Doop: method and evidence

Doop is a local-first engineering case study and experimental reference implementation built from an independently validated and extended reverse-engineered WHOOP 4.0 BLE pipeline. The source shows raw-first packet capture, HR/RR decoding, RTC-aware buffered-history recovery, independent proxy metrics, confidence-gated MCP tools, confirmed workout/plan boundaries, and BROC's minimum-disclosure safety-gated agent path.

View the evidence-first project overview at doop-showcase.vercel.app.

This is not a turnkey or production health product. The narrow, owner-controlled Gen4 workflow is documented in Using this code, including configuration, hardware cautions, supported boundaries, and a no-hardware smoke path. The repository contains no owner database, exports, captured health packets, device identifiers, credentials, personal athlete profile, or private Git history. Test packets, plans, preferences, and health inputs are synthetic.

Start with the reverse-engineering case study. Its one-page static trace follows a generated packet through parsing, daily computation, MCP JSON, and BROC's visible answer. The hardware results ledger separates recorded owned-device evidence from checks reproducible in this data-free repository.

If you have ten minutes, use the claim → evidence → command checklist. It points each headline claim to one file and one focused verification command.

The evidence manifest is the release index: it ties headline claims and résumé language to the corresponding evidence and checks.

Architecture

flowchart LR
    W[Owned WHOOP 4.0] -->|BLE notifications| C[Capture + CRC validation]
    C --> R[(Raw packets)]
    C --> T[(Typed HR / RR)]
    R -. re-derive .-> T
    T --> M[Independent daily metrics]
    M --> G{D39 actionable?}
    G -->|yes| Q[Read-only MCP queries]
    G -->|no: withhold + reasons| Q
    Q --> L[Minimum-disclosure today / week relays]
    S[Pre-model safety + offline gate] --> B[BROC via OpenClaw]
    L --> B
    P[Plan + canonical logbook] --> Q
    D[D47 prepare → review → confirm] -->|explicit local write| P

Capture preserves evidence before interpretation. Compute retains audit values; D39 decides whether they may influence coaching. The accepted agent sees two preformatted relays, while safety/offline handling occurs before model inference. Canonical writes remain outside MCP behind explicit local review.

What this does and does not claim

• It demonstrates empirical protocol investigation, negative results, raw-first storage, replay-safe decoding, and testable domain-agent controls.
• Recovery, strain, and sleep are independent proxies. They are not WHOOP parity, medical outputs, or ground truth.
• OpenClaw supplies generic agent runtime and Telegram transport. Doop supplies the domain data model, bounded tools, confidence/confirmation policy, minimum-disclosure relays, deterministic safety handling, and evaluator.
• The protocol work extends cited prior art and independently verifies the shipped wire facts; it was not invented from nothing.

Verify the evidence

Requires Python 3.11+ and uv.

uv sync --all-extras --locked
python3 tools/privacy_check.py
uv run pytest -q
node --test integrations/openclaw/plugins/doop-safety-gate/test/*.test.js

The sanitized suite contains Python and OpenClaw safety-plugin tests that run without a database, band, credentials, or private runtime. Exact release-gate counts are recorded in RELEASE_NOTES.md after the final clean-clone run.

For demonstration-video recording only, create the isolated minimal seed. It refuses to overwrite an existing file, stores synthetic raw packets, and computes derived metrics through the production path:

uv run python -m tools.create_film_seed \
  --output tmp/film-seed.db --film-date 2037-07-06

Focused entry points:

• src/whoop_local/capture/proprietary_protocol.py and tests/test_proprietary_protocol.py: framing, CRCs, HR/RR and four-slot RR behavior;
• src/whoop_local/capture/history_sync.py and tests/test_history_sync.py: RTC-first, commit-before-ACK, replay-safe history recovery;
• src/whoop_local/compute/ and tests/test_metrics.py: independent metrics;
• src/whoop_local/mcp_server/queries.py and tests/test_training_insights.py: D39 confidence gates and minimized relays;
• src/whoop_local/integrations/openclaw_acceptance.py and the OpenClaw safety plugin tests: strict outputs, offline refusal, and pre-model safety handling.

The sanitized agent acceptance record publishes the scenario matrix, visible outputs, and both promotion revocations. The separate D47 confirmation record shows why draft preparation and review cannot silently become a logged workout.

The demonstration storyboard defines the four required agent beats, exact narration, Mac recording setup, and frame-by-frame privacy review for the v1.0 video.

Privacy and licensing

tools/privacy_check.py checks the working tree and every reachable Git blob, path, commit identity/message, and annotated-tag identity/message for generic private-artifact, PII-shape, captured-frame, and secret patterns. It reports only safe identifiers and categories, never matched content. Third-party contributions and their status are recorded in THIRD_PARTY_NOTICES.md. The project is released under Apache-2.0; see LICENSE.

Release evidence and traceable résumé language are recorded in EVIDENCE_MANIFEST.md, RELEASE_NOTES.md, and RESUME_BULLETS.md.