Remote MCP acceptance

These evaluations check whether a remote ask_okf call preserves the governed evidence package produced by Explorer's existing context assembler. They do not grade conversational prose or claim current legal accuracy.

Cases

• Imprisonment: reuses the existing A–H case. Required source identities, exact passages, directed relationships, traversal paths, declared requirements and authority boundaries are assessed. Keyword overlap is not a pass condition.
• Hospital admission: the hospital case must expose insufficient context for the approved snapshot. It checks missing evidence, unresolved terms, zero applicable requirements and no budget truncation. Selected custody passages must not be mistaken for a hospital answer.

The hospital source review identifies relevant frozen material outside the Ask index. Ten PDF identities and 52 complete page literals were reverified on 19 September 2026. It is a bounded research inventory, not an approved hospital profile. Read the coverage explanation before interpreting these results.

Run real MCP calls

The public test endpoint below has a retained HTTPS acceptance receipt. Its protocol-client results and actual ChatGPT observations remain distinct in the demonstration guide:

node --experimental-strip-types scripts/test_remote_mcp.mjs \
  --endpoint https://ask-okf.crpage.chatgpt.site/okf/mcp \
  --explorer-root ../okf-explorer

For local development, use --endpoint http://127.0.0.1:8787/mcp and a separate output directory, for example --output /tmp/okf-dwp-mcp-local. HTTP is accepted only for loopback hosts. Redirects, embedded credentials, query strings and fragments are rejected. Endpoint selection is a test-client option; the service's bundle selection remains an immutable allow-list.

The independent raw HTTP client:

1. initialises MCP protocol 2025-11-25, then sends notifications/initialized;
2. discovers ask_okf and checks its read-only annotations and bounded input schema;
3. makes one tools/call per exact question, using bundle okf-dwp and version efb05c66616a9cd4328a86cf412780fe7bc7cf0b;
4. requires structured content and the JSON text block to contain the same existing context package;
5. assembles locally through the unchanged Explorer core, with identical question, budget and immutable index binding;
6. compares canonical package bytes and runs the evidence/path/provenance assessment;
7. records the real transport observations and raw tool results.

The case expectations never become engine inputs. The response size and request time are bounded. No model call, web search or source refresh is part of this procedure.

Retained evidence

The default output is validation/remote-mcp/:

Raw HTTP envelope digests are recorded separately from decoded tool-result and canonical package digests. A context identifier binds the evidence package, not network duration or observation time. Compression is checked against its retained archive hash and the decompressed bytes; replay does not assume that different operating systems produce identical gzip streams.

The receipt preserves source version, scope, provenance, missing evidence and truncation information. A loopback result is labelled local-loopback-mcp; it cannot establish remote HTTPS access. A successful protocol client call does not establish ChatGPT, another AI client or Voice invocation. Their separate observations must identify the client surface and actual tool use.

Replay without contacting the endpoint

node --experimental-strip-types scripts/test_remote_mcp.mjs \
  --check --explorer-root ../okf-explorer

Replay rechecks the retained archive, raw result, case, source-index and implementation hashes; reruns the unchanged core; compares the complete packages; and reassesses evidence requirements. It does not rerun the network request and cannot show current endpoint availability. Use the real-call command for that purpose.

Changing the engine, rubric, client, source index or approved version requires new evidence. Do not edit receipts to make stale results pass. Ordinary CI uses offline replay, so it does not depend on public endpoint uptime.

Bounded client demonstrations

Actual ChatGPT invocation exposed a separate host limit: the complete default imprisonment result was returned by the service but truncated for the model. The service's budget.truncated: false does not describe downstream host behaviour. The tested 32,768-byte hospital request instead deliberately returns an insufficient package with 8 authored concepts, no relationships and budget.truncated: true; the model reported access to that bounded result and declined a hospital answer.

The bounded receipt preserves both questions at three explicit budgets. All six HTTPS packages match the unchanged core. They remain insufficient and deliberately truncated; increasing the budget does not introduce missing hospital knowledge. No service default or adapter summary was substituted.

Replay all six without contacting the endpoint:

node --experimental-strip-types scripts/check_bounded_remote_mcp.mjs \
  --explorer-root ../okf-explorer

These are bounded transport and boundary tests, not passes of the full imprisonment sufficiency rubric. A larger byte budget does not currently guarantee that the package retains more useful evidence: an additional local 131,072-byte imprisonment probe returned a metadata_budget failure. The receipt records that engine limitation; this adapter change does not fix it. Use the exact rehearsed question and budget and inspect the returned status rather than assuming a monotonic improvement.