Writing-as-thinking → process transfer
Goal
Harvest the still-unincorporated ideas from the four writing/thinking essays ingested 2026-08-10 and turn each into a Commonplace process change (review gate, FIX routing, connect/multistage step, writing convention) — or an explicit rejection with a recorded reason. The prior written-artifacts-in-learning-loops workshop (closed 2026-08-10) promoted two conjectures and closed, but its own ingests flagged ~20 extractable items and most were left as "scaffolding already covered." This workshop works that backlog.
Why a workshop, and the one hard rule
The prior workshop was killed because it accreted wrong ideas faster than it tested them. So the governing discipline here is discriminating-test-before-build: no idea gets promoted to a durable artifact until a cheap one-note (or one-pair) test shows it has signal over what Commonplace already ships. An idea that fails its discriminator is rejected in this README, not built. Distilling the counterexamples that kill an idea is fine to do inline; do not reopen the closed workshop to mine them.
What closes this workshop
Every candidate below reaches a terminal state: promoted (named library artifact) or rejected (one-line reason recorded here). When no candidate remains open, extract the survivors and delete this directory per kb/work/COLLECTION.md.
Evaluation boundary
- These are human first-person essays. Transfer to an agent/KB process is a hypothesis, not an inheritance —
human analogies can motivate functions without determining component boundaries. State each transfer as the behaviour/dependency a process must preserve, then test it. - Changed human understanding, a retained artifact, and changed agent behaviour are different outcomes; do not call written output "learning" without the retention/activation/effect links.
- A candidate that only restates an existing gate or note is redundant, not novel — check the shipped gate catalogue (
kb/instructions/review-gates/,composition-friction-gate.md,critique-note.md) before claiming novelty.
Candidate backlog
Status ∈ {open, testing, promoted, rejected}. Method and first target are deliberately left to the live session.
| # | Idea (essay source) | Residual increment over what ships / what to discriminate | Status |
|---|---|---|---|
| 1 | Premise-decomposition review (how-to-think #2) | composition-friction-gate already decomposes into inferential joints (edges) and attacks each. Increment: decompose into premises (nodes) and test each for external support / counterexample. Discriminator: on one note, does the premise-node check surface a real weakness the joint check misses? |
promoted → kb/instructions/premise-decomposition-gate.md (discriminator + generalization test passed 2026-08-10, see Test log). Registration into the review pipeline deferred. |
| 2 | Counterexample-scope routing (how-to-think #1) | Classify a defeated point local (defeats a premise → revise, keep the commitment) vs global (propagates to the commitment → retire). Reconsidered 2026-08-10: this is a review-layer property, not a discovery-lifecycle phase. Local/global is Lakatos argument-internal (premise-vs-conclusion within an explanation), while discovery-lifecycle tracks a conjecture's evidential status through phases and deliberately abstracts its internal premises away; its test phase checks consequences (forward), not premises (backward). The earlier "add a test-result branch to discovery-lifecycle" framing (and the ingest's recommended action) was over-scoped. |
mostly delivered — premise-decomposition-gate emits LOCAL/GLOBAL per premise; the full pass consumes GLOBAL→Disposition. Remaining: optional fix-routing-by-scope in FIX-SYSTEM. A discovery-lifecycle one-line cross-ref (its revise-vs-reject step may cite a scope verdict) is likely not worth it. |
| 3 | Value-of-information reading (learning-by-writing #3) | Route the next read by expected claim-change, not retrieval relevance, in connect / write-multistage. Operationalizes the open question in the claim-routed conjecture. |
open |
| 5 | Stance-reversal operator (learning-by-writing #4) | Routine "write the strongest note that contradicts this one" as an anchoring check in connect/review. Cheap; de-risks the claim-routed conjecture's own anchoring failure mode. | open |
| 6 | False-precision anti-gate / "accurately-vague" register (wordless #2) | A check for stating-as-settled what the evidence leaves open — the counter-pressure to constraining. AGENTS.md forbids invented precision but no gate enforces it. |
open |
(#4, a killed-conjecture ledger, is deprioritized per maintainer: distilling counterexamples is useful but the closed workshop is not to be reopened.)
Open discriminating tests carried in from the two promoted conjectures
- descriptive-link-labels may supply claim self-sufficiency — a label-ablation test (2026-08-10) was run: 6/6 cold-reader PASS, gate never fired even de-labeled; it reassigned the mechanism from the label convention to the body-premise convention and left the thin pre-connect-draft case open. That thin-draft test is the cheapest remaining experiment.
- claim-routed reading may beat reading-first for synthesis notes — its head-to-head test (build one synthesis note both ways, compare warranted update vs context cost) is deferred by maintainer to control scope.
Sources
- How to think in writing — Karlsson/Lakatos: conjecture, premises, local/global counterexamples (#1, #2)
- Learning by writing — Karnofsky: hypothesis as inquiry-control state (#3, #5)
- Putting ideas into words — Graham: exact-word commitment + neutral-stranger reread
- When is it better to think without words — Karlsson: expansion/compression phasing, false precision (#6)
Test log
2026-08-10 — #1 premise-decomposition discriminator (premise-node vs joint-edge)
Setup. Two title-as-claim notes resting on empirical premises — llm-generation-confidence-tracks-typicality-not-soundness.md (A) and weakly-discriminated-qualities-tend-to-be-underselected.md (B). Each read by two fresh, cold (no linked notes), no-tools adversarial sub-agents: a JOINT reader running the shipped composition-friction-gate (test each inference/edge; default UNSUPPORTED) and a PREMISE reader (assume inferences valid; decompose the claim into premise-nodes; hunt a counterexample per premise; route failures LOCAL/GLOBAL). Question: does PREMISE surface a real weakness JOINT misses?
Result — yes, on both notes, and the two are complementary. - A: PREMISE-only found (i) an RLVR/post-training counterexample to the premise "training optimizes text likelihood" — a premise stated as background, which an edge-check has no joint to attack — and (ii) a token-logprob vs verbalized-confidence conflation. JOINT-only found the automation-boundary corollary as an inference overreach. - B: heavy convergence (both caught "no-lift underselection" over-stated as "directional drift," and the Bug-That-Shipped activation-vs-selection confound). PREMISE still added: whether decorrelated maintainability signals exist to assemble (JOINT endorsed that reasoning as holding), and whether the weak quality has a stance-independent ordering at all ("underselection" as possible category error). - #2 evidence: PREMISE's LOCAL/GLOBAL routing converted findings into "scope the claim" vs "the headline must weaken" — actionable in a way JOINT's thin-joint ranking is not.
Verdict. Premise-node decomposition reaches a distinct, actionable defect class (false/doubtful background premises; scope-repairs) the joint-edge check cannot structurally reach; the two are complementary, not redundant. Build #1 as a sibling to composition-friction-gate with #2's LOCAL/GLOBAL routing built in — do not replace the joint check.
Caveats. Two notes, both theoretical/empirical-premise register (where node/edge divergence is most expected); single reader per cell, no replication. Medium confidence: justifies building, not a universal-lift claim. Untested on definitional and procedural notes.
2026-08-10 — #1 build + generalization test
Built kb/instructions/premise-decomposition-gate.md (experimental, run-by-hand sibling to composition-friction-gate.md; register-aware Step 1/2 for claim / definition / procedure; LOCAL/GLOBAL routing per premise; writes a report, mutates nothing; not wired into the review system).
Generalization test. Ran the gate (fresh cold readers) against a definition (directed-reading.md) and a procedure (critique-note.md) to check the defect class holds beyond title-as-claim/empirical notes.
- Definition: produced boundary-commitment premises and mis-sorting counterexamples — three LOCAL condition-repairs (mid-pass scoping vs "bounded before the pass"; short sources vs the behavioural check; FAQ/API-reference sources vs the "native structure" contrast) and a GLOBAL "this may mark a phase of task-lensed synthesis, not a kind" defeater reached by a sharper route than the note's own "may be redundant" hedge.
- Procedure: produced assumption-premises and misfire counterexamples — LOCAL repairs (register inference, link-scoped engagement judgement, output-destination guard) and three GLOBAL defeaters, incl. P4: critique-note mandates "the strongest case" with an ERROR path covering only inability, not correctness, so it has no sanctioned way to report "the note is sound," driving fabricated opposition on correct notes.
Verdict. Gate generalizes to definition and procedure registers, producing register-appropriate, actionable LOCAL/GLOBAL findings. Confidence: medium → good (three registers). #1 promoted.
Byproduct — fixed 2026-08-10: the P4 finding (a real gap in critique-note.md — no sanctioned "note is sound / no strong critique exists" outcome, forcing fabricated opposition) was fixed by adding a first-class no surviving attack outcome. A verification re-run of the premise gate confirmed P4 RESOLVED but surfaced the inverted risk — a weak critic declaring no-surviving-attack on an unsound note — so the fix was hardened with a low confidence — needs domain check flag that makes that residual visible for the other five lenses to catch rather than silent. Gate calibration stays a recurring problem; visibility, not elimination, is the win.
Next. (1) ~~Decide whether to wire the gate into the review pipeline~~ Resolved (maintainer): the gate does NOT belong in the wired review pipeline — that pipeline is for verdict-producing, easy-to-auto-fix findings. Hard, human-routed, no-verdict methods (composition-friction-gate, critique-note, compression bundle) are set aside from it and orchestrated by kb/instructions/run-full-improvement-pass-on-note.md. Integrate premise-decomposition-gate there as the node-check companion to composition-friction-gate (edges), carried unresolved in "Routed attention", with a GLOBAL-defeated load-bearing premise feeding the note-level Disposition. Integration pending. (2) #2 is mostly delivered at the review layer — the gate emits LOCAL/GLOBAL and the full pass consumes GLOBAL→Disposition. Reconsidered: local/global is a review-layer property, not a discovery-lifecycle phase, so do not graft a test-result branch there; the only remaining option is optional fix-routing-by-scope in FIX-SYSTEM. (3) Optionally write the theory note behind premise-vs-joint decomposition and link the gate to it via rests-on.
Bookkeeping
- Move each candidate to
testingwhen its discriminator runs; record the verdict and the promoted/rejected outcome in this table's row before closing it. - Run cold-reader / adversarial tests in fresh sub-agents (no sympathy for the framing), mirroring the label-ablation setup.
- Validate any promoted artifact with
commonplace-validatebefore it leaves the workshop.