liberprimus-gpu

Mission

liberprimus-gpu is a reproducible research workbench for conservative Liber Primus cryptanalysis experiments. The project keeps corpus provenance, solved baselines, transform metadata, run records, and CI gates ahead of any exploratory search or GPU acceleration work.

Current boundaries and deferred work

These are not permanent project exclusions unless marked as safety rules. They describe the current implementation boundary after Stage 5N solved-fixture-safe Gematria CUDA parity reporting and controlled expansion gate. Future experiments must stay bounded, reviewable, and reproducible before larger campaigns begin. CUDA and broad campaigns are deferred, not permanently excluded.

Permanent safety rules

• No generated output is a solve by itself.
• No Liber Primus page is claimed solved; material that is still unsolved must not receive a solve claim without a pinned corpus, manifest, transform chain, reproducible output, tests, and review.
• Raw data must not be overwritten or committed.
• Generated outputs and SQLite databases must not be committed.

Current boundaries

• Canonical corpus: inactive.
• Page boundaries: reviewable.
• Broad unsolved-page search campaigns: not started.
• Scoring campaigns: not started; Stage 3A/3B minimal triage scoring exists only for sorting and inspecting bounded 841-candidate CPU runs, Stage 3C calibration uses small local controls only, Stage 3D applies that scorer to a four-key explicit Vigenere preview only, Stage 3F applies it to the bounded 48-candidate LP evidence-key Vigenere pack only, Stage 3G applies it to a bounded 256-candidate p56-local prime-minus-one offset sweep only, Stage 3H applies it to a bounded 64-candidate reset/advance ablation with 100 negative controls only, Stage 3I applies it to a bounded 56-candidate historical motif Vigenere pack only, Stage 3J applies it to a bounded 192-candidate Mersenne/perfect-number stream probe only, and Stage 3S applies it to the bounded 72-candidate Onion 7 explicit seed pack only. Stage 4I consolidates these score labels and calibration records as triage-only metadata, not solve evidence.
• Visual/image-derived observations: registry and deterministic feature summaries only, plus deterministic review transforms; Stage 3K records source locks and reviewable observations, Stage 3M records deterministic local image features, and Stage 3P generates ignored review transforms/contact sheets. No image-derived text experiments are executed.
• Cookie/hash preimage work: Stage 3L tests two explicit SHA-256 packs only, Stage 3U tests the manifest-declared signed/public string variant pack only, and Stage 4G refreshes source-backed Stage 4B strings only. All use exact byte-string logging with no fuzzy, partial, dictionary, hashcat, GPU, or solve claims.
• Stego/OutGuess work: Stage 3V adds a deterministic OutGuess regression harness only. Stage 4F adds fixture source-lock metadata and toolchain requirements only. Missing tools/assets are explicit skipped outcomes, generated payloads stay ignored, and no broad image scan, stego extraction, audio scan, or payload interpretation is performed.
• Discord source discovery: Stage 3N scans admin-provided local HTML exports only and commits aggregate/redacted records only. Stage 3O promotes a bounded, public-safe subset of redacted source-discovery records. Stage 3Q builds ignored redacted topic shards for local AI/deep-research review. Stage 3R audits those leads, promotes only corroborated public/source-observation records, preserves false positives as negative controls, and queues disabled post-Discord manifests. Raw logs, message bodies, usernames, and private attachment URLs are not committed.
• Full Discord review bundles: Stage 4A builds redacted chronological streams, channel shards, topic shards, indexes, an LP page gallery, and an SFTP-ready static site under ignored paths for Deep Research handoff. Raw Discord logs, usernames, user IDs, message IDs, private URLs, generated static site files, copied LP page images, thumbnails, archives, and generated bundle outputs are not committed.
• Source-lock and visual observation intake: Stage 4B promotes allowlisted public-source records only, records source-health metadata, preserves cuneiform/delimiter/dot/number-square/cookie observations as review-only and non-canonical, and stores false-positive classes as negative controls. No Stage 4B visual observation is an experiment seed.
• Visual annotation: Stage 4C creates cuneiform, delimiter, dot-pattern, number-square-reference, and visual negative-control annotation tasks plus a generated local annotation site and blank coordinate templates. Coordinates and readings are separate; no Stage 4C visual task is verified, canonical, or usable as an experiment seed.
• Bounded numeric verification: Stage 4D runs only no-fudge numeric and metadata audits. GP/rune batch002 skips without exact new spans, number-square routes skip without locked raw values, delimiter/dot audits infer no meaning, cuneiform seed execution remains deferred, and cookie pack v2 is deferred to a future explicit stage.
• Source-delta audits: Stage 4E records selected cicada-solvers/iddqd tree metadata only. It does not blind-mirror external repositories, commit raw images/audio/fonts/binaries, run stego tools, or infer meaning from compression artefacts.
• Public source-lock snapshots: Stage 4K locks a small allowlisted public-source subset with metadata, canonical URLs, retrieval status, hashes where fetched, copyright notes, explicit snapshot policies, and GitHub commit-addressed references. It does not broad crawl, blind-mirror repositories, commit full HTML by default, or commit binary/image/audio/font/archive artefacts.
• Observation promotion ledger: Stage 4L joins Stage 4J review decisions with Stage 4K source locks and records ready, blocked, deferred, quarantined, rejected, source-reference-only, and control-only states. Ready does not mean executed; all future manifests remain disabled.
• Image source-variant and compression preflight: Stage 4M scans ignored local LP page images for metadata and deterministic metric-only compression summaries, records source-variant comparison readiness, keeps star-like/compression-like artefacts review-only, and keeps the bigram/Fibonacci-421 observation blocked pending reproducible matrix and null controls. It does not commit raw images, generated visualisations, or execute image/bigram experiments.
• Stego/audio positive-control readiness: Stage 4N records fixture readiness, cache policy, expected-output requirements, toolchain state, and synthetic controls. Historical OutGuess/OpenPuff/MP3/audio cases remain blocked until assets, exact expected outputs, and tools are ready.
• CPU batch transform API: Stage 4H provides deterministic CPU-only input stream, transform candidate, result, summary, adapter coverage, and parity-contract records. Stage 4O expands this with solved-fixture-safe streams, adapter coverage records, deterministic parity expectations, and score-summary compatibility checks. Stage 4P makes those result and score surfaces comparable with older result-store summaries. Stage 4Q adds benchmark and future parity planning gates. Stage 5A adds CUDA target plans, non-targets, parity scaffolds, and implementation gates. Stage 5B adds CUDA parity harness plans, parity fixtures, backend capability profiles, and future-kernel matrix rows. Stage 5C adds no-GPU-safe CUDA build profiles, toolchain detection, device detection, and optional smoke-build readiness metadata. Stage 5D adds native C++ CPU backend, deterministic threading, and native/Python parity records. Stage 5E selects shift_score_kernel as the first future CUDA kernel contract and maps it to the native synthetic shift parity adapter. Stage 5F implements only that synthetic-only CUDA parity kernel and records matching synthetic hash metadata. Stage 5G reports the Stage 5F hash match and hardens .cu/.cuh CUDA-facing code to a conservative C-style subset. Stage 5H defines the separate numeric Gematria mod-29 contract and native fixture while keeping solved-fixture CUDA execution blocked. Stage 5I adds Gematria CUDA preparation records for a CUDA-C ABI and validation vectors. Stage 5J adds only the synthetic numeric gematria_mod29_shift_score_kernel and matching parity records. Stage 5K reports that parity, audits device-code subset compliance, and records solved-fixture-safe blockers without CUDA execution. Stage 5L maps committed solved-fixture-safe streams into exact Gematria token buffers and native output-token hashes. Stage 5M runs the existing kernel over exactly those five mapped buffers and matches all native hashes. It is infrastructure for future CUDA parity, not a broad experiment runner.
• Scoring contract: Stage 4I provides scorer inventory, finite confidence-label records, compatibility mappings, calibration-profile/report records, and CPU batch score compatibility checks. Score labels can create review leads only and cannot imply solved plaintext.
• Post-Discord experiment execution: Stage 3S executes only EXP-3R-003, the bounded Onion 7 explicit seed pack. Stage 3T executes only EXP-3R-004, the GP/rune claim verifier. Stage 3U executes only EXP-3R-001, the cookie SHA-256 signed-variant pack. All keep generated records under ignored paths and make no solve claim.
• CUDA experiment campaigns: not started.
• Normal bounded local CPU experiments: allowed automatically when they pass experiments/policies/operator-policy-v0.yaml.
• Broad unsolved-page campaigns: not started.
• Approval packets: optional/high-risk audit tooling, not the default path for policy-passing bounded CPU items.
• Existing CUDA code is scaffold, smoke-test infrastructure, the Stage 5F synthetic uppercase-Latin shift_score_kernel, Stage 5G reporting/device-code hardening, Stage 5H metadata/fixture contract records, Stage 5I preparation metadata, and the Stage 5J synthetic numeric gematria_mod29_shift_score_kernel only. Stage 5K adds reporting/preflight records only; Stage 5L adds token-mapping/native-fixture records; Stage 5M adds host-runner-only plumbing plus exact solved-fixture parity records and adds no new kernel.

Deferred future work

• Stage 5O solved-fixture-safe Gematria CUDA repeat verification and result-store preflight.
• Website expansion is deferred to Stage 6.
• Future visual numeric observations for base-60 or cuneiform-like numbers, binary dot patterns, symmetry/asymmetry, and page imagery must remain reviewable before becoming experiment seeds.
• Search campaigns.
• CUDA kernels after CPU references and parity tests exist.
• Benchmark campaigns after stable CPU/GPU baselines exist.

Already implemented since Stage 0A

• Profile and corpus-candidate infrastructure.
• Ten known solved baseline fixtures.
• CPU transform registry.
• Solved-baseline manifest runner.
• JSONL/SQLite result-store foundation.
• Raw-data-free GitHub Actions CI.
• CI-gated consistency checks.
• CPU exploratory experiment dry-run planner.
• Bounded CPU execution harness for synthetic and solved-fixture-only runs.
• Exploratory experiment proposal and human-approval workflow.
• Approval-gated execution path for approved synthetic/solved controls.
• First real bounded exploratory approval-readiness packet.
• Standing bounded local CPU operator policy and queue scaffold.
• Minimal CPU Caesar plus affine executor and triage scoring for the first 841 candidate bounded queue item.
• Candidate lead inspection, refined triage scoring, reranking, and reverse-direction bounded comparison.
• Scoring calibration with positive controls, null controls, negative controls, tiny crib checks, and a conservative Stage 3D queue recommendation.
• Small explicit-key Vigenere preview for the four declared known-motif keys, with calibrated scoring and ignored generated outputs.
• Deep Research method backlog ingestion, deterministic Stage 3E queue counts, and dry-run executor-support classification.
• Reset/advance-aware Stage 3F LP evidence-key Vigenere pack executor and 48-candidate run.
• Bounded Stage 3G p56-local prime-minus-one offset sweep executor and 256-candidate run, plus a queued future Mersenne/perfect-number probe.
• Shared Stage 3H reset/advance state machine, 64-candidate ablation run, and 100 family-specific negative controls.
• Bounded Stage 3I historical motif Vigenere pack run for 14 declared keys and 56 candidates.
• Bounded Stage 3J Mersenne/perfect-number stream probe for the finite declared exponent sequence and 192 candidates.
• Stage 3K historical archive/source-lock, local image metadata, visual numeric observation, and cookie/hash artefact registry.
• Stage 3L bounded SHA-256 cookie-hash preimage packs for the two archived 2013 cookie artefacts.
• Stage 3M deterministic local image-analysis CLI and visual-feature summaries.
• Stage 3N admin-approved Discord HTML archive ingestion and source-discovery index.
• Stage 3O privacy-preserving Discord source promotion, expanded tutorials, and GitHub Wiki mirror source generation.
• Stage 3P deterministic image transform suite, contact sheets, and local visual review index.
• Stage 3Q redacted Discord AI-review bundles and topic shards.
• Stage 3R Discord lead promotion audit, negative controls, and first disabled post-Discord experiment manifests.
• Stage 3S bounded Onion 7 explicit seed-pack execution.
• Stage 3T bounded GP/rune claim verifier execution.
• Stage 3U bounded cookie SHA-256 signed-variant pack execution.
• Stage 3V OutGuess regression harness.
• Stage 3W project-state consolidation and anti-drift checks.
• Stage 3X CLI modularisation without behavior change.
• Stage 3Y result synthesis, staged plan, and method-retirement ledger.
• Stage 3Z source-of-truth / newcomer map.
• Stage 4A full Discord research-bundle extraction and static review site generation.
• Stage 4B website-derived source-lock triage and visual observation intake.
• Stage 4C cuneiform and dot annotation pack.
• Stage 4D bounded numeric verifier pack.
• Stage 4E cicada-solvers/iddqd source-lock delta audit.
• Stage 4F historical OutGuess/audio fixture source-locking.
• Stage 4G cookie exact-candidate refresh.
• Stage 4H CPU batch transform API extraction.
• Stage 4I scorer consolidation and calibration report.
• Stage 4J observation review workflow hardening.
• Stage 4K allowlisted public source-lock snapshots.
• Stage 4L reviewed observation promotion ledger.
• Stage 4M image source-variant and compression preflight.
• Stage 4N OutGuess/audio positive-control completion.
• Stage 4O CPU batch adapter expansion.
• Stage 4P result-store and score-summary unification.
• Stage 4Q CPU benchmark and parity planning.
• Stage 5A CUDA planning and parity scaffolding.
• Stage 5B CUDA parity harness skeleton.
• Stage 5C CUDA build and device-detection scaffold.
• Stage 5D native C++ CPU batch backend and deterministic threading baseline.
• Stage 5E first CUDA kernel contract and CPU/native parity adapter selection.
• Stage 5F first synthetic-only CUDA parity kernel implementation.
• Stage 5G shift_score CUDA parity reporting and solved-fixture-safe adapter preflight.
• Stage 5H Gematria mod-29 shift_score contract and native parity fixture preparation.
• Stage 5I Gematria mod-29 shift_score synthetic CUDA parity preparation.
• Stage 5J Gematria mod-29 shift_score synthetic CUDA parity kernel implementation.
• Stage 5K Gematria shift_score CUDA parity reporting and solved-fixture-safe preflight.
• Stage 5L solved-fixture-safe Gematria shift_score token mapping and native parity fixture preparation.
• Stage 5M first solved-fixture-safe Gematria shift_score CUDA parity run.
• Stage 5N solved-fixture-safe Gematria CUDA parity reporting and controlled expansion gate.

Architecture summary

The CPU side owns corpus management, manifests, hypothesis generation, branching search, result provenance, and manual review. The GPU side will only accelerate large regular batches of transform-and-score work after a CPU reference implementation, parity tests, and benchmarks exist.

Where To Start

• Start Here: plain-English current-state overview.
• Source Of Truth Map: which file answers which question.
• Staged Plan: completed stages, current direction, planned work, deferred work, and method-retirement context.
• Tutorial Index: public workflow guides.

Current status

• Stage 2A: CPU transform registry and manifest-addressable solved-baseline runner complete.
• Stage 2B: JSONL/SQLite experiment result-store foundation complete.
• Stage 2C: raw-data-free GitHub Actions CI complete and lock-hash hardened.
• Stage 2D: CI-gated schema/docs consistency and manifest/result-store hardening complete.
• Stage 2E: CPU exploratory experiment manifest scaffold and dry-run planner complete.
• Stage 2F: bounded CPU execution harness for synthetic and solved-fixture-only runs complete.
• Stage 2G: exploratory experiment proposal and human-approval workflow complete.
• Stage 2H: approval-gated execution path for approved synthetic/solved controls complete.
• Stage 2I: first real bounded CPU exploratory experiment approval packet complete.
• Stage 2J: standing bounded CPU auto-run policy and queue scaffold complete.
• Stage 3A: minimal CPU Caesar plus affine executor and triage scoring complete.
• Stage 3B: Stage 3A lead inspection, scoring refinement, rerank, and reverse-direction comparison complete.
• Stage 3C: scoring calibration, null controls, positive controls, and tiny crib checks complete.
• Stage 3D: small Vigenere known-motif key-list preview complete.
• Stage 3E: Deep Research method backlog ingestion and bounded queue dry-run complete.
• Stage 3F: LP evidence-key Vigenere pack execution complete.
• Stage 3G: p56-local prime-minus-one offset sweep complete.
• Stage 3H: reset/advance ablation and family-specific negative controls complete.
• Stage 3I: historical motif Vigenere key-pack run complete.
• Stage 3J: Mersenne/perfect-number tiny stream probe complete.
• Stage 3K: archive and visual observation registry complete.
• Stage 3L: bounded cookie-hash preimage packs complete.
• Stage 3M: deterministic local image analysis complete.
• Stage 3N: admin-approved Discord HTML archive ingestion complete.
• Stage 3O: Discord source promotion and Wiki tutorial mirror complete.
• Stage 3P: deterministic image transform suite and visual review index complete.
• Stage 3Q: redacted Discord AI-review bundles and topic shards complete.
• Stage 3R: Discord lead promotion audit and first post-Discord manifest queue complete.
• Stage 3S: bounded Onion 7 explicit seed-pack execution complete.
• Stage 3T: bounded GP/rune claim verifier execution complete.
• Stage 3U: bounded cookie SHA-256 signed-variant pack execution complete.
• Stage 3V: OutGuess regression harness complete.
• Stage 3W: state consolidation and anti-drift hardening complete.
• Stage 3X: CLI modularisation without behavior change complete.
• Stage 3Y: result synthesis, staged plan, and method-retirement ledger complete.
• Stage 3Z: source-of-truth / newcomer map complete.
• Stage 4A: full Discord research-bundle extraction for Deep Research complete.
• Stage 4B: website-derived source-lock triage and visual observation intake complete.
• Stage 4C: cuneiform and dot annotation pack complete.
• Stage 4D: bounded numeric verifier pack complete.
• Stage 4E: cicada-solvers/iddqd source-lock delta audit complete.
• Stage 4F: historical OutGuess/audio fixture source-locking complete.
• Stage 4G: cookie exact-candidate refresh complete.
• Stage 4H: CPU batch transform API extraction complete.
• Stage 4I: scorer consolidation and calibration report complete.
• Stage 4J: observation review workflow hardening complete.
• Stage 4K: allowlisted public source-lock snapshots complete.
• Stage 4L: reviewed observation promotion ledger complete.
• Stage 4M: image source-variant and compression preflight complete.
• Stage 4N: OutGuess/audio positive-control completion complete.
• Stage 4O: CPU batch adapter expansion complete.
• Stage 4P: result-store and score-summary unification complete.
• Stage 4Q: CPU benchmark and parity planning complete.
• Stage 5A: CUDA planning and parity scaffolding complete.
• Stage 5B: CUDA parity harness skeleton complete.
• Stage 5C: CUDA build and device-detection scaffold complete.
• Stage 5D: native C++ CPU batch backend and deterministic threading baseline complete.
• Stage 5E: first CUDA kernel contract and CPU/native parity adapter selection complete.
• Stage 5F: first synthetic-only CUDA parity kernel implementation complete.
• Stage 5G: shift_score CUDA parity reporting and solved-fixture-safe adapter preflight complete.
• Stage 5H: Gematria mod-29 shift_score contract and native parity fixture preparation complete.
• Stage 5I: Gematria mod-29 shift_score synthetic CUDA parity preparation complete.
• Stage 5J: Gematria mod-29 shift_score synthetic CUDA parity kernel implementation complete.
• Stage 5K: Gematria shift_score CUDA parity reporting and solved-fixture-safe preflight complete.
• Stage 5L: solved-fixture-safe Gematria shift_score token mapping and native parity fixture preparation complete.
• Stage 5M: first solved-fixture-safe Gematria shift_score CUDA parity run complete.
• Stage 5N: solved-fixture-safe Gematria CUDA parity reporting and controlled expansion gate complete.
• Known solved baselines: 10 passing through the registry/manifest path.
• Fixture breakdown: direct translation 4, Atbash-family 3, explicit-key Vigenere 2, p56 prime-minus-one / phi-prime 1.
• Canonical corpus: inactive.
• Page boundaries: reviewable.
• Broad search/scoring/CUDA campaigns: not started.
• Latest CUDA kernel stage: Stage 5J wrote 1 implementation record, 1 build record, 1 synthetic parity record, and 1 summary record for gematria_mod29_shift_score_kernel. Local optional CUDA build and synthetic numeric parity passed with hash a6d5d5161145fd31ab429a8e955e0412d7b0af6089f06ee8b33baf8cd00b27a0; no real Liber Primus CUDA data, solved or unsolved page CUDA execution, GPU benchmark, speedup claim, or solve claim was added.
• Latest Gematria solved-fixture CUDA reporting stage: Stage 5N wrote 5 parity report records, 5 controlled expansion gate records, 1 boundary review record, 2 result-store/score-summary preflight records, 9 no-unsolved guardrail records, and 1 summary record. It carries forward Stage 5M parity pass/fail/skip 5/0/0, approves only exact-repeat verification for Stage 5O, blocks broad solved-fixture and unsolved-page CUDA, adds 0 new CUDA kernels, modifies no CUDA source, runs no CUDA, and makes no benchmark, speedup, real Liber Primus CUDA-data, website-expansion, canonical-corpus, page-boundary, or solve claim.
• Latest solved-fixture mapping stage: Stage 5L wrote 5 token-mapping records, 5 native parity records, 1 output-hash contract, 1 score-summary shape, and 1 summary record. It maps all 5 candidate streams, prepares 5 native output-token hashes, and supplies Stage 5M's exact approved input set.
• Latest bounded hash review: Stage 4G tested 4 source-backed deduplicated SHA-256 candidate byte strings against the two archived cookie/hash targets for 8 exact comparisons and found 0 exact matches; no solve claim.
• Latest image-analysis stage: Stage 3M analysed 58 ignored local page images, producing 406 component records, 58 symmetry records, 464 bitplane records, and 71 review-only feature candidates in ignored outputs. No OCR, AI/ML interpretation, image-derived search, or solve claim is made.
• Latest source-discovery stage: Stage 3O promoted 500 public source links, 200 method-claim candidates, and 200 numeric-observation candidates from the Stage 3N extraction into redacted review records. It rejected private/unsafe links, expanded public tutorials, and generated Wiki source pages. Raw Discord logs, message bodies, usernames, private URLs, and generated review indexes remain uncommitted; no solve claim.
• Latest visual review stage: Stage 3P processed 58 ignored local page images, emitted 2077 derived review images, 59 contact sheets, 58 review pages, and 6 review-only visual transform candidates under ignored outputs. No OCR, AI/ML interpretation, image-derived search, or solve claim is made.
• Latest Discord review stage: Stage 3Q generated 1700 redacted stream records, 17 topic shard files, and 900 review leads under ignored outputs, with an aggregate-only committed summary. Raw logs, message bodies, usernames, private URLs, AI upload, live API use, scraping, and solve claims remain absent.
• Latest Discord promotion stage: Stage 3R promoted 13 public source records and 11 review-only observation records, preserved 11 negative-control records, counted 25 unsafe/private or quarantined records, and created 3 disabled post-Discord manifests. No experiment was executed and no solve claim was made.
• Latest post-Discord experiment stage: Stage 3S executed EXP-3R-003 only, producing 72 bounded Onion 7 candidates from 3 value spaces, 6 routes, 2 directions, and 2 reset modes. The top candidate scored 1.460714 with calibrated confidence inconclusive; generated outputs remain ignored and no solve claim is made.
• Latest GP/rune verifier stage: Stage 3T executed EXP-3R-004 only, loading and deduplicating 25 exact claims. It classified 23 as verified and 2 as unsupported, with no unverified, boundary-sensitive, missing-source-span, malformed, or duplicate claims in the committed input set.
• Latest cookie signed-variant stage: Stage 3U executed EXP-3R-001 only, generating 156 candidates before deduplication, testing 105 deduplicated byte strings against 2 cookie targets for 210 exact SHA-256 comparisons, and finding 0 exact matches.
• Latest stego regression stage: Stage 3V added the OutGuess harness, detected no local OutGuess binary, and recorded 6 missing-tool skips plus 1 disabled case across 7 manifest cases. No raw artefacts or payloads were committed.
• Latest consolidation stage: Stage 3W refreshed persistent project context, defined the source-of-truth hierarchy, and added anti-drift checks so long-lived docs cannot drift back to obsolete current-state claims.
• Latest maintainability stage: Stage 3X split the Python CLI into cli_commands domain modules while preserving python -m libreprimus.cli and adding command-surface tests.
• Latest synthesis stage: Stage 3Y added docs/roadmap/staged-plan.md, research synthesis records, method-family status and retirement ledgers, Deep Research influence records, direction-change records, and validation CLI commands.
• Latest onboarding stage: Stage 3Z added source-of-truth, Codex navigation, Deep Research handoff, contributor module, task-lane, and private/generated data maps.
• Latest Deep Research handoff stage: Stage 4A processed 43 local ignored Discord HTML exports into 520009 redacted messages, 1327 channel shards, 12 topic shards, and an ignored SFTP-ready static review site with 58 LP page images included as generated gallery assets. Raw logs, raw images, private URLs, and generated outputs remain uncommitted.
• Latest source-lock stage: Stage 4B promoted 20 allowlisted public source records, recorded 19 source-health records, added 6 review-only visual observations and 17 negative controls, and queued 7 disabled future manifests. No experiments were executed and no solve claim is made.
• Latest annotation stage: Stage 4C created 15 visual annotation tasks, including 1 cuneiform task, 1 dot task, 2 delimiter tasks, and 10 visual negative-control tasks. Generated annotation outputs remain ignored and no visual meaning is inferred.
• Latest bounded numeric stage: Stage 4D discovered 7 manifests, audited 3, deferred or skipped 4, verified 0 GP/rune claims because no exact new spans were present, audited 2 delimiter observations and 10 visual negative controls, skipped number-square routes because raw values are not locked, and deferred cuneiform/cookie execution.
• Latest source-delta stage: Stage 4E inspected 310 public cicada-solvers/iddqd tree paths, recorded 1 source-delta record, 12 source-health records, 1 image artefact future-preflight record, and 4 disabled future manifests without committing raw artefacts.
• Latest stego/audio fixture stage: Stage 4F recorded 5 OutGuess fixture source records, 5 audio fixture source records, 10 source-health records, 5 toolchain requirement records, and 4 disabled future manifests without running OutGuess, OpenPuff, MP3Stego, hexdump/strings, audio analysis, or committing raw artefacts.
• Latest cookie refresh stage: Stage 4G generated 4 source-backed candidates before and after deduplication, marked 2 previous-pack duplicates, ran 8 SHA-256 exact comparisons, and found 0 exact matches without fuzzy matching, hashcat, GPU/CUDA, raw Discord processing, raw page-image processing, or solve claims.
• Latest CPU batch API stage: Stage 4H executed 6 synthetic CPU-only candidates, supported all 6 current registry transforms, wrote 6 ignored result records, and created the CPU/CUDA parity contract without GPU code.
• Latest scorer consolidation stage: Stage 4I recorded 3 scorer records, 9 confidence labels, 11 compatibility mappings, and 1 calibration profile. Positive/null/negative controls are available, CPU batch score compatibility is true, and scoring remains triage only.
• Latest observation review stage: Stage 4J loaded 96 observation-family records, wrote 96 review decisions, 96 promotion-gate records, 23 quarantine records, and promoted 0 observations to manifests. Review-only observations remain non-seeds.
• Latest source snapshot stage: Stage 4K considered 43 source candidates, wrote 15 allowlisted source-lock snapshot records, locked 8 GitHub commit-addressed references, fetched 1 ignored local snapshot, committed 0 raw snapshots, rejected 22 unsafe/noisy or non-priority sources, and recorded 6 duplicate sources.
• Latest observation promotion stage: Stage 4L loaded 97 reviewed observations, created 97 ledger records, 97 readiness records, 115 blocker records, and 13 manifest-readiness records. It marked 0 observations ready for manifest execution, 17 control-only, 14 source-reference-only, 48 blocked, 2 deferred, 15 quarantined, and 1 rejected. The added bigram/Fibonacci-421 community claim remains blocked pending reproducible matrix regeneration, rune-order declaration, indexing convention, null controls, and multiple-testing controls.
• Latest image preflight stage: Stage 4M scanned 58 ignored local LP page images, wrote 58 source-variant records, 58 deterministic compression metric records, 1 review-only image artifact candidate, and 1 blocked bigram/Fibonacci-421 readiness record. It committed no raw images or generated visualisations and ran no image/bigram experiment.
• Latest stego/audio readiness stage: Stage 4N wrote 11 OutGuess readiness records, 5 audio readiness records, 16 fixture-cache records, 16 expected-output records, 7 toolchain readiness records, 0 historical execution-ready fixtures, 8 blocked historical fixtures, and 2 synthetic-ready controls. It committed no raw artefacts or extracted payloads and ran no stego/audio tools.
• Latest result-store unification stage: Stage 4P recorded 18 source inventory records, loaded 11 committed summaries, saw 6 optional ignored generated outputs locally, wrote 82 unified result records, 82 unified score-summary records, and 82 method-status joins. It ran no new experiments, added no scorer semantics, committed no generated result bodies or SQLite databases, and added no CUDA code.
• Latest CUDA planning stage: Stage 5A wrote 14 target-plan records, 9 ready planning targets, 2 blocked targets, 8 explicit non-target records, 9 parity scaffold records, and 10 satisfied implementation gates. It made no CUDA source change, GPU benchmark, speedup claim, broad experiment, raw-data processing, or solve claim.
• Latest CUDA parity harness stage: Stage 5B wrote 14 harness plan records, 14 parity fixture records, 3 backend capability records, and 9 future-kernel matrix records. It made no CUDA source change, GPU benchmark, speedup claim, broad experiment, raw-data processing, or solve claim.
• Latest CUDA build/device stage: Stage 5C wrote 3 build profiles, 3 toolchain records, 3 device records, and 1 optional smoke-build record. It made no CUDA source change, CUDA kernel, GPU benchmark, speedup claim, broad experiment, raw-data processing, website expansion, or solve claim.
• Latest native CPU backend stage: Stage 5D wrote 1 backend capability record, 5 threading records, 1 native/Python parity record, and 1 diagnostic record. It tested thread counts 1,2,4,8,16, produced matching one-thread and multi-thread hashes, preserved Python as orchestration, and made no CUDA source change, GPU benchmark, speedup claim, broad experiment, raw-data processing, website expansion, or solve claim.
• Latest CUDA reporting stages: Stage 5G reports the Stage 5F shift_score_kernel native/CUDA synthetic hash match and Stage 5K reports the Stage 5J Gematria CUDA/native synthetic hash match. Both verify CUDA-facing device-code subset policy, record solved-fixture-safe blockers, and keep real Liber Primus CUDA data use, GPU benchmarks, speedup claims, broad execution, raw-data processing, website expansion, and solve claims out of scope.
• Latest Gematria CUDA kernel stage: Stage 5J implements gematria_mod29_shift_score_kernel for the Stage 5H synthetic numeric fixture only. It shifts transformable tokens with (token + shift) % 29, preserves masked separator placeholders, matches the native fixture hash a6d5d5161145fd31ab429a8e955e0412d7b0af6089f06ee8b33baf8cd00b27a0, keeps Stage 5F uppercase Latin parity separate, and adds no real Liber Primus CUDA data use, solved/unsolved page CUDA execution, GPU benchmark, speedup claim, website expansion, or solve claim.
• Latest Gematria solved-fixture CUDA parity/reporting stage: Stage 5N reports Stage 5M's five CUDA/native hash matches, records controlled expansion gates, and keeps unsolved-page CUDA blocked.
• Durable staged plan: docs/roadmap/staged-plan.md.
• Next: Stage 5O solved-fixture-safe Gematria CUDA repeat verification and result-store preflight.

How To Use This Repo

1. Set up Python 3.12 and a local virtual environment using the Windows or Linux tutorial.
2. Run the local validation stack before trusting a change:

.\.venv\Scripts\python.exe -m ruff check python/libreprimus tests/python
.\.venv\Scripts\python.exe -m pytest -q tests/python
.\.venv\Scripts\python.exe -m libreprimus.cli consistency check-state-drift
.\.venv\Scripts\python.exe -m libreprimus.cli research-synthesis validate --data-dir data/research --staged-plan docs/roadmap/staged-plan.md
.\.venv\Scripts\python.exe -m libreprimus.cli consistency check-all --allow-warnings

3. Keep local raw material local. Raw Discord logs, page images, transcript drops, workbooks, and Pastebin files are ignored by design.
4. Use generated outputs for local review only. Candidate dumps, image-analysis records, Discord extraction JSONL, SQLite databases, and local review indexes must not be committed.
5. Start with the tutorial index for repo tours, solved baselines, bounded queues, image analysis, Discord archive ingestion, and source/observation registry workflows.
6. The GitHub Wiki mirrors tutorial pages for public browsing, but repository tutorial files are the source of truth.

Safe bounded CPU experiments use experiments/policies/operator-policy-v0.yaml; they remain CPU-only, budget-limited, and unable to claim solves. CUDA work waits for CPU references, parity tests, and explicit future scope.

CI status

GitHub Actions runs at NoxxGames/LiberPrimus-GPU Actions. CI is raw-data-free, CUDA-free, secret-free, and does not upload generated corpus or result artifacts by default. Real-source smoke checks remain local-only because ignored raw sources are not present on GitHub-hosted runners.

Tutorials

Start with tutorials/README.md. The tutorials cover Windows and Linux setup, local data handling, current CLI tools, transcript alignment, hardware expectations, and Codex-assisted development.

GitHub wiki

Wiki source pages live under docs/wiki-source/ and are generated from tutorials/. The repository tutorials and docs are the source of truth; the GitHub wiki is a public mirror and must not contain raw data, generated dumps, or solve claims. Use scripts/github/validate-wiki-source.ps1 and scripts/github/sync-tutorials-to-wiki.ps1 --DryRun before publishing.

Issues and backlog

Issue templates live under .github/ISSUE_TEMPLATE/. Seed issues for future work live under docs/github/issues/ and can be created idempotently with scripts/github/create-issues.ps1.

Public-readiness status

The project is public-readable for documentation and scaffold review. It is not ready for unsolved-page cryptanalysis campaigns, canonical corpus release, or GPU acceleration claims.

Quick start on Windows

.\scripts\verify-toolchain.ps1
.\scripts\configure-windows.ps1
cmake --build build\msvc-debug
ctest --test-dir build\msvc-debug --output-on-failure

For Python:

py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -U pip setuptools wheel
.\.venv\Scripts\python.exe -m pip install -e ".[dev]"
.\.venv\Scripts\python.exe -m pytest -q
.\.venv\Scripts\libreprimus.exe smoke

Toolchain requirements

• Windows 11 or compatible Windows 10 development host.
• Git for Windows.
• CMake 3.26 or newer.
• Ninja.
• Python >=3.12,<3.14, with Python 3.12 preferred.
• Visual Studio 2022 Build Tools with the Desktop C++ workload.
• Optional CUDA Toolkit for CUDA smoke builds.

Configure and build CPU-only

cmake -S . -B build\msvc-debug -G Ninja -DCMAKE_BUILD_TYPE=Debug -DLPGPU_ENABLE_CUDA=OFF -DLPGPU_BUILD_TESTS=ON
cmake --build build\msvc-debug
ctest --test-dir build\msvc-debug --output-on-failure

If cl.exe is not visible in the current shell, run through scripts\configure-windows.ps1, which locates VsDevCmd.bat.

Configure and build with CUDA

CUDA builds remain optional and are scaffold/smoke only at the current stage.

cmake -S . -B build\cuda-debug -G Ninja -DCMAKE_BUILD_TYPE=Debug -DLPGPU_ENABLE_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=89 -DLPGPU_BUILD_TESTS=ON
cmake --build build\cuda-debug
ctest --test-dir build\cuda-debug --output-on-failure

Python environment setup

The Python package exposes smoke/toolchain commands plus legacy ingestion, transcript alignment, profile validation, corpus-candidate generation, and solved-fixture reproduction.

Direct fixture smoke:

.\.venv\Scripts\python.exe -m libreprimus.cli solved-fixture stage1a-smoke --fixture-dir data/fixtures/solved-pages/direct-translation-v0 --candidate-dir data/normalized/corpus-candidates/rtkd-master-v0-candidate --out-dir data/normalized/solved-baselines/direct-translation-v0 --allow-pending --allow-warnings

All-known solved-baseline registry smoke:

.\.venv\Scripts\python.exe -m libreprimus.cli solved-baseline stage2a-smoke --manifest experiments/manifests/solved-baselines/stage2a-all-known-solved-baselines.yaml --out-dir experiments/results/solved-baselines/stage2a --allow-warnings

Stage 2B result-store smoke:

.\.venv\Scripts\python.exe -m libreprimus.cli result-store stage2b-smoke --solved-baseline-manifest experiments/manifests/solved-baselines/stage2a-all-known-solved-baselines.yaml --result-store-manifest experiments/manifests/result-store/stage2b-solved-baseline-import.yaml --solved-baseline-out-dir experiments/results/solved-baselines/stage2a --result-store-out-dir experiments/results/result-store/stage2b --replace --allow-warnings

Local Stage 2C CI reproduction:

.\scripts\ci\run-python-ci.ps1
.\scripts\ci\run-schema-manifest-checks.ps1

Repository map

• src/: C++20 native scaffold.
• cuda/: optional CUDA smoke scaffold.
• python/: Python orchestration package.
• tests/: C++ and Python smoke tests.
• data/: immutable raw-data placeholders and future corpus areas.
• experiments/: manifest-driven experiment policy and smoke manifest.
• docs/: methodology, CUDA, corpus, scoring, and Codex notes.
• scripts/: Windows verification, bootstrap, configure, cleanup, and GitHub helper scripts.
• tutorials/: public user-facing tutorials.
• .github/: issue templates and pull request template.

Data policy

data/raw/ is immutable. Do not overwrite raw evidence, normalize in place, or commit real raw corpus files. Corpus work must use explicit SHA-256 locks and transcript version metadata.

Experiment policy

Experiments are manifest-driven. Candidate outputs must never be treated as solves without pinned corpus data, full transform chains, score metadata, null controls, reproducible tests, and manual review.

Testing policy

Current tests cover the C++ skeleton, Python package, manifests, schemas, result stores, lock hashes, public documentation status, and consistency gates. Future CUDA kernels must have CPU reference implementations, CPU/GPU parity tests, and benchmarks before optimization.

Next milestones

Stage 2J replaced per-experiment approval as the default path with the standing policy in experiments/policies/operator-policy-v0.yaml and the queue in experiments/queues/stage2j-bounded-cpu-queue.yaml. Normal local CPU items can run automatically when they stay within the hard limits: candidate upper bound 100000, runtime estimate 600 seconds, generated output budget 250 MB, CPU only, no CUDA/cloud/paid services, no generated-output commit, no canonical corpus activation, no page-boundary finalization, and no solve claim.

The first Caesar plus affine reviewable-slice queue item has candidate upper bound 841 and is policy-eligible. Stage 3A adds the minimal CPU executor and deterministic triage scoring for that item. Full candidate outputs remain ignored under experiments/results/bounded-auto-runs/stage3a/; committed research logs summarize counts and top score metadata only.

Stage 3B inspected Stage 3A top candidates, refined the scorer, reranked the 841 candidates, and ran the reverse-direction comparison. Both refined and reverse-direction top leads remain noisy.

Stage 3C calibrated scoring against positive solved-fixture controls, deterministic null controls, negative controls, and tiny crib checks. Stage 3D ran the conservative small Vigenere known-motif key-list preview for exactly four declared keys. The top key was LIBER, calibrated as noisy.

Stage 3E ingests the Deep Research method backlog, commits experiments/queues/stage3e-method-backlog.yaml and experiments/queues/stage3e-bounded-cpu-queue.yaml, and dry-runs bounded methods with deterministic counts: LP evidence Vigenere 48, p56 local prime-minus-one offsets 256, historical Vigenere 56, family-specific negative controls 100, reset/advance ablation 64, prime mod/gap 256, and Mersenne/perfect-number probe 192.

Stage 3F implements the reset/advance-aware evidence-key Vigenere pack executor and runs only stage3e_vig_lp_evidence_pack_v1. It executes 48 candidates, records key/reset/advance metadata, leaves generated outputs ignored, and makes no solve claim.

Stage 3G implements the p56-local prime-minus-one offset sweep executor and runs only stage3e_prime_minus_one_offsets_v1. It executes 256 candidates, records offset/direction/reset metadata, leaves generated outputs ignored, adds a future 192 candidate Mersenne/perfect-number probe to the queue without executing it, and makes no solve claim.

Stage 3H implements the shared reset/advance state machine and runs only stage3h_reset_advance_ablation_v1. It executes 64 bounded candidates across Vigenere and prime-stream adapters, records reset/advance and metadata-support status, writes 100 ignored family-specific negative controls, and makes no solve claim.

Stage 3I reuses the bounded Vigenere key-pack executor and runs only stage3e_vig_history_key_pack_v1. It executes 56 candidates for the 14 declared historical motif keys across reset modes none and line and advance modes runes_only and token_break_preserving. The top key SELFRELIANCE remains calibrated noisy; generated outputs stay ignored and no solve claim is made.

Stage 3J implements the bounded Mersenne/perfect-number stream probe and runs only stage3j_mersenne_prime_stream_tiny_v1. It executes 192 candidates from the finite declared exponent sequence, reports 96 duplicate stream signatures, leaves generated outputs ignored, and makes no solve claim.

Stage 1B Atbash-Family Fixtures

The workbench now includes known-solved reverse Gematria and rotated reverse Gematria reproduction fixtures. These run through:

.\.venv\Scripts\python.exe -m libreprimus.cli solved-fixture stage1b-smoke `
  --direct-fixture-dir data/fixtures/solved-pages/direct-translation-v0 `
  --atbash-fixture-dir data/fixtures/solved-pages/atbash-family-v0 `
  --candidate-dir data/normalized/corpus-candidates/rtkd-master-v0-candidate `
  --direct-out-dir data/normalized/solved-baselines/direct-translation-v0 `
  --atbash-out-dir data/normalized/solved-baselines/atbash-family-v0 `
  --allow-pending `
  --allow-warnings

These fixtures are regression baselines, not new solve claims. Generated outputs remain ignored, and canonical_corpus_active=false.

Stage 1C Vigenere Baselines

Stage 1C adds explicit-key Vigenere known-solved fixture reproduction for DIVINITY and FIRFUMFERENFE, plus reference-source locks for selected scream314/cicada3301 and lipeeeee/gematria files. These are provenance and test fixtures only: no new page is solved, no key search is implemented, and canonical_corpus_active=false remains required.

Stage 1D Prime-Stream Baseline

Stage 1D adds p56 An End known-solved reproduction using a CPU-only prime_minus_one_stream transform with phi_prime_stream recorded as an equivalent alias for prime inputs. The p56 hex block is preserved as a payload check, not merged into plaintext. No prime-stream search, scoring, CUDA, or corpus activation is implemented.

Stage 2B Result Store

Stage 2B adds generated JSONL and SQLite result stores for solved-baseline regression imports. Run records preserve manifest SHA-256, registry SHA-256, git commit, profile/source provenance, and explicit false flags for canonical corpus activation, page-boundary finalization, search, scoring, CUDA, and canonical trust. Generated result-store outputs under experiments/results/result-store/ remain ignored and are not publication artifacts.

Stage 2C CI

Stage 2C adds .github/workflows/ci.yml plus local scripts under scripts/ci/. CI is raw-data-free, CUDA-free, secret-free, and does not upload generated corpus or result artifacts by default. The Python job runs Ruff, pytest, package smoke, transform-registry validation, solved-baseline manifest validation, and result-store manifest validation.