Status: third-phase investigation completed, 15 September 2026; linear and nonlinear publications preserved. Read future-use/FINDINGS.md for the new cross-event composition results and future-use/README.md to reproduce them. Reader repair recovers substantial utility; exact relational uncertainty remains in three of five narrow writers. Broad and explicit full retention answer every tested composition. FUTURE_USE.md preserves the commission.
Read FINDINGS.md for the preserved linear study and followup/FINDINGS.md for the completed nonlinear experiment. Learned nonlinear codes exhibit both reader failure and exact collision uncertainty; broad reconstruction and explicit full records retain all fields. The follow-up reproduction guide links protocols and evidence. FOLLOWUP.md preserves the original commission.
This independent ancillary study asks what learned memory retains for later uses that differ from the goals anticipated during memory formation. It contributes to Construct-2's S4 question. Preparation ran no experiments; the subsequent investigator session performed methods reading, acquisition calibration, a fresh bounded comparison, and an explanatory explicit-compression control. Source studies and root documents were left unchanged.
The experiment separates three cases: expected-goal acquisition, information unavailable through a capacity-limited write projection, and a reader unable to recover retained information. A simple linear test-time memory permits both actual gradient writes and an analytic Gaussian diagnostic. It is a local mechanism study, not a reproduction of a complete language-model method.
| Stage | Protocol / methods | Evidence |
|---|---|---|
| Focused reading | Versioned sources and author-code inspection | Provenance, metadata cache |
| Acquisition | Design and resource envelope, development observations | 12 fits, gradient/write checks |
| Fresh comparison | Frozen protocol | 60 fits, all-condition tables, summary |
| Explicit derived records | Supplement protocol | 30 fresh-history evaluations, no new training |
| Publication | Findings, exportable figure | Verification, costs |
The first evaluation was frozen in Git commit 1081c78 after development;
the supplement in 39d6095 after the first evaluation. Failed-capacity cases
remain included. All learned matrices, diagnostic readers, curves, hashes,
seeds and configuration are retained locally in versioned JSON. No excluded
checkpoint, source checkout or remote service is required to inspect evidence.
Use uv with the committed lockfile. Runtime: Python 3.14.7, PyTorch 2.14.0,
NumPy 2.4.6; local CPU, float32 writes, float64 diagnostic readers, one thread.
uv sync --locked
uv run --locked scripts/verify.py
uv run --locked scripts/summarize.pyVerification regenerates data and recomputes stored-state hashes and metrics, including the supplement, without retraining. Summarization regenerates the complete tables and PNG/PDF comparison. Timing fields and PDF metadata can vary on rerun. The original data-generation/learning runner and protocols are hash-checked; small numerical differences across libraries/hardware may require inspection rather than silently relaxing the checks.
For an optional full training replay, supply a new destination that does not already exist:
uv run --locked scripts/replay.py --stage evaluation --destination /tmp/memory-goal-shift-replayThis copies the frozen recipe to that directory, generates the 60 fits there,
and leaves original evidence intact. Change --stage to development and use
a different new destination to replay calibration. Original runners refuse to
overwrite existing manifests. The supplement is executable with
scripts/supplement.py in a fresh study copy containing the evaluation donors.
The source ledger records exact versions, reading scope, retrieval details, and the distinction between published methods and our local simplification:
- PERK, 2507.06415v3: context adaptation trained for a distribution of later reasoning questions.
- TTCD, 2608.01672v1: longer-window teacher supervision of a shorter-window student's fast weights.
- Self-Guided TTT, 2607.09415v1: question-known span selection, full-context answering, per-instance reset.
- TTT layers, 2407.04620v4: learned label views and gradient-updated recurrent state; the elementary implementation lead.
The sibling neural-memory-depth supplied pinned reading artifacts, but no executable dependency. The two publications establish distinctions within linear and finite nonlinear settings. The completed third phase tests later cross-event compositions beyond directly covered targets while preserving those publications and evidence. AGENTS.md retains the operating instructions and available resources. Root synthesis remains separate under the ancillary-study approach.