REVIEWED PUBLIC SOURCE / HISTORICAL SNAPSHOT
Memory-stage findings
This is the exact UTF-8 source text at the recorded commit. The implementation repository is private. Older records may describe earlier access, hosting or release decisions; their original wording is preserved. Links and commands inside the record remain inert source text and do not require repository access to read this record.
- Original path
docs/MEMORY_FINDINGS.md- Commit
e6d9e0584066ce0fb4cfb82088c1e7490a886e7b- SHA-256
11dc9b67a17d727e51138729269e57bdf1e366ba2fb77be1f3b14fad298e1a67- Size
- 16,753 bytes
Download the exact original bytes · Read the source
# Memory-stage findings and evidence This is the historical sealed memory-stage snapshot. The subsequent grammar review and affine-scan work has a new runtime fingerprint and separate [continuation findings](CONTINUATION_FINDINGS.md). Original receipts and archives retain the source identity reported below. Research snapshot, 2026-10-02. The full 15-trial main comparison, trained-model diagnostics, execution benchmark, and local workbench verification are complete. Softmax passed all five main seeds; sparsemax and 1.5-entmax each passed none. All 15 trials completed their requested 2,000 updates. One additional interrupted attempt is preserved separately. The first accepted main checkpoint supports an exact local memory workbench within the finite vocabulary and grammar below. ## What was built Supported text becomes ordered write, read, and reset events. A trained key-to-slot map supplies addresses; a standard-library integer interpreter executes them. The interface exposes current bindings, overwritten writes, source spans, unknown answers, editable source history, and validated JSON export/import. It runs at `http://127.0.0.1:8765/` on this machine while its local server is running. The model has 1,024 learned logits, 32 keys, 32 slots, and 64 supplied values plus UNKNOWN. Training batches have 128 independent episodes. No language encoder, learned value representation, output head, or unseen-key controller was trained. Human labels receive finite existing IDs. Recording a statement preserves its content and provenance; it does not verify the statement's external truth. [Architecture and dimensions](ARCHITECTURE.md), [requirements](SPEC.md), [memory protocol](MEMORY_PROTOCOL.md), [diagnostic protocol](MEMORY_DIAGNOSTICS.md), and [language contract](MEMORY_LANGUAGE.md) define the scope. ## Evidence identity and GPU execution All official memory trials use source SHA-256 `52cb4af5792561f006fc7031421dbc262a298ae12f5032bfa3c5c83b1256e686`. The source/configuration lock was maintained during training. Documentation and reporting scripts are not part of that runtime fingerprint; release manifests separately hash delivered files. The selected device is the local RTX 5070. The separate external-drive environment uses Python 3.12.10 and PyTorch 2.14.1+cu130. Actual synchronized CUDA operations, backpropagation, and a CUDA profiler trace passed; hardware inventory alone was not used as evidence. CPU/GPU memory-model forward and gradient parity passed for all three selectors at rtol 1e-4 and atol 1e-6. The dedicated parity probe's configuration was 8 keys, 8 slots, 16 supplied values plus UNKNOWN, batch 4, and 8 body events after 8 initialization writes. This is a small numerical check, distinct from the main-size resource probe below. Its largest forward/state absolute error was 5.96e-8 and largest gradient error was 1.16e-9. Float32 IEEE math and deterministic algorithms were selected; TF32 was off. Main-size preflight exercised all three selectors through 288 events, with batch 128 retained. The selected first completed run took 648.36 seconds including its training/evaluation metric scope, peaked at 349,388,288 allocated GPU bytes and 364,904,448 reserved GPU bytes, and had supervised host-tree RSS of 1,692,598,272 bytes. GPU allocator measurements exclude some driver/context allocations. Windows WDDM monitoring is not an OS-enforced isolation guarantee. ## Calibration: three treatments, one seed All three completed 2,000 updates. Initialization, generated training stream, evaluation corpus, and budget matched. The independent hard-map replay covered 1,406 reads. These are empirical learning results, not the earlier untrained diagnostic fixture. | Selector | Scientific result | Exact hard replay | Soft behavior | |---|---|---:|---| | Softmax | Failed | 1,406/1,406; no collisions | Six errors in 2,048-write constant interference; that family scored 2/8 | | Sparsemax | Passed | 1,406/1,406; no collisions | All families 100%; exact one-hot addresses; no soft/hard disagreement | | 1.5-entmax | Failed | 1,239/1,406; one collision | 1,261/1,406 overall; each retention/interference family scored 3/8 | Softmax's valid hard map does not make its observed soft execution lossless. Entmax's learned collision was preserved in its artifact and report. One seed at eight keys establishes neither a general ranking nor a scaling law. Calibration and main configurations also differ in value count, batch size, and training sequence lengths; their comparison does not isolate a single scaling factor.  ## Main comparison and selected demonstration The registered matrix is softmax, sparsemax, and 1.5-entmax, each at seeds 0–4, 2,000 updates, and the same generated training stream. All 15 completed results passed evidence-integrity verification, including the ten scientific failures. The frozen source still matches the implementation delivered with this snapshot. | Selector | Accepted seeds | Hard correct / 1,598 | Soft correct / 1,598 | Worst soft-family accuracy | Hard-map collisions | |---|---:|---:|---:|---:|---:| | Softmax | 5/5 | 1,598 | 1,598 | 100% | 0 | | Sparsemax | 0/5 | 1,461–1,517 | 1,413–1,502 | 65.625–81.25% | 2–3 | | 1.5-entmax | 0/5 | 1,470–1,552 | 1,441–1,540 | 62.5–87.5% | 1–3 | Ranges are across the five fixed seeds, not pooled acceptance scores. Every evaluation family was checked separately. Every softmax seed had zero soft/hard or relabel disagreements. Its initial and random hard controls failed in all five seeds; the dictionary control passed. Sparsemax's collision counts by seed were 2, 3, 2, 2, 2; entmax's were 2, 2, 3, 1, 1. The exact per-seed counts and retained interruption are in the [verified summary](evidence/memory-main-summary.json). No hyperparameter changes or selective retraining were made in response to failures. These treatments share mean-squared read loss, Adam at learning rate 0.03, and the fixed update budget. Their results do not rank these selectors across other objectives, architectures, or data. Independent row normalization alone imposes no injective allocation constraint. The first invocation stopped at its two-hour ceiling. Entmax seed 0 attempt 11 retains a periodic checkpoint with 1,800 saved updates and no completed evaluation. The resumed invocation restarted that seed from its original initialization and data stream; it did not resume optimizer state. Both attempts remain recorded. Completed scientific failures were terminal and were not retrained.  Before any main result completed, the demonstration rule was fixed to the first scientifically accepted checkpoint in protocol order. Softmax seed 0 passed first: all 1,598 frozen reads were correct in both soft and hard execution, with no collision, soft/hard disagreement, or relabel disagreement. This selects a working demonstration, not a representative seed or a preferred selector. Its checkpoint SHA-256 is `1e770a373b401411837764de8b3d2ac237bd898ccbf1f653e9adcaf07ec3257e`; its exact artifact SHA-256 is `2959dbcff0c81431e0c10a5b8f01e7566bd225d35a8dd4ed213652c0e3191c2e`. The same checkpoint was used for the diagnostics below, execution benchmark, and workbench. An injective shared read/write map is dictionary-equivalent for every finite valid event stream in this fixed vocabulary. That finite mechanism statement does not imply learned language semantics or unseen-key generalization. The [architecture's recurrence argument](ARCHITECTURE.md) explains the retention boundary: in real arithmetic, repeatedly writing one distractor through strictly positive addresses eventually drives every slot toward that distractor value. Finite soft retention success therefore supplies no indefinite-retention proof. The argument concerns this update rule and decoder, with explicit floating-point qualifications; it is not an additional executed experiment. ## Topology: observable geometry, limited interpretation CPU float64 diagnostics recomputed the frozen model from saved float32 logits. Sixteen independent seeded streams produced 296 reads. Soft and hard variants both answered 296/296 correctly; uniform addressing answered 32/296 correctly, the empty/reset unknown reads. There were 2,112 coordinates per query point. The same fixed sample of 128 query positions was used across variants. | Variant | Unique sampled points | Coverage radius | H0 intervals | H1 intervals | |---|---:|---:|---:|---:| | Soft | 115 | 0.129972 | 115 | 71 | | Hard | 115 | 0.130558 | 115 | 34 | | Uniform | 107 | 0.017963 | 107 | 54 | These are Euclidean Vietoris–Rips diagrams through H1 over F2, after the fixed coordinate scaling and sample. Interval counts alone do not measure meaningful memory: equally accurate soft/hard models had different counts, and incorrect uniform memory also had nontrivial persistence. Coordinate permutation preserved distances to numerical precision, and reversing the sampled point order preserved the diagrams. Permuting value channels while keeping the decoder fixed changed 114 of 128 sampled answers in both the soft and hard variants while preserving geometry. Topology therefore cannot certify the decoded meaning or sequence order of these states. No topology score entered training or model selection.  ## Diffusion: the imposed graph damaged binding The fixed cycle-graph heat operator was applied after writes. The identity, constant-preservation, and independent SciPy matrix-exponential checks all passed at the registered 1e-10 absolute tolerance. The largest reported discrepancy was 1.55e-15; a minimum coefficient of about -2.60e-16 is numerical roundoff. Soft and hard variants had identical answer counts in this diagnostic: | Heat strength | Fixed cycle | Permuted slot assignment | Matched global mean mixing | |---:|---:|---:|---:| | 0 | 296/296 (100%) | 296/296 (100%) | 296/296 (100%) | | 0.1 | 270/296 (91.22%) | 272/296 (91.89%) | 253/296 (85.47%) | | 1 | 108/296 (36.49%) | 104/296 (35.14%) | 76/296 (25.68%) | There was no semantic basis for neighboring storage slots to share values. The observed degradation supports keeping diffusion off for exact binding in this system. It does not establish that diffusion or PDE methods are generally unhelpful. Useful smoothing would need a task-specific graph, a new hypothesis, and independently registered controls.  Every total includes 32 empty/reset UNKNOWN queries, correct in every arm. Excluding those probes, the cycle's counts were 264, 238, and 76 correct out of 264 nonempty reads at strengths 0, 0.1, and 1 respectively. ## Execution cost: the GPU did not accelerate this workload EXEC-001 used the selected learned hard map on 128 episodes of 96 events each (32 initialization writes plus 64 body events), totaling 12,288 events and 2,185 reads. All six implementations matched the independent dictionary before timing. Three warmups preceded ten measured repetitions per implementation, in fixed order. Training was stopped during this benchmark. These are warm-process measurements of this workload, not training speed, browser latency, or a general hardware ranking. | Implementation | Median end-to-end batch time (ms) | Median CUDA event interval (ms) | |---|---:|---:| | Standard-library dictionary | 3.791 | — | | Standard-library integer slots | 4.566 | — | | CPU dense tensor | 56.218 | — | | CPU indexed tensor | 26.651 | — | | CUDA dense tensor | 56.224 | 47.074 | | CUDA indexed tensor | 75.468 | 66.263 | Tensor arms used the same float32 one-hot value representation. End-to-end times include fresh input packing, transfer where applicable, memory initialization, execution, CUDA timing/synchronization overhead, and decoded CPU answers. CUDA event intervals exclude input transfer and CPU decoding; they span memory initialization and execution, including gaps between dispatched operations. The dictionary and integer-slot arms use compact symbolic values, so they also compare representations. The observed result supports exact CPU replay in the local workbench. Sparse address coefficients and indexed operations alone supplied no GPU advantage here. A separately registered affine-scan or fused-kernel study could test whether reducing sequential dispatch changes costs while preserving answers, gradients, event order, and memory limits. It is a proposed experiment, not an implemented speedup or an inferred explanation of these timings. ## Human workflow and verification The actual browser used the learned artifact, visibly showing its checksum. A synthetic notebook binding used learned slot 23; a separate owner binding used slot 27. Updating the notebook value changed the subsequent answer and linked it to the correct Unicode source span. An unwritten key returned Unknown. The JSON roundtrip restored the original trace; an altered answer was rejected without replacing the valid session. A later source correction retained the prior revision. [Browser evidence](evidence/memory-workbench-validation.json) records those checks; [the synthetic portable session](evidence/memory-workbench-session.json) also passed a separate canonical replay. No personal conversation was used as a fixture.  The final implementation and reporting/package suite passed 276 tests and 17 subtests, including actual CUDA checks, with no skips. The one Ripser warning concerned an intentional point cloud with more coordinates than samples. Mathematical cycle, projector, ordered-composition, and coprimality checks passed. Cross-file review used Codex subagents; Claude Code replication has not been performed. The [final reconciliation record](evidence/memory-review.json) links the checked receipts, tests, and selected workbench configuration. ## Reproduction and local evidence locations Artifacts use the external drive. Original CPU research and its release remain preserved. This memory-stage snapshot is not remotely published. - Main matrix: `G:/RenigmaLLM_Artifacts/memory-runs/memory-1c9f9dbed9ec/receipt.json` - Calibration: `G:/RenigmaLLM_Artifacts/memory-runs/memory-b1f4f2a983ce/receipt.json` - Selected checkpoint: main batch `softmax-seed-0-attempt-1/memory-checkpoint.json` - Actual trained diagnostics: `G:/RenigmaLLM_Artifacts/runs/memory-diagnostics-20261002-170109-e4e057b2/receipt.json` - Actual trained execution benchmark: `G:/RenigmaLLM_Artifacts/memory-performance/exec-0a0c041a8b62/receipt.json` - Dedicated memory parity: `G:/RenigmaLLM_Artifacts/preflights/cuda-cdb7589d6c7c/receipt.json` - Main sizing: `G:/RenigmaLLM_Artifacts/preflights/memory-sizing-9e8c527b5aeb/receipt.json` - Source and tests: `docs/evidence/memory-source.json`, `memory-tests.xml` - Calibration table extraction: `docs/evidence/memory-calibration-summary.json` - Full main table extraction: `docs/evidence/memory-main-summary.json` The main receipt SHA-256 is `5f31b65f86becb44015b0786d8f48b921f3a7c9454966954e5cca0f16aef96de`. The separate local archive is named `renigma-0.1.0-memory-stage-source-and-evidence-52cb4af57925-5f31b65f86be.zip` under `G:/RenigmaLLM_Artifacts/release`. Its snapshot manifest hashes every payload file; the adjacent `.zip.sha256` covers the entire archive, including the manifest. The archive preserves original receipt bytes and all 16 main attempt records. Use `memory-verify` to check bytes and independently replay hard behavior. Historical soft metrics remain hash-verified observations, not a new training run. `scripts/summarize_memory.py` derives compact tables from verified receipts; `scripts/plot_memory.py` produces the family-level PNG/SVG figures. `scripts/plot_memory_diagnostics.py` checks diagnostic result hashes and plots the saved observations without recomputing or modifying the model. Dependency versions and declared licenses are in [MEMORY_DEPENDENCIES.md](MEMORY_DEPENDENCIES.md). The optional topology chain includes GPLv3-declared metadata; it is not exclusively permissive. No dependency binaries or unrelated HDD research are vendored. Project licensing, dependency metadata, and mathematical abstractions do not establish universal IP clearance. Early setup timeouts and a Windows atomic-file sharing race were retained as incomplete development probes. The race was fixed before the frozen official trials. The hand-authored, untrained diagnostic fixture established tool execution only; none of its metrics are counted as learned findings above.