## Compound routing follow-up

This experimental candidate is [SDG PR #658](https://github.com/Gramscii-Git/semantic-deterministic-graph/pull/658), stacked on the corrected benchmark runner in #656. It is not deployed and does not retrain or replace the model. The complete final review tree at `4b640443` equals the measured tree at `020b8e6c` (`62304e00dc89963ed7bc2a018e3e5edc56246a07`). Model weights, decoding and workspace thresholds are unchanged.

The candidate checks lexical acquisitions for separate requests, keeps every confident in-scope repaired part, and asks for clarification when any part remains unresolved. Repair abstention cannot execute the original winner or continue history. Writing and conversation readers retain their subject. The territory descriptions explicitly name administrative and ISTAT municipality codes.

| Native admission measure | Before | Candidate |
|---|---:|---:|
| Original bank: complete desired set or correct refusal | 54/160 | 68/160 |
| Original bank: three capabilities | 0/20 | 5/20 |
| Original bank: two capabilities | 1/20 | 6/20 |
| Original bank: admissions with unwanted capabilities | 54/160 | 32/160 |
| Additional diagnostic bank: complete desired set or correct refusal | 3/24 | 7/24 |
| Additional diagnostic bank: three capabilities | 0/12 | 0/12 |
| Additional diagnostic bank: admissions with unwanted capabilities | 8/24 | 3/24 |

The Pisa map/ISTAT-code/population request now admits the three desired capabilities in Italian. All 12 additional triple requests remain unresolved. The original noisy family regresses from 8/20 to 4/20. This is a partial engineering correction, not evidence of reliable general multi-intent handling or production readiness. An unresolved request is not counted as a correct positive result.

The original 160-question bank is development regression material after its outputs are examined. The additional 24 questions have frozen gold before candidate edits and initial inference, SHA-256 `ab34aff9e955f30e95b8de9a0de6c65018fe5950230b400489f8f6428a3448f1`; their first outputs are subsequently used in diagnosis before the final gate changes. They are therefore diagnostic data, not an untouched holdout. Labels are authored by Atlante, without independent review; translated pairs remain dependent. No model training occurs, and no graph tools or external generation providers execute.

The combined 184-row bank has SHA-256 `4d0cb0f4f43c9460d92d6d496cc58828b48f9890c6443a10dc989e220764cec9`. Complete traces, graph snapshots, raw completions and refusal details are retained in `candidate/`. `comparison.json` pairs the original historical results and additional diagnostic baseline with the candidate; it does not replace the original published results or make population significance claims.

Local gates: 1655 unit tests, 73 focused routing/artifact checks, 51 planning/repair consumer checks, Ruff, and one real native integration test processing all 184 rows. Platform: macOS 26.6.2 arm64, Python 3.13.12, llama.cpp build 10621. Full product/database/browser suites, populated-source search, tool execution and other platforms remain untested for the candidate. Integration with the later coordinated release remains pending before merge.

### Artifact ownership correction

Use the corrected runner at `2fc0cbaf` or the candidate tree, with the included `source/bench/semantic_repair/run.py`. It copies model and engine files to independent inodes and verifies that source qualification metadata stays unchanged. The historical runner at `34996e76`, retained in the original `source/` bundle for audit, creates hardlinks: those change source ctime and have invalidated deployed SDG artifact qualification. Do not use that historical runner against live artifacts. Existing linked directories remain intact pending coordinated cleanup and requalification.

The final rerun verifies all 84 monitored source files unchanged, all 84 copies on distinct inodes, seven complete owner cleanup receipts, four released reserved listeners, an exact database backup and drop, and unchanged source workspace content. See `copy-and-cleanup-verification.json`. The source metadata incident does not change model bytes or rewrite historical measurements.