REVIEWED PUBLIC SOURCE / HISTORICAL SNAPSHOT
Language review and ordered scan 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/CONTINUATION_FINDINGS.md- Commit
e6d9e0584066ce0fb4cfb82088c1e7490a886e7b- SHA-256
a8fc748471ee59a028027c44baf3510b7b63cf7cc16dcdc6ea1a574ffdad4a91- Size
- 8,988 bytes
Download the exact original bytes · Read the source
# Reviewed language and ordered tensor execution The practical goal is useful local memory whose interpretation and evidence remain under the user's control. This continuation implements LANG-002 and tests a separate EXEC-002 execution hypothesis. The earlier MEM-001, topology and diffusion results remain historical evidence with their original source snapshot. ## Language behavior The workbench now recognizes ten explicit phrase forms and separates a draft interpretation from the accepted record. It shows the original wording, canonical operation, actual-engine answer, overwrite source and unresolved clauses. Editing invalidates a review. Adoption replays the source and checks that its complete document matches the reviewed digest. Only the accepted record is exported. The frozen corpus results are: | Synthetic category | Cases | Version 2 exact | Version 1 against the new contract | |---|---:|---:|---:| | Supported phrase examples | 200 | 200 | 80 | | Unsupported or ambiguous examples | 200 | 200 | 200 | | Multi-step episodes | 80 | 80 | 0 | | Total | 480 | 480 | 280 | All 96 cases marked held-out pass. The templates were authored and known, and no encoder was fitted. This establishes the finite grammar contract, not learned language understanding or unseen-phrase generalization. Exactness covers all event fields except explanatory reason text, plus complete vocabulary and final bindings. There were zero unintended changes on the negative corpus. See the frozen [language protocol](LANGUAGE_REVIEW_PROTOCOL.md) and [API contract](MEMORY_LANGUAGE.md). Compatibility is assessed separately: all 480 entire version-1 documents match the original parser loaded from the verified sealed ZIP, including original reasons. Every historical document also validates under the current version-1 importer. This checks the synthetic corpus rather than proving equivalence for all strings. Actual browser checks used the selected softmax-seed-0 learned map, with source key 0 mapped to slot 23. A Unicode key was stored, overwritten, and recalled with the correct source highlight. An unquoted `it` remained unresolved, and an unwritten key returned Unknown. Version-2 export/import agreed; a forged answer was rejected without replacing the accepted record. Importing a version-1 `Set color to green` record left it unsupported; explicit review/adoption created the new interpretation and retained both prior source revisions with language version 1. A temporary local wrapper delayed preview, adoption and import responses by two seconds. Editing source, canceling adoption, and resetting while requests were in flight prevented late replies from replacing current input or accepted state. The wrapper was used only for synthetic UI testing and stopped afterward. ## Execution experiment EXEC-002 tests a chronological affine prefix scan of the same memory recurrence, using public PyTorch tensor operations and the exact prior workload/checkpoint. The frozen [scan protocol](MEMORY_SCAN_PROTOCOL.md) specifies correctness gates, six implementations, timing scopes and resource limits before measurements. The official run `exec-80c6f756439a` passed on the local RTX 5070. It completed in 16.704 seconds overall, including a 15.484-second supervised child. All six implementations returned the exact 2,185 reference answers across 128 episodes and 12,288 events, before timing and on every warmup/measured repetition. | Implementation | Median end-to-end ms | Min–max ms | |---|---:|---:| | CPU dictionary | 3.741 | 3.643–4.478 | | CPU integer slots | 4.603 | 4.501–5.746 | | CPU dense recurrence | 56.764 | 54.745–95.837 | | CUDA dense recurrence | 56.584 | 55.833–57.894 | | CPU affine scan | 425.175 | 383.602–428.924 | | CUDA affine scan | 20.243 | 19.701–21.785 |  These are ten repetitions after three warmups, with rotated/reversed arm order. Full wall time includes packing, transfers, map construction, initialization, coefficient construction, execution, synchronization and CPU decoding. Different native integer versus float32 representations remain part of this comparison. The CPU dense outlier is retained. Cold initialization is outside the timed samples; CUDA import/initialization and preliminary checks took 2.967 seconds. On this workload, CUDA scan reduced median wall time by 64.2% versus the rerun CUDA dense recurrence (2.795× faster). Its CUDA-event interval was 10.971 ms versus 47.594 ms for dense. These event intervals exclude input transfer and CPU decoding, and can include host dispatch gaps; they are not summed kernel times. The scan remained 5.41× the dictionary latency, while CPU scan was 7.49× slower than CPU dense. Keep exact CPU replay in the interactive workbench. Peak allocated GPU memory was 420,921,856 bytes (401.42 MiB) for scan and 5,785,600 bytes (5.52 MiB) for dense, a 72.75× increase. Both reported a 512 MiB reserved peak because the benchmark retains allocator cache across arms. Reserved memory is therefore not an isolated per-algorithm requirement. Supervised host process-tree peak was 2,117,808,128 bytes (1.972 GiB). All registered limits held. Sixty small CPU/CUDA control comparisons covered all three selectors, lengths 17/96/129, order, resets, ragged padding and no-read cases. Each scan was checked against sequential CPU and its same-device sequential result. Largest recorded absolute errors were 5.96e-8 for final memory, 8.94e-8 for predictions, 1.12e-8 for loss, and 9.31e-10 for logits gradients, within rtol 1e-4 / atol 1e-6. Reversing overwrite order produced the wrong value as required by the negative control. This establishes bounded differentiability parity and inference timing, not a trained-model speedup, work-efficient scan, or general GPU superiority. ## Evidence and reproduction Runtime SHA-256: `69fbd692eb80407fbde3904bfad1c1a2c5b106b5f2f08c0824df82d04d35660b`. The run retained original checkpoint, map, input and EXEC-001 receipt bytes. Read-only verification checked all 13 new payload files, the source fingerprint, dictionary replay, control records and timing arithmetic. Separate Codex subagent reviews repeated those byte/arithmetic checks; they did not rerun the GPU study or constitute cross-provider/cross-machine replication. - [Full test receipt](evidence/continuation-tests.xml): 403 tests and 17 subtests passed, no failures or skips; one expected Ripser fixture-shape warning. - [Math checks](evidence/continuation-math.json). - [Language evaluation](evidence/lang002-evaluation.json) and [historical parser compatibility](evidence/lang002-compatibility.json). - [Source manifest](evidence/continuation-source.json). - [Review ledger](evidence/continuation-review.json), including the browser checks and exact hashes of the supporting artifacts. - [Synthetic accepted session](evidence/lang002-workbench-session.json), [original v1 example](evidence/lang002-legacy-session.json), and [explicitly upgraded example](evidence/lang002-upgraded-session.json). - [Actual GPU receipt](G:/RenigmaLLM_Artifacts/memory-scan/exec-80c6f756439a/receipt.json). Reproduction commands are in [README.md](../README.md). The scan launcher starts a new unique attempt and preserves failures; `--verify` performs read-only evidence verification without rerunning timing. The prior sealed memory-stage ZIP remains unchanged at SHA-256 `44b57c986a24559791aafcdf37554f0cc5a5cf782140657fa7f31c70c008b254`. The separate [continuation source/evidence snapshot](G:/RenigmaLLM_Artifacts/release/renigma-0.1.0-continuation-69fbd692eb80-1c31a068f539.zip) contains this source, the new receipts and synthetic fixtures, and the unchanged earlier ZIP. Its payload manifest and detached `.zip.sha256` identify the bytes. It is a local artifact, not a remote publication. ## What follows from this stage Versioned interpretation, visible uncertainty, review, provenance and portable records give the user concrete control over this bounded memory workflow. Numerical coordinates alone do not establish human meaning. A GPU kernel remains an implementation of an operator, not a biological neuron or a mathematical fiber. A trained language encoder remains a separate next gate: define its phrase and composition splits, abstention rules and human-reviewed gold before fitting it; compare against this explicit grammar. Unseen-key routing, learned graph transport, topological objectives and different scan algorithms need separate controlled experiments. The earlier diffusion intervention damaged recall, so diffusion remains disabled in the workbench. No new model downloads, optimizer updates, remote publication, or cross-provider replication are part of this continuation. The workbench remains local, with explicit exports and no server conversation persistence. The existing dependency and license inventory remains applicable; this work does not establish universal IP clearance.