Prior work: the epistack framework-additions menu
Extracted 2026-10-24 from the deleted workshop
kb/work/epistack-framework-additions/ (added 028c9278c on 2026-07-08,
deleted 511c8f198 on 2026-07-23 after the FLF Epistemic Case Study
Competition; last state at 511c8f198^). The casework ran in the sibling
epistack-casebooks repo. Only the decisions that bear on the conjecture
store are kept here; the workshop is not restored.
Decisions the store reaches again
| Epistack decision (2026-07) | Conjecture store counterpart |
|---|---|
| The neutral map is the durable object; a verdict is a separate, labelled, downstream artifact computed over it and never merged in. | The store is primary state; theory and current design are projections computed from it. |
Bare confidence and status fields rejected: a credence is not recomputable from ground truth, so as frontmatter it is a stale trusted cache. |
The level is computed from the record fields and never stored. |
| Independence clustering: ten papers sharing a dataset, method, authorship or citation lineage are one piece of evidence, not ten. | Witnesses carry a source so independence can be judged; corroborated needs two independent sources. The epistack version is more developed: it names the shared dependencies to check. |
| Gap register (what is not known) distinct from crux (what would move the assessment, verdict-relative). | missing is the gap; trigger plus the refuters are the crux. |
| Belief ledger: previous state, current state, the update, the evidence responsible, what would reverse it. | The witnesses and criticism record, appended and never rewritten. |
| Support and rebuttal live on attributed edges, not as a stance scalar on the claim. | Criticism entries name the critic; the conjecture text carries no verdict. |
| Convergence through decorrelated adversarial review: a model checking its own claim under the same prompt is a correlated, weak oracle. | Same-source agreement counts as one witness; a critic from a different model family or a person. |
What the store adds that the menu lacked
The menu was built for contested domains, where motivated actors produce
the evidence. It had no account of the uncontested bulk: many small
decisions bound to a project's own choices, decided once and reused. The
choices field and the design projection are that account. The menu's
machinery still applies there, with the threshold set low.
Decisions that survived into the library
- Source genre as one open field on the snapshot, ADR 045.
- The verbatim-quote verifier, ADR 046.
- Register demoted to a default profile under open-ended text contracts,
ADR 042, which is why a casework collection needs only its own
COLLECTION.md, not a taxonomy change. The same route serves the conjecture collection.
Unfinished business worth carrying
- The silently-averaged mechanism was never tested. The menu named it as its most load-bearing untested assumption and designed a cheap experiment. It is now a candidate conjecture in candidates.
- The sibling repo is a second project. The first candidate conjecture
asks for a project other than Commonplace whose working notes factor into
bound and unbound claims.
epistack-casebooksis one, with its choices in itsAGENTS.mdand its casework inkb/<case>/notes/. - Independence as shared dependencies. When the witness record is revised at stage 1, add the dependency kinds the menu listed: dataset, method, authorship, citation lineage, and for agents the model family and prompt.