[codex] test code-specific multi-vector evidence frontiers

[TheGreenCedar]

Benchmark-only prospective test of one pinned LateOn-Code-edge token-level relevance mechanism against the retained development fragment universe. Preserve ETR-1, STR-1, and reader failures. No production, public API, model sweep, selector adoption, or release work. Refs #2116. Full frozen contract follows before corpus output.

[TheGreenCedar]

MVR-1: one code-trained multi-vector frontier

Development diagnostic only. Refs #2116; closes no earlier experiment. ETR-1 and STR-1 remain failed. Reader diagnostics remain 3/8 complete, not an accuracy pass. This tests one different relevance mechanism, not a model sweep or an automatic planner.

Fixed inputs and method

• Base: reader head 35e4e27b5413703498693d06f99a7dd5eed9de4e; stacked benchmark-only child [codex] test code-specific multi-vector evidence frontiers #2131.
• Preparation: codestory-etr1-final.8ttvNp/prepared/preparation.json, SHA-256 30b84d4d848f96bd4fe799f2e0f28b9114971da0e47bf98ebe54fe36242199fd.
• All 10,369 retained source fragments, 72 wording rows, 24 cases, 12 repositories. Same complete-line source representation, questions, BM25 seeds, pinned publications and source-only exact oracle.
• Questions SHA-256 8e7219a59c973c02f8ea93120bb680da46a75b8272153986c76e55bfb73ca3b6; annotations SHA-256 52b0cc223292bc70f1e4fa3f52b67bf42a91e4d4b9ed997aa12c648c068e9ade.
• One model: lightonai/LateOn-Code-edge@4bcdf5ed93f791259eb130b577a240f753d68dd8, Apache-2.0. Reference PyLate 1.3.4, Transformers 4.56.2, Torch 2.8.0, MPS with fallback disabled. No ONNX/int8/native-runtime claim.
• Published model normalization strips outer whitespace and lowercases semantic text. Raw source and identities are unchanged. Prefixes [Q] / [D] , query/document limits 256/2048, 48-dimensional unit token vectors, reference punctuation masking, no pooling, query expansion, generated queries, source expansion or graph.
• Preflight every exact text against the reference tokenizer without truncation. Overflow invalidates execution; do not alter questions or fragments.
• Encode documents once in batches of 32. Query text is exactly the raw question. Exhaustive MaxSim: sum over query tokens of the maximum normalized dot product over each document's tokens. Score descending; fragment ID sole tie-break.
• Same BM25 seed list S. For m seeds, take the highest-scored 8m non-seed fragments. L is S plus these successors, unique and independently selectable, at most 144. No threshold, quota, name/path bonus or annotation-based restriction.
• Compare against retained ETR-1 unconditioned frontiers for descriptive quality only. This changes both encoder and scoring objective; it does not isolate either separately. No timing comparison against reused ETR rows.

Execution and analysis

Authenticate repository commits, complete source digests, ranges, fragment identities and serialization costs before inference. Bind source/tree, Python executable, packages, model assets, inputs and method before the run. Freeze all token receipts, document/query vectors, complete score vectors, seed/successor pools and outputs before annotation access. Validate source bindings, exact score reconstruction and pool equations before evaluation. Any missing or inconsistent row invalidates the complete run, with no quality aggregate. Read-only source reuse is allowed; partial query results are not reused after a validity repair.

One whole run; 25-minute preparation-plus-request deadline, excluding validation/oracle. No hidden warmup: report all 72 measured requests. Request interval includes query tokenization/encoding, exhaustive scoring, deterministic selection and selected source authentication/mapping; one-time document encoding and receipt serialization are separate. Report enclosing time and exclusive phases, not a sum invented as total. Reference scoring may use fixed 128-document padded batches; this is a memory bound, not search pruning. No approximate index.

After outputs freeze, use the existing exact optimizer: 16 rows/16 KiB including identical publication metadata; whole atoms; alternatives separate; seeds optional; no relation credit. Aggregate three phrasings within cases first. Frontier qualifies only at >=85% mean recall, >=70% in every group, >=75% complete-set rate, 100% source validity. Prepared frontier p95 must be <=1.25 seconds. Preserve all unfavorable results.

Also report a non-oracle baseline fixed now: score-descending candidates from L, greedily admit each full row if the exact serialized packet still fits, stop at 16 rows. Seeds receive no protection. This baseline is diagnostic only and does not authorize selector adoption. Its gap to the oracle identifies whether relevance ranking is enough for packet selection.

No output is a working packet claim. No production routing, schemas, plugin, compiler, model family swap, graph, confirmation/builder/sealed corpus, release or version changes. A frontier failure ends this mechanism. A pass permits a separately specified selection/usefulness experiment only. One independent verifier checks the fixed mutation matrix and reconstructs results; no extra review lane.

Hostile matrix

Reject changed source/commit/range/fragment, query or model; token truncation; vector shape/norm/nonfinite values; missing score; swapped document identity; wrong MaxSim; duplicate or missing row; altered seeds/exclusions/tie-break; pool overflow; altered metadata/row cost; early or unbound annotation analysis. Exercise natural underfill, UTF-8, equal scores, successors selected without roots, exact budget boundaries and complete-line source authentication. Synthetic canary uses the real encoding/scoring and analysis path before corpus execution.