionally truncated fixture that must be flagged as undeliverable. In the old probe, force constructs full padded tokens, but in_prompt is a branch-set boolean, not an independent end-to-end truncation audit [S3 L57–60, L176–198].

**Separating outcomes/limits.** Good retrieval + absent/wrong/masked-out value → delivery failure. All spans verified + bad answers later → beyond this boundary. Presence does not demonstrate selection, comprehension, or general reading. The completed force result already warns against equating presence with success.

**Metric/gate.** Receipt agrees with actual input for all 48/48 configurations; 100% of required deliverable spans present/visible and all designed non-deliverable cases explicitly flagged. Any silent truncation, false receipt or version substitution is no-go. No model-recall CI belongs on this deterministic assembly audit.

**Design cap.** 180 CPU-seconds, 5 wall-minutes, 512 MiB RAM, 10 MiB artifacts; at most 48 assembly checks, zero model forward calls/optimizer steps. If the real assembly boundary is unavailable without new repository instrumentation, stop and propose the needed access/change rather than editing it here.

### Stage 3 — GENERAL page-conditioned reading, before a retention claim

**Mechanism.** Freeze available weights. Use in-vocabulary entities/attributes/values but disjoint, never-optimized entity–attribute combinations; reserve the evaluation generator/map before training in any future protocol. Do not score unseen out-of-vocabulary tokens as if they were ordinary held-out facts. For each of 100 held-out keys, construct two otherwise identical pages: query K with value v, then query K with alternative v′. Balance/permutate alternatives so one default answer cannot win. Randomize declared positions and distractors without breaking query binding. Answer correctness must follow the currently supplied page, not the old trained map. Match layout to Stage 2.

**Smallest diagnostic to select next.** Evaluation-only paired value swaps on existing checkpoints, if available. First use the existing checkpoint(s); no trained linear probe, no fresh optimizer or large training run. Distinguish the primary held-out set from a smaller practiced-fact descriptive check. Full proposed screen: per checkpoint/seed, 100 held-out pairs (200 forward sequences) plus 100 corresponding no-page controls; at most two seeds × before/after checkpoints = 1,200 sequences total. If those checkpoints do not exist, this is a design needing an explicit later collection decision, not a license to rerun teaching.

**Separating outcomes.** (a) Both page versions answered correctly on held-out keys while no-page controls fail: evidence of conditional use for these synthetic mappings. (b) Original value wins even after swap: memorized map/default dominates. (c) Practiced passes, held-out fails: narrow curriculum behavior, not general reading. (d) Before and after both fail: cannot say interference destroyed a skill never established. (e) Before passes and after fails with verified delivery: loss of behavioral dependence on supplied context, with selection/representation/output causes still unisolated. No-page success alone is not evidence against reading, but high performance there weakens claims that provided pages were necessary.

**Metric/gate.** Primary unit: pair success = both alternatives correct, denominator 100 distinct keys per checkpoint per seed; publish the two marginal accuracies, paired flips and no-page baseline separately. Proposed progression gate: pair-success Wilson 95% lower bound ≥ .90 **in each seed** before wash; page-conditioned marginal accuracy at least .30 above no-page accuracy, with paired exact/McNemar test of discordances reported as a secondary check. If no-page is unexpectedly high or the lower bound fails, no retention-mechanism claim; refine eval/teaching question first. This is a screen, not a general-language reading certification. It needs future tasks and sizes before broad claims.

**Design cap.** 900 CPU-seconds, 20 wall-minutes, 2 GiB RAM, 25 MiB artifacts; 1,200 forward sequences, zero training steps. The practiced subset must fit within the cap by predeclaring its allocation (e.g. replace part of the descriptive no-page allocation before viewing results), not by expanding after seeing failure. Prefer an initial smaller predeclared descriptive subset if Kirk selects it, explicitly without a confirmatory pass claim. Never change a confirmatory denominator mid-run.

### Stage 4 — Retention while learning, with a price on rehearsal

**Mechanism.** Only after a general page-conditioned behavior is established, ask whether consolidation preserves it during new learning. Use WM01's shared teach checkpoint per seed and exactly seeds 1 and 2; compare ordinary interference wash versus a predeclared fraction of wash updates replaced by rehearsal of teach-set reading examples. Record parameter count, optimizer updates, tokens, sequence lengths and estimated FLOPs. Match update/compute budgets as closely as the task permits; equal updates alone are not equal work if sequences differ. This is a refinement proposal for Kirk's review, not a silent rewrite of WM01a/b [S10 L105–158].

**Small falsifiable diagnostic.** Reuse valid existing teach/eval artifacts first. If new work is later authorized, two arms × two seeds only, identical shared checkpoint/optimizer/RNG start, wash length 150 updates and predeclared cadence (illustrative: replace every tenth wash update by one rehearsal update; 15/150, not an extra 15). Keep batch/context/token counts equal; rehearse only teach keys, never held-out labels. Score Stage 3-style held-out swaps before/after, practiced facts separately, and fixed never-trained held-out text loss. Add a small disjoint new-fact learning check or explicitly identify held-out text from the new-learning distribution, so a retained reader that learns nothing new cannot win by freezing. Historical WM01 text eval is already required; do not mistake low training loss for new-learning generalization.

**Separating outcomes/limits.** Rehearsal preserves held-out swap behavior but damages new-task learning → stability/plasticity tradeoff, not unqualified success. It preserves only practiced answers → memorization, not retained reading skill. Both arms retain → no need to claim a special replay effect; both fail → null under this budget. Because rehearsal **replaces** wash steps, matched total updates are not matched interference exposure. Even a positive result cannot distinguish rehearsal benefit from reduced wash damage without a later exposure-matched control. Do not add that third arm without a new decision. Rest/replay is consolidation by mechanism, not proof of a gain merely because the schedule is called “rest.”

**Metric/gate.** Per seed, paired before/after held-out pair success (100 keys), practiced accuracy and new-task held-out loss/NLL, plus new-learning baseline progress. Predeclare proposed engineering screen: both seeds lose ≤ .05 absolute held-out pair success in rehearsal; rehearsal-minus-control retention gain ≥ .10 in each seed; new-task held-out NLL no more than .05 nats/token worse than control. Report uncertainty and per-key discordances; gains must have a positive paired lower confidence bound within each seed for a task-level advantage claim. With two training seeds this still is not a population-level training-mechanism conclusion. Fail/inconclusive if bounds overlap unacceptable damage or new learning is absent. A future independent-seed replication would need its own authorization.