data: manifest.json with SHA256 hashes for all 41 bgz7 shards Release v0.1.0-bgz-data created with 41 assets (685 MB): qwen35-9b-base (4 shards), qwen35-9b-distilled (4 shards) qwen35-27b-base (11 shards), qwen35-27b-distilled-v1 (11 shards) qwen35-27b-distilled-v2 (11 shards) hydrate --download MODEL now works against this release. hydrate --verify MODEL checks SHA256 from manifest. https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK#71
Merged
Conversation
- data/.gitignore: *.bgz7 never committed (reproducible from HuggingFace) - data/manifest.json: SHA256 + source URLs for all 6 models (committed) - palettes/: PAL8 files committed (non-reproducible NARS artifacts) - src/manifest.rs: feature-gated (hydrate) — serde + sha2 for manifest IO - src/hydrate.rs: binary — --list, --download, --reindex, --verify - Library stays zero-dep. Hydrate deps behind `--features hydrate`. cargo check compiles clean. cargo check --features hydrate adds serde+sha2. bgz-tensor is in workspace exclude — workspace never sees hydrate deps. https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK
…y default No feature = palette-only (4 KB). Consumer picks what they need: qwen35-9b 80 MB — quick thinking, shallow HEEL routing qwen35-27b-v1 174 MB — Opus 4.5 behavior (deep reasoning) qwen35-27b-v2 174 MB — Opus 4.6 precision (code/format) qwen35-full 430 MB — all variants Railway deploy with `features = ["qwen35-9b"]` downloads 80 MB. Without any feature flag: zero download, palette routing only. hydrate --download (no model arg) fetches all enabled models. hydrate --list shows enabled/disabled/hydrated status per model. https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK
hhtl_cache.rs: palette + distance table saved as {model}_hhtl.bgz
Format: "HHTL" + k(u16) + k×Base17 + k×k×u16 + k×u32 radii
k=256: 140,294 bytes — fits L2 cache
load_or_build(): try cache file first, build from Base17 rows if missing.
Furthest-point sampling for palette construction (CLAM-style coverage).
Deployment:
PAL8 (4 KB) → HEEL routing (always present, committed)
HHTL (140 KB) → HIP distance table (built on first use, saved alongside)
bgz7 (17+ MB) → TWIG per-row lookup (feature-gated download)
BF16 (never) → LEAF (stays on HuggingFace)
https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK
HIP level: 64 archetypes, 10 KB total — L1 cache resident. 9B model: ~640 unique patterns → k=64 gives ~93% coverage 27B model: ~4096 patterns → k=64 gives ~76%, use k=256 HHTL instead build_hip() for 9B, build_full() for 27B. as_p64_distances() exports 64×64 matrix for Palette64::attend(). https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK
… data Route table: precomputed cascade decision for every archetype pair. Skip(0): pair doesn't interact (60%) Attend(1): direct attention, score = distance lookup (35%) Compose(2): multi-hop through intermediate archetype Escalate(3): HIP can't decide, needs TWIG Base17 L1 (5%) Inference: route(a, b) → O(1) action lookup. No cascade at runtime. The prefetch loads decisions, not scent. k=256: 206 KB (distances + routes + palette + radii) k=64: 14 KB (fits L1 cache for p64 HIP level) https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK
SavantDispatch: Core (14KB L1) → Psychology (206KB) / Linguistics (206KB)
Core always loaded, specialists lazy-attached on escalation.
Dispatch via ScentByte S/P/O plane analysis:
S agrees, P doesn't → Psychology (behavioral patterns)
P agrees, S doesn't → Linguistics (structural analysis)
Both agree → merge from both backends
Not user-facing agents — infrastructure backends like database indexes.
ThinkingStyle (contract) = user control knob → CascadeConfig → Savant routes.
manifest.json: savant entries with k, size, description.
SAVANT_INTEGRATION.md: full backend infrastructure plan.
37 tests passing.
https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK
Savant agents are spawned Claude agents with domain expertise, not Rust structs with route tables. The 34-tactic prompt + savant analysis outputs are the artifacts. No code wrapper needed. https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK
Release v0.1.0-bgz-data created with 41 assets (685 MB): qwen35-9b-base (4 shards), qwen35-9b-distilled (4 shards) qwen35-27b-base (11 shards), qwen35-27b-distilled-v1 (11 shards) qwen35-27b-distilled-v2 (11 shards) hydrate --download MODEL now works against this release. hydrate --verify MODEL checks SHA256 from manifest. https://claude.ai/code/session_01M3at4EuHVvQ8S95mSnKgtK
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
AdaWorldAPI
added a commit
that referenced
this pull request
Apr 30, 2026
…ring F1 (MySQL <-> SPO oracle parity) shipped via MedCareV2 PRs #1, #2, #3, medcare-rs PR #71, and lance-graph PR #309. The vision doc still claimed F1 was "the next concrete deliverable". Rewrite section 7 to: state F1 has shipped, describe the LanceProbe -> ParityWitness -> DriftSink flow, name the contract DTO (lance-graph-callcenter::transcode::parallelbetrieb::DriftEvent), list F1's known gaps (no latency claims; in-memory ring buffer), and state F2 RBAC+audit wiring (medcare-rs adopting RlsRewriter) as the next posture. No other sections touched.
3 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.