Add measured native routing fidelity development and validation results

Keep historical scores, training data, model artifacts and their licenses unchanged.

Co-authored-by: Codex <noreply@openai.com>

README.md

@@ -73,16 +73,43 @@ model-index:
 
 # SemanticRepair-270M
 
-A 270M rewriter that sits behind an embedding router. When a question does
-not land on any capability with enough margin, this model restates it in
-the plain form the capabilities are described in, and the router tries
-again on the restatement. When it reads no request at all, it says so.
+A 270M model that proposes request lines for an embedding router, or
+NO_REQUEST when it finds no request. It does not answer the user's question.
+The host application owns capability admission and tool execution; model
+output alone does not authorize an operation.
 
-That is the whole job. It does not answer questions, it does not decide
-anything, and nothing it writes is ever executed: the router runs on the
-restatement, the tool runs on the original.
+SDG's evaluated native path admits individual operations with their own
+queries, which can contain the model's repaired request text. This report
+checks those admitted per-operation queries. It does not execute graph tools
+or establish the correctness of their results.
 
-## Latest measured routing candidate — 5 October 2026
+## Frozen routing-fidelity validation — 5 October 2026
+
+**The frozen candidate reduces unwanted automatic admissions but does not improve overall completion on this validation bank.** Complete supported requests fall from 20/80 to 9/80, while correct refusals increase from 9/16 to 15/16. The result does not establish a generally improved replacement or 100% routing accuracy. Model weights are unchanged. These are native routing/admission measurements; no graph tools or answer generation execute.
+
+| Measure and denominator | Live-release baseline | Frozen candidate |
+|---|---|---|
+| Complete operation multiset + required literals, or correct refusal | 29/96 (30.21%) | 24/96 (25.00%) |
+| Complete supported requests | 20/80 (25.00%) | 9/80 (11.25%) |
+| Correct refusal on unsupported/no-request cases | 9/16 (56.25%) | 15/16 (93.75%) |
+| Unwanted capability or excess-operation admission | 34/96 (35.42%) | 5/96 (5.21%) |
+| No operations returned for supported requests | 15/80 (18.75%) | 65/80 (81.25%) |
+| Complete requests requiring multiple operations | 0/32 | 1/32 |
+| Noisy family: complete operation/argument result | 5/8 | 2/8 |
+| Native/raw model measurement errors | 0/96; 0/96 | 0/96; 0/96 |
+
+
+The less strict capability-set measure is 33/96 for baseline and 26/96 for candidate. The candidate's 16 automatic admissions comprise 9 complete supported requests, 5 cases with an unwanted capability, 1 incomplete operation set and 1 correct operation set with a failed required literal. “Correct refusal” includes unresolved route choices. This table uses actual native Chat; direct-router baselines from different harness versions are not compared.
+
+**Model-only request-presence classification is 88/96 (91.67%) in both conditions on this bank:** 82/88 messages containing a request are recognized and 6/8 nonrequests are rejected. Protected literal retention is 56/70. These unchanged model-only scores are separate from native routing and from the historical 140/160 score on a different bank; model weights are unchanged.
+
+The 504-row development bank and the separately frozen 96-row validation bank are distinct. Labels are agent-authored, without independent human review; translated pairs are dependent. A post-hoc audit finds two history-dependent messages in the fresh-conversation fixture. Original labels and 96-case scores remain intact; excluding only those two cases yields 27/94 versus 23/94 complete results/correct refusals, so the coverage tradeoff remains. No candidate tuning follows this validation round.
+
+On the separate, already exposed 184-case development regression, complete capability sets or correct refusals improve from 57/184 to 74/184; supported sets from 50/156 to 52/156, negative refusals from 7/28 to 22/28, and multi-capability sets from 1/52 to 8/52. That bank has no operation multiplicity or argument annotations. Its development improvement does not override the stricter independent validation result above.
+
+Measured candidate source: `3f6ee71f95902607e490a1f8824b43575273f409`. Baseline production source: `8f773b82`, run in isolation with benchmark-only instrumentation. The candidate is experimental and this evaluation does not establish successful tool execution, answer quality, deployment readiness or a generally improved replacement. [Full tables, all four development experiments, frozen validation traces, operation labels and limitations](https://huggingface.co/datasets/Gramscii-IT/semantic-repair-routing/blob/main/eval/20261005-routing-fidelity/README.md) are maintained in the existing `Gramscii-IT/semantic-repair-routing` dataset. The additional evaluation retains PolyForm Noncommercial; training data retain CC BY 4.0 and model artifacts retain Gemma terms.
+
+## Earlier measured routing candidate — 5 October 2026
 
 The completed evaluation measures the unchanged v22 Q8_0 model with an experimental SDG admission correction. **Model-only request-presence accuracy remains 87.50% (140/160). Native system exact sets or correct refusals reach 42.50% (68/160) on the development regression bank.** These are different measures: detecting a request does not establish correct routing.