Workshop: Epistemic Architectures

Goal

Compare the epistemic architectures of six research systems — how claims, evidence, warrant, and explanation route through each system's acceptance machinery, and what its operative oracle actually discriminates. The six cases:

  • Commonplace — this KB: natural-language artifacts, declared quality goals, verdict/report review pairs, freshness baselines.
  • Eigenius — typed kernel substrate: certificates, epistemic grades, path-specific commit gates, optional Lean proof checking; pinned authority in the code-grounded review.
  • ScienceFlow — long-horizon research-agent harness: evaluator-gated Stage acceptance, anchor selection, folded memory; pinned authority in the code-grounded ingest.
  • AI Research OS — skill-driven personal research wiki: immutable sources, LLM-maintained synthesis pages, deterministic indexes, structural lint, no reject-capable acceptance step; pinned authority in the code-grounded review.
  • The ontology draft — a private research-ontology design for an artifact-focused AI lab (v0.2, examined 2026-08-19): measurement-only acceptance over versioned objects, with explanation represented but unscored. Kept anonymous in all committed text until the authors approve naming; the source document is not ingested.
  • ARC skill — prediction-gated game-solving harness: pre-action admission, post-action consequence grading, event evidence, prose notes, executable replay, provenance-bound plans, and halt-on-surprise execution. Its code-grounded reading is pinned to ARC commit dba53c3.

"Epistemic architecture" deliberately names the comparison plane, not a class of systems: every system has one, possibly degenerate, so "system X's epistemic architecture lacks a claim object" is a sayable finding. Whether a given system's machinery is genuinely epistemic is an output of the comparison, not a membership condition.

The selection angle that seeded this workshop is weakly discriminated qualities tend to be underselected; the lineage angle was already run for two of the systems in the lineage-mechanisms workshop and is not this workshop's job.

Threads

  • four-system-baseline.md — the opening position from the 2026-08-19 analysis: per-system operative-oracle readings, the four-way spectrum of where explanation sits relative to the oracle, the route-level asymmetry variant, and the response taxonomy (exile / grade-cap / declare-and-audit)
  • ai-research-os-reading.md — the fifth case: retention without an acceptance gate, selection moving to read-time as a candidate scope extension of the conjecture, the mark-and-expose response, and the participation × containment 2×2 that reframes the taxonomy
  • arc-skill-reading.md — the sixth case: an authority-route ledger separating target, oracle, timing, force, epistemic authority, and operational authority; the system-level participation × containment 2×2 fails unchanged
  • Analyse an External System's Epistemic Architecture — promoted workshop result: a collection-neutral review procedure that inventories epistemic objects, separates truth-apt transformations from non-truth-apt adaptations, and traces checking, acceptance, integration, and behavioral force per material route
  • Response taxonomy (failed unchanged) — the fifth case's participation × containment 2×2 cannot assign ARC one system-level cell. Participation and containment remain useful only as route-qualified questions; no unchanged taxonomy note should be promoted.
  • Route-level oracle asymmetry (open, candidate addition) — asymmetry between commit routes into the accepted population, not just between qualities; Eigenius is the original witness. ARC strengthens the need to type each route's target and force, but is not positive evidence for explanatory-quality underselection.
  • Behavioral-authority fold (completed 2026-08-31) — the behavioral-authority decomposition proposal now uses ARC as a bounded worked case grounded in the pinned live-action implementation ingest. The case exercises applicability, path position, and consequence without selecting an ontology or warranting a standalone theory note.
  • operator-response.md — the draft's author responded (2026-08-19, corrected 2026-08-20): the draft is intended for publication, and the stated intent is lab tooling à la Neptune.ai, not theory production — reframing exile as deliberate scoping (the operator concurs that lab tooling is right to be evidence-heavy) while leaving the drift prediction and the tracking-vs-knowledge-production tension open
  • Attribution of the ontology draft (open, external) — the author intends to publish the draft; once it is public, snapshot and ingest it and restore the anonymized references here and in the weakly-discriminated note to full citations. Until then everything stays anonymous.

What closes this workshop

  1. The durable claims extracted: the failed system-level response taxonomy either replaced by a route-qualified comparison or explicitly dropped; the route-level variant either added to the weakly-discriminated note with its warranted evidence edges or rejected with a reason.
  2. The comparison itself consumed: anything worth keeping lands in library notes or an article; the workshop's tables are not promoted as-is, workshops being sinks.
  3. The attribution question answered or parked with the contact.
  4. The ARC worked-case fold into the behavioral-authority proposal either completed in a separate run or explicitly parked — completed 2026-08-31.

Evaluation boundary

Evidence grades differ by system and the comparison must not flatten them: Eigenius and ScienceFlow claims rest on pinned code-grounded reviews; ontology-draft claims describe an unpublished design, not a running system — no loop has run, so all selection effects there are predicted by construction; Commonplace claims about itself can cite operational episodes (the Popperian-maintenance record) but are this installation's behavior, not the framework's in general. ARC's architecture claims are pinned to inspected code; its campaign quantities remain repository-reported because no run directories or component ablation were supplied.

Bookkeeping

Plain markdown, workshop register. Started 2026-08-19 from the conversation that analyzed the ontology draft against Commonplace, added the designed-in example to the weakly-discriminated note, and re-read the Eigenius and ScienceFlow analyses through the selection angle. ARC was added on 2026-08-20 as the sixth case; it failed the system-level 2×2 unchanged and moved the working comparison to authority routes. The route-level method was promoted on 2026-08-20 as a reusable instruction after cold applications to ARC and held-out GBrain; the workshop remains open for its separate explanatory-underselection and attribution threads.

Complete file listing (generated at build time)