jbergner
f7f3460f6c
2.0 KiB
GLPI NeuroForge Mega v1.5.6
v1.5.6 hardens structured LLM output after a live BitLocker staging test exposed invalid character \\ in string escape code even after the syntax-repair pass. The failure was caused by local-model JSON strings containing literal Windows/registry/Markdown backslashes that were not valid JSON escapes.
Structured-output hardening
- Provider-native JSON mode: Ollama structured calls for staging synthesis, syntax repair, claim verification and grounded rewrite now send
format: "json". - Strict schema decoding: structured objects reject unexpected fields instead of silently accepting arbitrary model-added keys; arbitrary prose surrounding an object is no longer stripped/accepted.
- Narrow deterministic escape repair: when strict decoding fails, NeuroForge may preserve a literal backslash by JSON-escaping it only when it occurs inside a JSON string and is followed by a character that JSON does not define as an escape.
- Valid escapes are untouched:
\\,\",\n,\t,\r,\/and valid\uXXXXescapes retain their JSON semantics. - Fail-closed remainder: malformed quoting, control characters, multiple JSON values, invalid code fences, unknown schema fields and other syntax defects are not broadly normalized. They either enter the single bounded syntax-repair pass or fail staging.
- Shared verifier protection: the same strict decoder/escape handling protects claim-verifier JSON and grounded rewrite JSON, not just the first synthesis response.
Production-quality gates retained
All v1.5.5 source-authority, authoritative-action, claim-coverage, contradiction and human-review-only promotion gates remain enabled. This release changes structured transport/parsing robustness; it does not weaken evidence validation.
Regression coverage
Tests cover Windows paths such as C:\Windows\System32, registry-like paths, a model-emitted backslash before a Markdown backtick, unknown schema fields, Markdown fences, missing outer braces, one bounded repair call and Ollama native JSON mode.