Work

Experimental workshop space. Purpose-driven working artifacts that haven't codified into notes yet.

Each workshop is a directory exploring a specific workflow end-to-end: from question through sourcing and extraction to finished notes. The goal is to discover what patterns actually emerge from use, rather than designing structure upfront.

Active Workshops

  • sniadanie-ai-demo — preparing a five-minute Commonplace demonstration with staged structural and semantic failures, theory-collection methods, and a latency-safe rehearsal plan

  • informed-direction-and-autonomous-judgment — posed 2026-11-08: parked hypothesis that agents assume informed operator acceptance while operators rely on their advice; await local cases distinguishing missed objections, unreported difficulties, and knowingly exploratory work

  • link-label-premise-contribution — posed 2026-11-08: determine when descriptive link text supplies a missing premise; calibrate reconstruction assessment, design controlled comparisons, retain evidence, then write bounded claims and decide the old note's disposition

  • unresolved-work-reconsideration — posed 2026-11-08; reconciled 2026-11-09: single-task and batch selection trials retained; four batch recommendations led to retirement, rewrite, a separate experiment workshop and a declined naming-time check; next decide adoption, revision or another bounded pilot, with broader coverage and recurring-work selection untested

  • karpathy-gist-system-expansion — all 30 original targets integrated; later expansion and comparison refresh reconciled on November 9; Sourcebook acquisition remains failed; decide bounded continuation or closure, preserving historical pilot recurrence evidence

  • hidden-failures-review-fixes — reconciled 2026-11-09: scanner precision, Claude Code audit, review instruction and process-episode retention resolved; parcha analysis integrated; capped-evidence check awaits exact historical query from a Claude Code review, silent-edit recurrence remains watch-only, and ingest leftovers await keep/delete decision

  • curiosity-exploration — second five-turn exploration complete: bounded explanation-seeking and evidence-preservation requirements survive as candidates; no compression-steering advantage established; final findings and measured costs retained

  • bootstrap-exploration — first five-turn exploration complete; second exploration brief on curiosity, compression and retention recorded 2026-11-06 and executed separately in curiosity-exploration; preserves first-trial evidence and costs

  • invariants-under-composition — five rounds complete: concrete probes for preserving requirements through composition and local repair; next proposed test checks dependence carried through a shared cache

  • epistemic-relevance-and-conceptual-blending — observational watch and correction checklist for useful threads being given unsupported conceptual relationships; record the next few cases and counterexamples before refining the hypothesis; research retained as background, experiments deferred

  • operationalization-executability-automation — conceptual investigation of task-relative operationalization, executability, function-level automation and decision autonomy against Commonplace's existing actionability, closure and codification terms; nine sources ingested and turned into worked cases; the task-relative claim extracted into the actionable-methodology definition; the other candidates are covered by existing notes or wait for cases, so the workshop is closure-ready

  • adr-testing — preliminary size measurements and tests of whether agents recover consequential decisions for code and LLM-instruction changes; bounded trace pilot distinguishes late consultation from merely reading after an edit, without assuming ADRs can cover the whole design

  • windows-signed-python-field-use — back burner, waiting for a Windows user: observe the signed-Python install over days of ordinary use under Smart App Control and third-party antivirus (R9 of the closed windows-install workshop); closes with a week clean or a record that reopens the prerequisite

  • review-on-workjig — rebuilding the review and freshness system as Python workflows over Workjig: result files with note and criterion hashes replace the SQLite store, a gate sweep, the full improvement pass and a simplification revision become three workflows under one command, and the interactive parts of the simplification passes become stop-and-rerun points or stay interactive; the stored reviews are not migrated, and phase 0 measures before any code
  • community-scene — finding a first concrete step toward an intellectual community around deep questions about the new computing model; Borretti's failure model and Fabryka AI as the first two cases, an Application-Book analogue built on agentic-system analyses as the untested candidate
  • open-ended-work-economics — what should guide automatic search: expected returns, costs, alternatives and learning from allocation choices under uncertainty; the three-tier model is provisional supporting analysis, not an adopted objective; original investigation delivered library results, current question continues in the same workshop; posed 2026-10-30, resource-tier framing added 2026-11-05
  • conjecture-store — staged migration from a library of corroborated theory to a primary store of cheap refutable conjectures with the theory collection as a projection at a stated threshold; stage 0 adds the store and its type beside theory, skills check its triggers, and nothing consumer-facing changes before onboarding
  • recurring-task-runner — phase 1 writes recurring tasks that can run unattended (explanatory-reach reviewer in batches of ten, workshop dormancy report) and tests them with commonplace-task run; phase 2 builds one cron heartbeat that launches one Workjig run for the due set, committing each task's declared paths separately in the shared checkout; no worktrees or branches, tasks stay obvious, reversible and self-isolating

  • lens-directories — per-lens comparison directories under kb/agentic-systems/comparisons/ replacing Agent Memory Systems in the menu; all four lenses landed with build-time lists, memory with tables and a landscape synthesis; comparison fields for the other lenses remain open

  • agentic-system-completion — resolve the 16 remaining analysis targets through separately authorized evaluation, source/scope decisions or incident disposition; preserves exact predecessor evidence locators

  • use-and-outcome-records — working out how uncontrolled operational observations can guide decisions through compact use records, retained evidence, and summaries as histories grow; storage, retrieval, feedback, and maintenance costs remain open

  • code-simplification — backlog of verified-by-review duplication, test clean-up, and design-level simplifications left after the 2026-09-28 dead-code sweep; one commit per item
  • volunteer-compute-tasks — four contributor choices grounded in committed content: relocation stress search, validator defect-detection search, faster collection validation, and link recognition differential; shared checkout and .venv handoff
  • design-space-studies — exploring whether an explicit design-space study (frame, constraint provenance, axes, typed compatibility rules, decision-critical unknowns) should precede consequential proposals and ADRs; Lane, Ritchey and ATAM ingested; next is a retrospective pilot on Windows command delivery and one prospective pilot; the original LLM exploration is background only
  • decision-lifecycle-evidence — 2026-11-11: operator set provisional second-repository preservation of intermediates for debugging, with churn exceptions to reconsider; workflow-file ledger records current handling, deletion/replacement gaps and remaining coverage; main-record admission and producer changes stay separate
  • unattended-processing-failure-modes — cataloguing observable failure modes of chat-tuned models in unattended Commonplace processing, each admitted by a recorded instance and paired with a system-level countermeasure and a regression check; first entries are premature polishing (repair ordering), format-driven fabrication, and finding deference; the post-training cause stays a tentative frame
  • institution-theory-import — assessing institution theory and adjacent semantic frameworks for natural-language interpretation, conformance, and constraining; separates useful conceptual imports from an unproved institutional formalization
  • first-downstream-run — providing a template for the first experiment on Commonplace in outside use (operator interventions on one KB-building task, an earlier framework release against a revised one) for an outside experimenter to run; not yet handed over as of 2026-10-24, and Release A, the commit the enterprise trials used, is still unrecorded; the fuller scored-run protocol and constructed episodes stay as background
  • ideal-interpreter — exploratory sketch of the LLM modelled as an interpreter of semantics; lacks a functional definition, and nothing in the library may depend on it until one is adopted or the sketch is dropped
  • theory-refinement-interface — reconciling the derive, compare, locate, revise, evaluate, and apply interface with the theory-refinement literature, separating task-level descriptions from implementation guarantees; continues the internal interface investigation after the theory-builder workshop closed
  • operator-led-article-clarification — recording a two-phase method for clarifying draft articles (operator-flagged passages diagnosed by kind, then an agent sweep by pattern) and its working taxonomy of reader difficulties, until enough article runs exist to promote a procedure
  • commonplace-nearest-constructions — characterizing Commonplace through the systems nearest its current human-inclusive theory builder, its path toward less task-specific human input, and its conjectured software-house endpoint
  • cognitive-architecture-transfer-scan — breadth-first conjecture scan whose promising cross-architecture candidates still require deduplication, selective grounding, and extraction
  • collection-aware-full-improvement-pass — determining whether prose-improvement workflows should route by collection and artifact function rather than imposing the theory-note contract
  • dialectical-sample — illustrative dialectical/evidential collection awaiting an explicit keep, archive, or delete disposition after serving its prototype-checking purpose
  • error-catching — systematizing error-detection techniques, their catch rates, activation gaps, and promotion paths into a durable scheme
  • hanoi-experiment — OpenProse bookkeeping stress test with recorded results and remaining variations
  • review-evidence-and-disposition-hardening — resolving or rejecting the code-grounded review findings about result delivery, evidence identity, acknowledgement, recovery, and review-system contracts
  • skill-creator-distillation — comparing Codex and Claude skill creators as mixed-evidence distillation workflows and deciding what should be extracted
  • token-wiki-review — deciding and executing the durable promotion path for the completed Token Wiki analysis
  • full-pass-instruction-coherence-audit — repaired claim-change authority, phase/guard recovery, and closing completion after one instrumented keep exposed a schema-valid final capture with an edit-introduced parsing failure; scenario coverage remains incomplete
  • change-operations-catalogue — a working list of the operations by which Commonplace gets changed, each admitted by an observed instance, with the premises it must read and where they live; used to audit kb/reference/ in both directions under ADR 074. Carries no completeness claim
  • adr-routing — 71 ADRs, no index, and no instruction that routes a self-improvement run into them; the change loop is the only loop without an instruction, so the decisions that bind a change are consulted by luck. Shape is open: index, instruction, subsystem-keyed routing, or none of these
  • explanatory-theories-deployment-time-learning — testing whether explicit system theories improve candidate search, candidate choice, and evidence acquisition in deployment-time learning, including on-the-spot versus retained theories, SPADE-inspired procedure generation, and an Exo-specific compounding track
  • epistemic-architectures — comparing six systems (Commonplace, Eigenius, ScienceFlow, AI Research OS, a private research-ontology draft, and ARC); ARC failed the system-level participation × containment 2×2 unchanged and moved the working comparison to route-specific target, timing, force, and epistemic-versus-operational authority
  • popperian-maintenance-episode — four worked Popperian-maintenance episodes and the 23-run validation series for the machinery they produced (ADR 066). The series verdict: mode conversion is rare because mode landings require mode-appropriate warrant, and counterexample-shaped defeats route scope and category repairs instead. The genre-drift cohort question is closed and refuted; the maxim reconciliation and the episode record's disposition are what remain
  • natural-language-theory-human-agent-contribution — attributing the human and agent interventions behind the natural-language-theory warrant note, with frozen pre-revision evidence and an occurrence/revision-surface/compounding analysis
  • writing-as-thinking-process-transfer — harvesting the still-unincorporated ideas from the four 2026-08-10 writing/thinking essays into Commonplace process changes under a discriminating-test-before-build rule; backlog covers premise-decomposition review, counterexample-scope FIX routing, value-of-information reading, a stance-reversal operator, and a false-precision anti-gate, plus the two promoted conjectures' open tests
  • continual-harness-governance-experiment — designing a matched test of whether Commonplace-derived control, warrant, recovery, and artifact-lifecycle mechanisms improve reset-free harness adaptation without importing a compounding claim or presuming proposal selection
  • dsw-talk — preparing a 35-minute Data Summit Warsaw practitioner talk: one core hybrid-system claim, four engineering lessons, a triage of KB material into included/compressed/cut, and an evolving visual spine that closes on self-hosting
  • agent-runtime-design — deriving what an agent runtime must provide, starting with capability requests, scoped grants, and durable suspension for approval, while separating runtime guarantees from CLI, TUI, API, and host-application control surfaces
  • reflective-improvement-divergence — investigating when open-ended improvement search becomes revision drift, oscillation, moving-criterion instability, or operational non-termination, and what episode-closure rules let a reflective system return to object-level work without pretending it has converged
  • self-revision-design-space — explaining, in plain language first, aspect-bounded builder-loop reach: experimental pathways revise different roles and code while leaving different governing machinery supplied, Commonplace partially retains a human–agent redesign path, and the theoretical Gödel machine internalizes broad redesign behind a proof requirement
  • agent-curiosity-and-structural-coherence — testing whether locally acceptable but globally misplaced code and prose arise from weak role-model activation, anomaly-to-subgoal transition, structural candidate search, or global selection rather than treating “lack of taste” as an explanation
  • execution-channel-compatibility — rebaselined 2026-11-03 against 12 shipped skills, 25 commands and three instruction trees. Delivery/CI acceptance is separate from open procedure portability; next revise the breadth probe for Windows shims, then choose an E1 slice. Aggregate validation must reconcile the CLI's deliberate all rejection; non-promoted/reference coverage remains incomplete.
  • description-length-optimization — determining the shortest description allowance that preserves correct read/skip decisions across plausible hits and large scoped slices, then deriving the validator policy from retrieval quality and an explicit pointer-context budget
  • self-improvement-cluster-operationalization — turning the self-improving-systems cluster into methodology that guides Commonplace's own changes: close the theory's ambiguities first, then audit existing instructions and behavioral-authority artifacts (code, type specs, validators, contracts) against it; also decides the authority-path mix (wired, user-invoked, link-mediated)
  • connect-report-mining — mining the 45 connect reports from the last two weeks plus kb/log.md for recurring synthesis/gap patterns no single connect run surfaces
  • chatbot-goal-state — whether the computational-model notes need a second shape besides the goal-fixed tool loop, for conversations where the human's later messages revise the goal itself rather than supplying tool output; benchmark-construction implications are a downstream thread
  • db-native-reflective-system — what a complete DB-native content/type/link/review schema would need for a hypothetical Commonplace-like reflective self-improving system, starting from the freshness store as a skeleton; does not re-litigate Commonplace's own schema-substrate decision
  • theory-methodology-derivation — the two-layer theory↔methodology pattern (derived fast path, fallback to the generator, promotion by matching) and whether "derivation" should restructure the distillation vocabulary; framings from cognitive architectures, inductive bias, and effective theories in physics
  • freshness-module-review — code review of src/commonplace/freshness/; generic acceptance was withdrawn under ADR 065, while latent CAS/normalization asymmetries and a selector that re-hashes each criterion once per note remain
  • extensible-controlled-vocabularies — designing how code-enforced vocabularies become open-ended per installed KB while staying validator-checkable; source genre now has a severity-warn floor on the authoritative ingest, but its recurring-value/lens extension path and source-tier remain open
  • authority-ranking — testing whether source authority is a linear rank at all (partial order, domain-conditional, time-varying, non-additive under independence) and what consumers a ranking function actually has; framework-side companion to the epistack author-dossier casework
  • lineage-mechanisms — designing one derived-artifact lineage vocabulary across multiple storage weights; review keeps its purpose-built DB while generic lineage state stays deferred until a second churning mesh earns it
  • linking-contract-consistency — reconciling stale linking ADRs, collection-local grammars, asymmetric label directions, reciprocal-link procedures, semantic validation, and inbound-link delivery while leaving derivation lineage to the lineage workshop
  • linking-foundations — grounding authored links in argumentation, inferentialism, explanation and mechanism, discourse/relevance theory, and semantic-memory research; separates assertion, reader function, strength, revision consequence, and derived association before further vocabulary decisions
  • src-architecture-alternatives — alternative architectures for src/commonplace/ from a full code read; active thread is an append-only event log as review-store source of truth with acceptance events that embed their snapshots
  • relocation-move-map-engine — collapsing note and directory relocation around one move-map engine for link rewriting, file moves, redirects, and removal of review-store rekeying
  • kb-graph-loader — testing whether validation, generated indexes, docs hooks, and review targeting should share one loaded KB note/graph model
  • artifact-freshness-and-referential-checks — building artifact-neutral dependency baselines and reverse selection on the shipped review and full-pass version cases; the second consumer decides what comparison/guard code deserves extraction
  • agent-note-improvement — testing instructions that help agents improve weak existing notes by comparing an older weak revision against a later accepted revision
  • agent-memory-design — continuation workshop for discussing revisions and companion artifacts around kb/theory/designing-agent-memory-systems.md
  • bulk-operations — generalizing deep research, review reruns, connect triage, source refresh, validation sweeps, and corpus migrations into a reusable target-selection, sharding, execution, merge-back, and validation pattern
  • runner-execution-profiles — cataloguing model, effort, runner, isolation, and escalation policy across skills and delegated roles, then defining configurable Codex/Claude execution profiles for faster and cheaper workflows
  • pi-agent-zerostack-comparison — preparing a code-grounded comparison instruction for the two Rust coding-agent CLIs cloned under related-systems/
  • vocabulary-governance — deciding how global, collection-local, and type-specific vocabularies should be declared and used by shipped KBs
  • aris-full-trial — running a private full-ARIS paper-production trial while keeping only framing and lessons learned in the public KB
  • review-bundle-packing — measuring and deciding whether review prompts must stay bundle-local or may pack multiple bundles into one run
  • validation — making validation a reliable part of the workflow: when, what, and how to validate (hooks, skill upgrades, periodic revalidation)
  • obsidian-affordances — deciding which Obsidian-facing affordances are useful compatibility layers versus representation drift for a repo-native KB
  • philosophy-borrowing — evaluating Peirce's abduction, Quine's web of belief, speech-act theory, and Carnap's explication as operational borrowings for KB methodology
  • agent-complexity-theory — formal consequences of the bounded-context orchestration model; candidate theorem sketches for academic collaboration
  • latent-space-generation-without-training — exploring whether embedding-guided novelty-generation papers can become a practical no-large-training workflow
  • review-revise-gated — finding review/revise arrangements that reliably produce the manual-edit quality bar, then codifying as reusable instructions
  • auditable-llm-editing — testing whether sparse, anchored writing state prevents accidental claim drift across repeated LLM editing passes
  • lifecycle-management — mapping the full artifact life-cycle (intake, promotion, maturation, retirement); the agent-memory-design test case landed as a note + synthesis trait in kb/theory/designing-agent-memory-systems.md; the proposal disposal half shipped 2026-07-25 as ADR 056 (extract, then archive out of the frontier as a link sink), leaving note and source-snapshot retirement as the blocking gaps
  • condensation-faithfulness-experiment — designing an experiment to test whether our condensation methodology (write conventions + gate suite) produces more behaviorally faithful memory than naive auto-summary, using the Faithful Self-Evolvers perturbation protocol

Complete file listing (generated at build time)