{"path":"research/practices-for-interconnected-research.md","content":"# Practices for sustaining interconnected research\n\n**Researched:** 2026-10-10 by Codex.  \n**Status:** Comparative research; the suggested transfers to Deliberus are proposals, not founder rulings.  \n**Question:** Who else sustains a large, interconnected intellectual programme alongside the tools used to develop it, and what can Deliberus learn from their practice?  \n**The founder's question, in his words (2026-10-10):** *\"Are there others out there on the internet creating as intellectually complex, interconnected projects/research as I am with Deliberus? What tips and tricks and solutions have they found?\"*  \n**Extended:** 2026-10-10 by Claude, at his request: the last two sections before *Where to look first*.\n\nThere are substantial peers. Their relevance is not an alleged ranking of intellectual complexity. It is the combination of long-lived ideas, dependencies across fields, working software, and a need to preserve understanding as people and tools change. Some are individual researchers; others are laboratories or communities. None of the sources below demonstrates that the whole problem of maintaining such a programme has been solved.\n\nThis extends the infrastructure comparison in [science-reasoning-infrastructure.md](science-reasoning-infrastructure.md) into everyday research practice. Discourse Graphs and the mathematical lineage already have homes in that earlier comparison; they are not new discoveries for Deliberus.\n\n## How the evidence was read\n\nThe sources are researchers' papers, their own accounts of working practices, project documentation, demonstrations, and a founder's postmortem. Their evidential roles differ:\n\n- A field study can reveal patterns and failure modes without proving that a tool caused better research.\n- An inspectable prototype establishes that a workflow can be implemented, not that everyone benefits from it.\n- A practitioner's account is evidence of that person's experience, with selection and reporting limitations.\n- A proposal or research vision is a possibility to investigate, not a demonstrated result.\n\nRecent material was checked alongside older accounts. Historical descriptions retain their dates; present-day availability is not inferred from a 2024 announcement. Only public sources are included here. The project-specific proposals are Codex's inferences, separately identified below.\n\n## People and projects worth studying\n\n### Joel Chan and the Discourse Graphs collaborators\n\nThis is the closest substantive comparison: scholarly synthesis organized around questions, claims, evidence, and their relationships, embedded in researchers' existing work. The project's [August 2026 update](https://discoursegraphs.com/blog/dg-news-august-2026) reports continued lab use and development, including better node search.\n\nThe most useful evidence is the [30-month field deployment](https://arxiv.org/html/2407.20666v2). It included 41 survey responses plus interviews and examination of graphs. Users often captured nodes more readily than explicit edges; valuable relationships remained implicit in notebook structure. The study also reports demand for adaptable grammars and workflows. This is observational evidence from a selected community, not a controlled estimate of productivity gains.\n\nA complementary [research account from Protocol Labs](https://research.protocol.ai/blog/2023/discourse-graphs-and-the-future-of-science/) identifies another limit: a graph intelligible to close collaborators can still require substantial explanation for outsiders. Structured representation does not abolish context.\n\n**Transfer proposed:** evaluate whether a captured connection preserves its bearing and context, not merely whether both endpoint notes were found. Let useful structure emerge during real synthesis. Do not replace Deliberus's richer distinctions with a compulsory three-type ontology.\n\n### Andy Matuschak and Michael Nielsen\n\nTheir [tools-for-thought programme](https://numinous.productions/ttft/) combines an ambitious account of thinking and learning with working media experiments. Matuschak's [evergreen notes](https://notes.andymatuschak.org/Evergreen_notes) organize developing ideas around concepts and connections, making writing part of the thinking rather than only its final report.\n\nHis [2023 retrospective](https://andymatuschak.org/2023/) is particularly useful: close observation and bespoke experiments with individual learners exposed problems that earlier large-scale experiments had not clarified. Apparent forgetting could mask inadequate understanding. He explicitly distinguishes this exploratory learning from measuring general effects.\n\nHis [research-practice reflection](https://andymatuschak.org/stillness/) also describes the difficulty of judging progress on ill-defined questions through ordinary productivity measures. This is a personal account, not a validated routine to prescribe to other researchers.\n\n**Transfer proposed:** use a few difficult, fully understood cases to discover what an evaluation must measure. Broad philosophical ambition is compatible with very particular intermediate investigations. A clearer unresolved distinction can be real progress even when it produces no new feature.\n\n### Gwern\n\nGwern's [account of his website](https://gwern.net/about) describes a maintained corpus rather than a succession of disposable posts. Of special value is the separation between a page's completion status and the author's confidence in it. A coherent, finished essay can still be uncertain; rough notes can contain strong evidence. The account also documents writing and maintenance practices rather than treating publication as the end of a text's life.\n\n**Transfer proposed:** preserve the distinction between polished prose, evidential support, and endorsed positions when material travels into retrieval results or agent context. Deliberus already distinguishes asserted and ruled positions; the opportunity is to make that distinction survive every derived view, not to invent another competing taxonomy.\n\n### Jerry Michalski and Open Global Mind\n\nMichalski's [public Brain](https://www.jerrysbrain.com/) is a long-running example of a large personal network connecting ideas across domains. Its reported size establishes sustained accumulation, not the truth or usefulness of every connection.\n\nAn [Open Global Mind discussion from 2021](https://wiki.openglobalmind.com/conversations/conversations_and_knowledge,_fast_and_slow) supplies the more instructive failure account. Participants describe a gap between rich conversations and accessible shared knowledge. They discuss purpose-specific documents and maps developed during conversations, rather than relying exclusively on later harvesting. These are participant observations and proposals, not a completed evaluation of an intervention.\n\n**Transfer proposed:** judge a research session partly by what changed in a standing question, explanation, or decision record. Transcripts preserve evidence; maintaining a usable intellectual object requires additional work. Automation can assist that work without making participants' intended meaning self-evident.\n\n### Geoffrey Litt and Ink & Switch\n\nLitt's [July 2026 essay on understanding](https://www.geoffreylitt.com/2026/07/02/understanding-is-the-new-bottleneck) describes explainers, interactive environments, and other instruments for understanding AI-produced work. His [2025 essay on AI HUDs](https://www.geoffreylitt.com/2025/07/27/enough-ai-copilots-we-need-ai-huds) argues for information presented where the work happens. Neither establishes that interactive displays always outperform agents.\n\nInk & Switch's [Patchwork experiments](https://www.inkandswitch.com/patchwork/notebook/2024-version-control/) make alternative edits and AI contributions reviewable through branches, diffs, and history. The researchers report using these prototypes in their own writing. Their experiments also make the prompts behind editing bots ordinary versioned documents.\n\n**Transfer proposed:** show a proposed intellectual change together with its sources, consequences, and alternatives. An interactive example can make a contested distinction easier to inspect. Group related edits by the question they answer, even when they affect several files. Existing Git history can provide the underlying record.\n\n### Simon Willison\n\nWillison's [practice of retaining working examples](https://simonwillison.net/guides/agentic-engineering-patterns/hoard-things-you-know-how-to-do/) gives agents concrete, previously demonstrated ingredients to combine. His reports and small tools preserve implementation knowledge in reusable form. This is especially relevant to the experimental and engineering parts of an intellectual programme; it is not evidence that his subject matter duplicates Deliberus's philosophy.\n\nHis [interactive-explanation example](https://simonwillison.net/guides/agentic-engineering-patterns/interactive-explanations/) illustrates a related technique: when prose failed to convey an algorithm, an animated explanation made its operation inspectable.\n\n**Transfer proposed:** retain executable cases alongside conclusions about a model or method. A benchmark result should point to its inputs, context, version, and interpretation. Include known failures, so later improvements can be tested against the actual problem rather than against a remembered description of it.\n\n### Andrej Karpathy's LLM Wiki\n\nKarpathy's [April 2026 recipe](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) separates raw sources from an LLM-maintained Markdown wiki and conventions for ingesting, querying, and checking it. He describes personal involvement in ingestion and saving useful answers back into the wiki. His reported scale is moderate; this is a practitioner workflow, not a comparative benchmark.\n\n**Transfer proposed:** retain useful synthesis between sessions, tied to the sources from which it was constructed. Deliberus already does much of this. The unresolved issue is correction: a fluent derived page must not gain the standing of a ruling, and an upstream change must make affected synthesis discoverable. This recipe supplies no basis for assuming that search is unnecessary at Deliberus's scale, or that an LLM will maintain every dependency correctly.\n\n### Terence Tao, Patrick Massot, and the Lean blueprint ecosystem\n\nTao's [account of the PFR formalization](https://terrytao.wordpress.com/2023/11/18/formalizing-the-proof-of-pfr-in-lean4-using-blueprint-a-short-tour/) describes a human-readable proof blueprint linked to formal declarations. Its dependency graph shows which components are formalized and which are ready for work. This provides a concrete unit of collaboration inside a much larger argument.\n\nThe [2026 LeanArchitect paper](https://arxiv.org/html/2601.22554v1) addresses drift between informal blueprints and formal code by extracting dependencies and proof status from annotated Lean declarations. Its case study also reports that generic instructions to refine a failed blueprint did not reliably resolve the underlying problems. Unformalized nodes still require manual management.\n\n**Transfer proposed:** derive views from their authoritative records where possible, and surface missing dependencies or stale assertions. The transferable property is traceability. A proof checker verifies formal statements under formal assumptions; it does not certify the adequacy of a philosophical interpretation or its connection to lived experience.\n\n### Metagov and BlockScience's KOI work\n\nTheir [KOI project description](https://metagov.org/projects/koi-pond/project-details) treats organizational knowledge as a network whose objects can change and whose uses have local governance. The [FAQ](https://metagov.org/projects/koi-pond/faq) distinguishes references from the objects they identify, describes multiple granularities and perspectives, and separates deployed workflows from planned ones. It includes an ethnographic workflow connecting community material to researchers' contextual analysis.\n\n**Transfer proposed:** preserve where an item came from, what it refers to, and which context gives it its standing. Treat updates and withdrawals as relevant to downstream uses. Shared identifiers can connect different views without requiring that they all share one interpretation. These are architectural lessons, not a recommendation to adopt KOI or Neo4j for the personal corpus.\n\n### Bret Victor and Dynamicland\n\nVictor's [2024 introduction](https://dynamicland.org/2024/Intro/) describes a long-running research programme in which computation, learning, community, and the research medium develop together. Its demonstrations place data and working models in a shared physical setting so people can inspect and modify what they are discussing.\n\n**Transfer proposed:** for a difficult Deliberus question, consider whether participants can manipulate the same concrete example and see its consequences. The relevant lesson is shared inspectability, not a requirement for physical computing hardware. The demonstrations do not establish that this interaction model resolves philosophical disagreement.\n\n### Arbital: a close historical comparison that did not succeed\n\nArbital combined interconnected explanations, personalized paths, and discussion with unusually broad intellectual ambitions. In his [2018 postmortem](https://www.lesswrong.com/posts/kAgJJa3HLSZxsuSrf/arbital-postmortem), founder Alexei Andreev describes difficulty communicating how the parts fit together, demand for reading that did not translate into writing, and features whose usefulness remained untested. He proposes detailed interface walkthroughs as one way to make an integrated vision assessable. This is a retrospective from one participant.\n\nThe [2025 migration announcement](https://www.lesswrong.com/posts/fwSnz5oNnq8HxQjTL/arbital-now-lives-on-lesswrong) documents preservation of its content and some features on LessWrong; the original project had been discontinued.\n\n**Transfer proposed:** make the intended intellectual experience concrete enough to examine. This does not establish that Deliberus should abandon its review priorities or that all ambitious interconnected systems share Arbital's fate. The informative question is which dependency can be inspected now, not how quickly every part can be launched.\n\n## What follows for Deliberus — proposals, not decisions\n\nThe following synthesis combines those accounts with the project's existing distinctions. It is not a claim that the cited researchers have validated these exact designs.\n\n### 1. Make the current question the organizing unit\n\nA useful working view could bring together the question, existing positions, unresolved distinctions, sources, concrete examples, and the implications of each alternative. This should be a view over existing material. Creating a second manually maintained authority would reproduce the synchronization problem.\n\nFor the current reviews, this means presenting one claim as it passes from assertion through ruling, compression, and resident context. It could also show the constellation wording and code affected by it. This supports the existing ordering of work rather than replacing it with a generic product-development sequence.\n\n### 2. Capture why a connection matters\n\nFinding related text is a discovery step. A useful proposed connection also needs its endpoints, the particular bearing between them, and enough context to challenge that reading. Sometimes the context is a surrounding conversation; sometimes it is an unstated assumption or a contrast with a third position. A typed edge alone can still discard the crucial information.\n\nThis fits the distinction in [what-a-rule-rests-on.md](what-a-rule-rests-on.md) between having relevant material available and taking it into account. The appropriate test is whether the connection helps the next act of reasoning, including when it should be rejected.\n\n### 3. Separate finding, applying, and maintaining knowledge\n\nThese are three different failure points:\n\n| Operation | Example of failure | Evidence that would address it |\n|---|---|---|\n| Find | A relevant earlier position never reaches the agent | A retrieval or delivery test with identified sources |\n| Apply | The position is present, but the new proposal ignores its bearing | A contextual case testing the proposed action against the position |\n| Maintain | A correction reaches one document but an older derived form remains in use | A trace from the changed source to affected views, with checked updates |\n\nBetter retrieval can help the first without resolving the others. A hook can ensure a check runs without making its interpretation correct. A visible reviewer can support human judgment without proving that later agent behavior will follow it. The [agent-memory research note](what-agent-memory-research-says-about-agent-surfaces.md) supplies related questions; this comparison does not independently validate the benchmark numbers quoted there.\n\n### 4. Use concrete failures to design evaluation\n\nA historical review set can reveal distinct failure classes: explicit connections, context-dependent short replies, plausible but unwanted associations, and changes in a position's standing. Human labels should preserve disagreements and uncertainties. Once examples have shaped a prompt, they cannot also provide an untouched estimate of its performance.\n\nBeyond judging whether a reminder was warranted, an application test could present the earlier position and a proposed next step, then ask what that position permits, rules out, or leaves undecided. This is a proposed extension, not a reason to discard the simpler detection benchmark.\n\n### 5. Preserve standing while allowing useful derived views\n\nOriginal sources, endorsed positions, working hypotheses, and generated syntheses need distinguishable roles. Source links and version information make a derived view inspectable; they do not make its interpretation faithful automatically. A detected source change can flag dependent passages for reconsideration without pretending to know whether their meaning is now wrong.\n\nThis is compatible with [store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md): preserving the record and computing a reading over it is different from repeatedly overwriting the record with a new summary. The comparison does not settle any unruled design choices about that distinction.\n\n### 6. Add an interface when it enables a missing operation\n\nAn additional surface is useful if it lets the user inspect source context, correct an interpretation, compare alternatives, or explore a concrete model more effectively. Another place to converse is not by itself evidence of improvement. Existing files, search, hooks, and a small reviewer may already supply the needed operations.\n\nThere is no conclusion here that a particular Claude feature must be adopted. A new interface should be judged by the intellectual action it enables and by evidence from actual use. The programme's breadth does not itself require a more elaborate software framework.\n\n## Read in more depth, and set beside the corpus (2026-10-10, Claude)\n\nThe founder's request, the same day: *\"Research each of the strongest matches more in-depth for learnings.\nRelate their projects, ontologies, processes, UX and human insights to what is and has been going on in\nDeliberus, and make sure you've read our corpus deeply on each and every aspect, and think about unexpected\nconnections and analogies. Also for each of the five \"practical lessons\" you name, can you do in-depth work\nto prepare for me and Claude to act fruitfully on each of them?\"*\n\nTwo helpers read the primary sources and reported how deeply they read each\n([structure peers](practices-for-interconnected-research/structure-peers-as-a-helper-read-them.md),\n[practice peers](practices-for-interconnected-research/practice-peers-as-a-helper-read-them.md); kept word\nfor word, unchecked by Claude). Two of the eight matches were outside the helpers' reading: Claude read\nDynamicland's 2024 introduction in full, and of Michalski's Brain only its front page. The right-hand\ncolumn is Claude's reading and is unruled.\n\n| What the deeper read found | Where it meets this project |\n|---|---|\n| **Discourse Graphs: people recorded items and seldom relations.** Four exports averaged 1,400 nodes in a year with few explicit relations; users indented their notes to show a relation. No machine was in the loop, and recognising relations by machine is listed as future work. | This project has the opposite supply. Its linking step writes relations by machine in volume, and on 2026-10-07 four ways of having a machine judge a link failed. Each project holds what the other lacks, and writing out what a link rests on sits between them ([writing-out-the-unwritten.md](writing-out-the-unwritten.md)). |\n| A relation there is stored once and shown from both ends. Many *evidence* nodes were empty placeholders. In one lab an explicit *claimed by* field was filled on 69 of 130 claimed issues, and the rest had to be read from logs. | An empty placeholder is an accident in their study and an invitation here (the open slot that asks for a contribution). Reading who did what from the record is [store-the-uses-read-the-meaning.md](store-the-uses-read-the-meaning.md), and the ruling that assent is read from what people do. |\n| Sharing a graph worked mainly between people with much shared context, often in a live meeting. | The stranger test, and the choice to try this project first with two people side by side. |\n| **Lean blueprint: status is computed from dependencies and never typed**, and one record is the authority from which the other view is generated. | *A horizon kind is read, never set*, and the generated frontier page. It also names a standing weakness here: a count typed into always-loaded text decays, and two such counts in the instruction file have each gone stale twice. |\n| Change an upstream value and the checker lists what breaks (one changed constant broke about a dozen lines). | A correction here has no checker. Two partial ones exist: the ledger's row ids are stable names, so every file that names a row can be listed, and a sweep for the old wording was built this day. Neither finds a passage that depends on a ruling without naming it. A note written the same day reaches the same point from the Mathlib library: a library can overwrite because a change that matters breaks a build, and here nothing breaks ([alien-proofs-and-the-library.md](alien-proofs-and-the-library.md)). |\n| A machine-written proof that did not depend on the lemmas around it showed that its statement was wrong. Telling a model to *refine* a failed attempt did not help while the prover reported only which nodes were unsolved. | The collision of 7 October ([what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 14.6). Telling a session to read more did not help. What helped was a reader that named the settled item and quoted the words in conflict, which is a diagnosis and not a request to try again. |\n| **KOI: the reference is kept apart from the thing referred to.** A deleted item keeps its identifier and the deletion is logged. With no decision policy, those who show initiative shape the direction. | Contributions are added and never overwritten, a superseded link stays visible, and the constitution this project has not yet written. |\n| **Matuschak: a learner can answer when tested and still fail to use what was learned**, without recognising the occasion for it. He learned more in weeks beside single users than in a year of large experiments, and now tests in numbers only after a design works for one person. | The same failure was seen here in an assistant: the ruling's note was read, cited, and not applied. So it belongs to knowledge held as text to recall, and is no defect peculiar to machines. His order of work is the principle blessed here on 2026-10-06: a person reads what a pass produced before anyone proves how often it is wrong. |\n| Shared wikis decay as notes stop being atomic, authors cut concepts differently and terms drift. | The concept layer and the glossary pass. |\n| **Litt: an explainer for each change, ending in five questions he must pass before he hands the work on**, since reading is easy to fake. Ink & Switch: a machine's edits arrive on a branch with a reason for each, and its instructions are a versioned document. | [understanding-after-the-proof-machines.md](understanding-after-the-proof-machines.md) and the listening pieces. One use not tried here: the questions put to the *session*, about the settled item, before it proposes. Edits on a branch with reasons is *propose, never assert*. |\n| **Karpathy: sources kept unchanged, a model-kept wiki, an index read first, a log, and a periodic check for contradictions and stale claims.** He reports the index alone working to about a hundred sources. Not addressed: a source that changes, or a page read and not applied. | This corpus is past that size, at more than three hundred notes, and its index is half of the instruction file. It has what the recipe lacks: rulings in the founder's words, each with a state. |\n| Willison: keep runnable examples at known paths, and tell an assistant to fetch raw text, since the default fetch summarizes. Gwern: status, confidence and date as separate fields on every page. | Helper reports kept word for word; *a value shaped for display must not be used for judging*; and the habit here of stating how far built a thing is apart from how well it is recorded. |\n| Open Global Mind: harvesting afterwards cost two hours for each recorded hour and lost nuance, so they proposed mapping during the conversation. Arbital: only its designer used the relation feature, so it went untested, and it had about a dozen regular writers. | The live-session design, and capturing a connection in the turn it is made. Arbital's account is this project's own caution: no instrument here has yet met a user with a stake. |\n| **Dynamicland: people in one room, with the data and the model on the table between them**, each able to see and change what the others made. Bret Victor's introduction calls it \"Not a product, but a practice\", learned by being among people who do it. The system is small because the room and its social habits do much of the work. | The two-person rehearsal at one screen, and the finding that the feeling channel does not survive delay ([convictions-across-scales.md](convictions-across-scales.md)). Two differences matter. There anyone may change anything, and the room keeps that honest; here people are apart, so a contribution is added and nothing is overwritten. And they remove the software developer where this project removes the need to know its categories. The fourth condition of the success prior (the code maps a practice people already have) is their whole method. They also build their system inside itself; here the like is held by the founder's ruling, and nothing is proposed. |\n| **Michalski's Brain: one person, 28 years, more than 645,000 linked items**, by its own front page. Read no further than that page. | With the Open Global Mind account above (the same group asking who could use its records), it marks the limit of a register kept by one hand: legible to its keeper. The register of the founder's connections keeps his sentence for each line for that reason, since the sentence is what a stranger can read. |\n\n## The five lessons, prepared (2026-10-10, Claude; proposals, unruled)\n\nEach is set out the same way: what exists, what the reading adds, the first step, what is the founder's to\ndecide, and how we would know.\n\n*Since the same day, part of lessons 2, 3 and 5 is built as the corpus graph\n([corpus-graph/README.md](corpus-graph/README.md)): `corpus_graph.py around Q33` prints one ledger row with\nevery file that names it, its home notes and the open work on it (a first version of lesson 2's page,\nwithout his words); `add` keeps a connection with its reason (lesson 3, for connections between notes);\nand a note's corrections written elsewhere are handed to a session when it opens the note (lesson 5).\nFor lesson 1 it gave a first count of the failure to apply: three readers sent to 27 recent notes found 42\nplaces where a note states what an earlier ruling, measurement or design already answers or contradicts\n(corpus-graph/README.md § 9).*\n\n**1. Finding, applying and maintaining are three different failures.**\n*Exists.* Finding: corpus search, with its embeddings restored on 9 October and a line that says when a\nsearch is partial. Applying: the worked example of the 7 October collision, and one trial of a reader\noutside the draft, which named the conflict when given the settled items in full and missed it when given\ntheir headlines. Maintaining: the rule that a correction is fixed on every surface that repeats it, and a\nsweep command for it.\n*Adds.* Matuschak's learners and the Lean case say the same thing from two sides: recall is not use, and\n*try again* without a diagnosis does not help.\n*First step.* A test set for applying, drawn from our own record: the nine places where a proposal of\n7 October met something ruled or built ([what-a-rule-rests-on.md](what-a-rule-rests-on.md) § 14.2), with\nproposals the founder blessed as controls, put unlabelled to the outside reader. Not yet run.\n*His to decide.* Whether that reader becomes a standing step before a proposal reaches him, and what counts\nas settled for it.\n*How we would know.* How many of the nine it names, and how many blessed proposals it wrongly flags.\n\n**2. Make the current question the working surface.**\n*Exists.* The four reviews treat one position as it passes from asserted to ruled to compressed to resident,\nbut its parts sit in four files.\n*Adds.* The blueprint's page for one node (statement, status, what it depends on, where it is discussed)\nand the Discourse Graphs panel that shows a node's relations from both ends.\n*First step.* A page generated for one ledger row, with no model: the row and his words, every file that\nnames the row, the section of its home note, the constellation sentence and canon line if they exist, and\nthe open rows that name it. Row ids make this mechanical. It is a view over the files and stores nothing.\n*His to decide.* Which review item to try it on. It would not live inside the Deliberus graph: storing the\nproject's own beliefs there is held by his ruling.\n*How we would know.* Whether a ruling sitting goes faster and needs fewer re-readings with the page.\n\n**3. Capture why two things connect.**\n*Exists.* Since 9 October, a register with one line for each connection the founder draws (the two things\nand his sentence), a reminder that fires on his message, and a step at the end of a session.\n*Adds.* Discourse Graphs measured that relations go unrecorded unless recording one is the cheapest act\navailable, and Open Global Mind that they are lost if harvested afterwards.\n*First step.* His review of 96 of his own past messages, which asks two things apart: is this a connection,\nand would he want it kept. The reminder is then re-measured on his labels.\n*His to decide.* What counts. A link between two things already under discussion, or a tool proposed for a\nrole in one word, is missed by the reminder as it is defined today.\n*How we would know.* Catches and false alarms on the 32 messages held back from tuning.\n\n**4. Keep examples as reusable assets.**\n*Exists.* Every test of the decision model keeps its cases, labels, script and answers; the review set keeps\na manifest and a fingerprint; since 10 October each logged decision names the rule that made it.\n*Adds.* Willison's known paths, and Gwern's separation of how finished a thing is from how sure.\n*First step.* One index of the test sets: what each tests, who labelled it, and whether it has been tuned\non, so that a set used to shape a rule is never also quoted as its score.\n*His to decide.* Nothing.\n*How we would know.* A later session reuses a set where it would have rebuilt one.\n\n**5. Make corrections travel.**\n*Exists.* The rule, the ⚠ mark kept in place beside the old text, the `correction:` commit prefix, and the\nsweep command built this day. On its first use, on a statement corrected that day, three notes still\ncarried the old wording.\n*Adds.* The blueprint's checker, which lists what a change breaks, and KOI's notices of *new*, *updated*\nand *forgotten*.\n*First step.* Use the sweep at every correction, by wording and by ledger row id.\n*His to decide.* Whether a `correction:` commit must list the surfaces it checked.\n*How we would know.* The count of surfaces found late, which today is only noticed by accident.\n\n## Where to look first\n\nFor the closest research-method and representation peers, start with [Discourse Graphs](https://discoursegraphs.com/). For knowledge governance and interoperation across communities, examine [Metagov](https://metagov.org/). For long-term personal and collective sensemaking, explore Michalski's public map and the Open Global Mind archive. For instruments that make difficult ideas inspectable, read Litt, Ink & Switch, and Dynamicland. These are suggested places to learn from; no outreach was made.\n\nThe most useful short reading sequence is the Discourse Graphs field study, Matuschak's 2023 retrospective, Litt's July 2026 essay, and the LeanArchitect limitations and case study. Together they expose four different problems: relationships left implicit, misunderstood tasks, insufficient human understanding, and drift between representations.\n\n## Related Deliberus work\n\n- [Science reasoning infrastructure](science-reasoning-infrastructure.md): scholarly systems and historical lineage.\n- [What agent-memory research says about agent surfaces](what-agent-memory-research-says-about-agent-surfaces.md): delivery, provenance, contradiction, and the limits of existing evidence.\n- [What a rule rests on](what-a-rule-rests-on.md): context, warrants, and this project's instance of relevant material read but not applied.\n- [The canon distillation pass](the-canon-distillation-pass.md): the existing programme for deciding what must remain resident.\n- [Store the uses, read the meaning](store-the-uses-read-the-meaning.md): preserving records and computing interpretations.\n"}