docs: align README and docs claims with default behavior; correct telemetry privacy, churn and funnel wording (CLAIM-01..05, TEL-01/06/08 docs)

[Dumbris]

Problem

README and docs entry pages promised more than default behavior, quoted unattributed percentages, and the telemetry page contradicted receiver behavior and over-read two fields. Docs-only; no code or schema changes.

Findings and revalidation (on origin/main c9c1ad2)

• CLAIM-01: README "scans ... before anything leaves the machine" was wrong. Detection is async and flags the activity log only; response redact/block is opt-in, text-only, and does not cover arguments. README, intro, sensitive-data-detection, activity-log and output-sanitisation now say so.
• CLAIM-02: README/intro called quarantine a TPA blocker. Reworded to withholding unapproved servers/tools and holding changed definitions; added "What quarantine does and does not cover" (detector misses, implementation changes outside the definition hash, trust modes scan/auto, quarantined:false, clean scan is not proof).
• CLAIM-03: removed the unattributed 99% / 43% accuracy figures. Added a README section with first-party bench conditions. Revalidation found bench/README.md's 17%/43% table is stale: the current go run ./bench/cmd/bench shows the proxy menu larger than the 45-tool baseline (2625 and 1874 vs 1730 tokens). README and bench/README.md now carry the current numbers and say it measures initial menu size only (no latency, workflow or accuracy claim).
• CLAIM-04: "one function" replaced by "a small set of discovery and execution tools". tools/list checked per mode: retrieve_tools 12, code_execution 7, direct exposes upstream tools plus describe_tool (README notes the direct exception).
• CLAIM-05: README quick start now follows add -> review fetch -> rescan -> approve -> harmless read -> activity, links the published quick-start, and qualifies the activity log (bounded retention and per-record size cap, not a compliance guarantee).
• TEL-01 (docs): telemetry.md no longer says IPs are stripped; it states the receiver does not store the raw IP, that older rows may still hold one, and that clearing history is a separate step.
• TEL-06 (docs): previous_shutdown/last_error_code described as prior-session evidence, not the cause of terminal silence.
• TEL-08 (docs): funnel guidance separates prevalence from conditional conversion, notes direct routing can activate without retrieve_tools, and asks for denominators and unknown counts.

Local verification

Clean scratch instances (scratch HOME/data, ports 18985-18989, telemetry off). Full transcripts, raw outputs and screenshots under docs/qa/assessment-2026-10-11/evidence/c4/ (not committed).

• README path executed on a rebuilt binary against a clean instance: add, review, approve, harmless read (Context7 resolve-library-id) succeeded, activity row visible.
• Gate: calls to a quarantined server return QUARANTINED_SERVER_BLOCKED with a policy_decision row; unapproved tools return TOOL_BLOCKED.
• Fake canaries: default config forwards argument and response unchanged and flags both in activity; response_action: redact redacts; response_action: block blocks with isError. Redact replaced only the PEM BEGIN line and left the key body (documented as best-effort; a possible separate product defect).
• Docs site built with npm run build (no broken links), served locally, screenshots taken.
• go test ./cmd/release-gate/... ./bench/... pass.

Compatibility and rollout dependencies

• Docs only. The telemetry privacy wording describes the receiver contract from the assessment (raw IP not persisted). The receiver change (smart-mcp-proxy/mcpproxy-telemetry, branch fix/assess-ip-minimization) had no commits yet when this was written, so merge this after that receiver PR is deployed, or reconcile the exact stored fields (geography granularity) with it first.
• Not changed here: docs/getting-started/quick-start.mdx approve step (security approve can refuse on current main), docs/img/mcpproxy-architecture.jpeg text, and the public website repo (src/pages/telemetry.astro:104,135 still say IPs are stripped; src/pages/docs/features.astro:29 still cites ~99%).
• Other PRs edit the telemetry.md activation section; this one touches only privacy, churn and funnel text.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying mcpproxy-docs with    Cloudflare Pages

Latest commit:	57b450c
Status:	✅  Deploy successful!
Preview URL:	https://156ecc95.mcpproxy-docs.pages.dev
Branch Preview URL:	https://docs-assess-claims-and-telem.mcpproxy-docs.pages.dev

View logs

[github-actions]

📦 Build Artifacts

Workflow Run: View Run
Branch: docs/assess-claims-and-telemetry-docs

Available Artifacts

• archive-darwin-amd64 (31 MB)
• archive-darwin-arm64 (28 MB)
• archive-linux-amd64 (19 MB)
• archive-linux-arm64 (17 MB)
• archive-windows-amd64 (31 MB)
• archive-windows-arm64 (27 MB)
• frontend-dist-pr (0 MB)
• installer-dmg-darwin-amd64 (27 MB)
• installer-dmg-darwin-arm64 (24 MB)
• smart-mcp-proxymcpproxy-goCW1JG4.dockerbuild (0 MB)

How to Download

Option 1: GitHub Web UI (easiest)

1. Go to the workflow run page linked above
2. Scroll to the bottom "Artifacts" section
3. Click on the artifact you want to download

Option 2: GitHub CLI

gh run download 38206388404 --repo smart-mcp-proxy/mcpproxy-go

Note: Artifacts expire in 14 days.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!