Skip to content

Chapter 17 — Opt-in surfaces: watch, batch, crosscheck, atlas

Status: EXPERIMENTAL (all four surfaces) · Prerequisites: Chapter 10, Chapter 14, Chapter 18

Mental model

Four surfaces the default deployment does not have, each with a named gate and a named ceiling. They are opt-in because of what they lack, not because they are new. Enable only when you accept those limits.

Explanation

Default product tools/list uses OCCAM_PROFILE=reader (8 tools). Set OCCAM_PROFILE=full for all 15 core tools. Four additional env flags add 6 more MCP tools (opt-in are not profile-filtered — OCCAM_PROFILE=reader with OCCAM_CONSENSUS_MCP=1 still exposes occam_crosscheck). A fifth flag adds browser interact.

Env gate Tool(s) Class
OCCAM_WATCH_MCP=1 occam_watch EXPERIMENTAL
OCCAM_BATCH_MCP=1 occam_batch_submit, occam_batch_status, occam_batch_results EXPERIMENTAL
OCCAM_CONSENSUS_MCP=1 occam_crosscheck EXPERIMENTAL — multi-source comparison only
OCCAM_ATLAS_MCP=1 occam_failure_atlas EXPERIMENTAL
OCCAM_BROWSER_ACTIONS_MCP=1 occam_browser_interact EXPERIMENTAL — declarative browser actions + materialize

occam_watch (change monitoring)

  • No daemon — cadence is the agent's job.
  • No un-watch — IWatchStore.Remove has no product caller; URL set is uncapped while per-URL history caps at 64 entries.
  • Corrupt store can silently reset to empty; multi-process writes race.
  • History entries may carry signatures when receipts are enabled; history_verified requires all entries signed and verified (chain integrity is reported separately for unsigned chains).

Batch jobs

  • Produces no Receipt v1 on job results.
  • Retains full markdown in jobs.json indefinitely — no delete API.
  • Last-writer-wins across processes.
  • Related but distinct: --batch-server HTTP mode (loopback, no auth) vs MCP batch tools.

occam_crosscheck (multi-source comparison)

  • Same host, same process, same egress IP, same proxy — vantages differ by engine (HTTP vs browser) and session cookies.
  • Agreement excludes one cloaking axis (bot-vs-browser, anon-vs-authed).
  • Verdict is unsigned observation — no shipped tool re-derives it from vantage receipts.
  • Never "consensus proof." Never multi-node attestation. Never proof of correctness.

Per owner decision OD-8: canonical concept is multi-source comparison / source agreement.

occam_failure_atlas

  • Per-session in-memory telemetry only — not durable analytics.
  • Enabling it replaces the host telemetry sink.
  • Not a cross-session leak (atlas is session-scoped).

Should you enable?

Need Consider Reject if
Notice page hash changes over time watch You need a daemon, guaranteed eviction, or multi-process safety
Bulk URL jobs via MCP batch You need signed results or retention control
Compare HTTP vs browser fingerprints crosscheck You need proof of genuineness or third-party attestation
Session failure patterns atlas You need durable ops analytics

CHECK

LOCAL. Enable OCCAM_CONSENSUS_MCP=1 together with OCCAM_PROFILE=reader. Call tools/list and assert occam_crosscheck appears — opt-ins are not profile-filtered.

NETWORK (optional). Run crosscheck on a URL; inspect that the verdict JSON has no signature and cannot be verified by any occam_verify mode.

Common misconception

"Crosscheck proves the content is genuine." All vantages leave one process, one egress IP, and one proxy configuration. Agreement is an unsigned observation that narrows one cloaking hypothesis — not consensus proof, not multi-party attestation, not cryptographic evidence of correctness.

Limitations

  • All four surfaces are EXPERIMENTAL — behavior and retention gaps are known.
  • Batch: no receipts, unbounded retention, no delete API.
  • Watch: no daemon, no un-watch, concurrency races.
  • Crosscheck: unsigned verdict; DO NOT_DOCUMENT_AS_FEATURE for trust claims.
  • Atlas: session memory only; enabling swaps telemetry sink.
  • None of these surfaces establish truth, origin, identity, or trusted time.