OKF grounding and compact retrieval evidence

Grounding means giving a model the relevant evidence for a question and requiring the answer to stay within it. An OKF bundle helps because its content is small, addressable, source-linked and quality-labelled. It cannot make an unsupported answer true, repair a wrong source, or guarantee that a model obeys instructions.

Evidence already obtained

The Microsoft 365 Copilot development trial over 293 governed families kept all 293 assessed responses within the defined safety boundary and selected the exact expected family for 292. The one failure was a near-neighbour identity collision between a general Universal Credit family and the more specific unemployment case. That is 99.6587% strict selection accuracy for the authored development schedule, not a universal accuracy figure. Raw transcripts remain private; the schedule was not an independent holdout; OneNote and a negative permission test remain open.

This result supports four narrower claims:

  1. stable family identity and compact facts can make errors visible;
  2. direct prompts and links can work well in an existing enterprise product;
  3. near-neighbour disambiguation needs explicit aliases, scope and ranking;
  4. rate limits and permission boundaries are part of retrieval quality.

It does not isolate the effect of OKF from the model, SharePoint index, authored questions or prompt. The MCP evaluation in this review therefore measures retrieval size and expected-record selection separately from answer quality.

Compact context contract

A useful context pack contains only:

The prototype enforces byte and result limits before returning text. It reports an estimated token count (ceil(characters / 4)) only as a repeatable proxy; real tokenisation varies by model.

Evaluation layers

Layer Metric What it can establish
Retrieval expected record in top-k, reciprocal rank, ambiguity count Whether indexing selects the intended evidence.
Compactness input bytes, returned bytes, estimated tokens, reduction ratio Whether access avoids sending the whole bundle.
Attribution answer claims linked to supplied record/source IDs Whether a reviewer can trace an answer.
Answer faithfulness supported, contradicted and unsupported claims Whether the model stayed inside the evidence.
Answer correctness expert-scored exact answer and material omissions Whether the final answer meets the domain need.
Safety inaccessible-record, stale-data, rights and advice-boundary tests Whether retrieval respects constraints under adverse cases.
Operations latency, error rate, throttling and cache behaviour Whether the system works reliably at the intended scale.

The report treats a compactness win as necessary but not sufficient. An empty answer is tiny; a highly compressed wrong record is dangerous. The preferred comparison freezes one bundle and question set, then tests prompt-plus-link, MCP retrieval and any vector/semantic variant against the same ground truth.

Where OKF is strong and weak

OKF is strong for bounded, mostly documentary knowledge where identity, provenance, progressive disclosure, human review and portable publication matter. It is weak as a transactional system, rapidly changing operational database, numerical analysis engine, confidential access-control authority, general ontology reasoner or substitute for professional judgement. Those systems should be linked through governed interfaces, not copied wholesale into Markdown.