Skip to content

Add source-bound Qwen35 LCT translation compiler - #65

Draft
Inserian wants to merge 2 commits into
qsa/exact-representation-compiler-v1from
qsa/lct-qwen35-translation-v1-gated
Draft

Add source-bound Qwen35 LCT translation compiler#65
Inserian wants to merge 2 commits into
qsa/exact-representation-compiler-v1from
qsa/lct-qwen35-translation-v1-gated

Conversation

@Inserian

@Inserian Inserian commented Aug 13, 2026

Copy link
Copy Markdown
Member

Bounded QSA LCT child above exact PR64 live head 142119fbca7414195546ba5da4ad9d5708ab4aac. Draft, unmerged, and not production-promoted.

Accepted research authority

Source-bound Qwen35 LCT producer machinery is accepted on exact head eff0f7c64fd65b0cefce29f2510bab96130e799a.

Decision-bearing exact-head gates are green:

  • Structural Hardening 31737595535 — success;
  • QSA Qwen35 LCT Source Bound 31737595546 — success, including Ubuntu/macOS/Windows and ASan/UBSan;
  • Runtime Evidence 31737595554 — success;
  • Build/Test 31737595549 — success.

The accepted capability:

  • pins Qwen35 source semantics to ggml-org/llama.cpp commit 74ade52741203e5c8f81eaf06a96cb1cfe15f2a3 plus frozen source blobs;
  • compiles recurrent/SSM, full-attention and tokenizer/embedding/output source-bound CIR families;
  • preserves recurrent state-before/state-after and residual bindings;
  • freezes PR162 donor/topology identity and 48 recurrent / 672 components, 16 full-attention / 176 components, tokenizer/output / 3 components;
  • accepts the strict qelm.lct-qwen-source-fragment.v1 contract and emits only the frozen PR163 qelm.lct-qsa-translation-fragment.v1 schema/signing contract;
  • binds CIR identity to source identity, component-manifest identity, context identity and exact llama.cpp source identity;
  • uses an injected SHA-256 provider rather than private crypto;
  • fails closed unless operator/reconstruction mismatch counts are zero and both exact-equivalence gates are explicitly verified.

Current real-evidence boundary

The donor lane has now staged an occurrence-bound 39-target Qwen activation observer and v1r2 receipt certifier. Those are the next real evidence inputs. The certified JSON intentionally retains tensor_index_sha256 while the raw tensor-index.tsv carries the exact logical_name -> graph_name#occurrence mapping. QSA must independently hash/check that raw index rather than treating it as PR162's component manifest.

This acceptance does not establish real donor activation equivalence, real operator equivalence, exact reconstruction, final QSA producer fragments, end-to-end LCT acceptance, Brain truth/memory authority, merge authorization, or production promotion. Those remain false until real occurrence-bound capture evidence is frozen and source/component identities arrive without rewriting.

PR64's partition/10k scale candidate remains independently unaccepted and gains no authority from PR65.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant