Skip to content

Experimental features

Features that ship but are not default product promises. Experimental ≠ invisible.

Status vocabulary: EXPERIMENTAL — enable deliberately; expect sharper edges.

Feature Enablement What it does Limitations Why not default
Exam MCP OCCAM_EXAM_MCP=1 occam_exam_submit grades a harness JSON; may change the tool surface + list_changed Does not run the four tasks for the agent; OCCAM_PROFILE_LOCK=1 freezes surface; clients must re-list after notify Capability research surface; day-to-day profile switch uses occam_client_capabilities(profile=…) without exam
Watch OCCAM_WATCH_MCP=1 Re-check a URL over time; history chain No daemon (agent polls); store races; unsigned history ≠ history_verified Costly; persistence/concurrency limits
Crosscheck OCCAM_CONSENSUS_MCP=1 Multi-source / multi-vantage comparison (source agreement) Verdict is computed, not a “consensus proof”; same-process/egress limits Expensive (2+ extracts); easy to overread
Batch OCCAM_BATCH_MCP=1 Queue many URLs No Receipt v1 on the batch envelope; store races; retention limits Operator/server mode; not a casual agent default
Failure atlas OCCAM_ATLAS_MCP=1 Session-local failure telemetry Not proof a host is a “dead end” Diagnostic, not a trust layer
Browser interact OCCAM_BROWSER_ACTIONS_MCP=1 Declarative click/type/scroll then materialize Max 16 steps; typed text redacted; no raw page JS; never cached Automation surface; keep off unless needed
Crawl OCCAM_CRAWL_MCP=1 Same-domain map→transcode under max_pages, caller-held resume_token, optional JSONL Caps 32 pages/call; no CAPTCHA unlock; not in default reader Separate product surface; keep opt-in (ADR-0018)

Forbidden readings

  • Crosscheck agreement is not proof of correctness (“consensus proof” forbidden).
  • Watch history_verified requires every entry signed and verified; hash-chain integrity alone is a different signal.
  • A typed wall is not a CAPTCHA bypass. Use a session profile or accept ok:false.