AI Research OS reading: retention without an acceptance gate

Working note, 2026-08-19. Fifth case, read against the baseline's standing question — does the response taxonomy fit or break. Pinned authority is the code-grounded review (commit dc66605).

The reading

The operative oracle is deterministic scripts plus four structural lint checks: they discriminate structure (orphans, missing hubs, broken links) and origin (seed vs discovered, with fixed scores 1.0/0.8/0.5). Weakly or not discriminated: source-page faithfulness to raw evidence, synthesis quality, explanatory quality — the review is explicit that there is "no repository-wide validator for those promises" and that lint "does not compare a hub or thesis back to raw evidence."

Explanation's position is the opposite corner from ScienceFlow: it is the medium. Wiki syntheses, comparisons, contradictions, and open questions are the system's main content — fully represented, marked (> Synthesis: prefix, per-claim wikilinks), and never scored.

The conjecture's second condition barely applies, though, because there is no reject-capable acceptance step anywhere: every written page is retained, and later ingests rewrite pages in place. Two consequences worth keeping:

  1. Selection moves to read-time. With universal retention, whatever discrimination the loop has lives in attention routing — index ordering from fixed origin scores, the query read ladder, word budgets, read caps. These discriminate origin, recency, and length, not quality. The weakly-discriminated conjecture is stated over acceptance and retention; this case suggests the same mechanism operates over read probability when retention is universal — a candidate scope extension rather than a counterexample.
  2. The dominant failure channel is compounding, not drift. Downstream writers consume source pages and are forbidden from reading raw material, so an unfaithful source page propagates into hubs and syntheses unchecked — the review's "cumulative abstraction error." Where the other four systems risk a population slowly enriched for form-fit, this one risks unfiltered error amplification through a mandatory derivative path: a no-oracle corner rather than an unequal-oracle case.

Taxonomy test

A fourth response to the form-vs-explanation asymmetry, not in the baseline's three: mark-and-expose — label LLM judgment, cite per claim, keep raw/ immutable, and defer discrimination to the eventual consumer. The catch: the label is instruction-level and the loop itself consumes marked content as its authoritative input, so marking bounds nothing operatively. That is the read-time analogue of the baseline's schema corollary: a marker, like a first-class slot, is representation, not selection.

Five systems now suggest the taxonomy is two-dimensional rather than a single list:

explanation participates in the loop explanation excluded/absent
authority contained Eigenius (grade cap), Commonplace (declare-and-audit, advisory force) ontology draft (exile)
authority uncontained AI Research OS (mark-and-expose: labeled but fully operative) ScienceFlow (unrepresented, unmanaged)

Participation (is explanatory content inside the production loop?) and containment (is its authority bounded relative to its oracle strength?) vary independently. If this 2×2 survives the next case, it — not the flat response list — is the candidate note.

Neighbor note

AI Research OS is Commonplace's nearest neighbor in medium (markdown, source/synthesis separation, inspectable files), and the review already locates the divergence at where governance lives. In this workshop's terms: near-identical declared contracts, very different operative oracles — Commonplace backs part of its contract with validators and verdict gates; AI Research OS backs its contract with instruction compliance plus structural lint. A useful pair for eventually testing whether operative-oracle strength, not contract quality, predicts corpus quality over time.

Relevant notes: