v1.0.9

[1.0.9] - 2026-05-11

Source-fidelity diagnostics: per-claim unsourced_items

The source-fidelity ratio (Layer 4: digit-substring match of a
document's numbers against the source it was supposed to be
grounded in) is the capability that distinguishes Frame Check
from framing-only tools. Pre-v1.0.9, the wire format reported
the COUNT of unsourced numbers
(source_fidelity.not_in_source: 2 out of 25) but not WHICH
specific numbers. Adopters reading "23/25 in source" got the
headline but couldn't act on the 8% unsourced rate without
manually diffing source vs. summary side-by-side — which
defeats the tool's purpose for AI-output auditing.

The internal data was already computed:
clarethium_measure.source_matching() builds
unsourced_details as [{value, type, context}, ...] where
context is the claim-sentence the number appeared in
(clarethium_measure.py:843-846). The MCP composer at
mcp_compose.py:2708 was reading the summary fields and
dropping the items list.

This release surfaces verification.source_fidelity.unsourced_items
on the wire when source_text is provided. Adopters now see:

{"value": "100000000", "type": "integer",
 "context": "NVIDIA RTX now serves 100 million gamers..."}

against the load-bearing demo (Grok-on-NVIDIA worked example):
the LLM smuggled a "100 million users" figure from training data
into a summary of a press release that disclosed no user count.
That class of insertion is exactly what the source-fidelity
capability exists to surface; the headline ratio is necessary
but not sufficient — the per-claim diagnostic is the actionable
half.

Wire-format addition only (no removed or renamed fields). The
new field is omitted when no source_text is supplied (same
schema-stability discipline as the rest of the verification
block). Pinned by
tests/test_mcp_server.py::test_source_fidelity_carries_per_claim_unsourced_items.