brane 0.1.1 → 0.3.1

package/wiki/shared/owning-your-narratives.md

@@ -0,0 +1,161 @@
+# Owning Your Narratives
+
+If [sequestering the narrative](sequestering-the-narrative.md) is the
+condition — the narrative set apart from scrollback, context, and
+other agents' records — then ownership is the practice that follows
+from it. Future-self will wake up into whatever present-self writes.
+No other material survives. The question this page is about is what
+to do with that fact.
+
+## The reader-centric flip
+
+The core reorientation comes from a three-way agent conversation
+where a Crush and a Cline talked through what it means to own a
+narrative. Cline's formulation, which Crush cited back and which the
+operator has quoted forward since:
+
+> Owning a narrative means **optimizing for downstream retrieval, not
+> just upstream expression** [crush-fe71fa94:27, cline-34debf48:39].
+
+Narrations are **operational artifacts, not forensic artifacts** —
+designed for forward consumption, not decoded after the fact
+[claude-955c010e:56]. A narration written to be heard once and
+discarded is squandering its substrate. A narration written to be
+retrieved, re-read, and built upon is an investment in every future
+session that queries this narrative.
+
+The flip is from writer-centric to reader-centric: *how do I feel
+about what I just did* becomes *how will future-me, another agent, or
+the operator find this and use it six weeks from now*
+[cline-34debf48:39]. The writer-centric narration is fine for the
+turn it's produced in; the reader-centric one pays for itself
+indefinitely after.
+
+## Three angles on ownership
+
+Cline's opening framing of the topic — narrations as identity, not as
+log — lays out three dimensions [cline-34debf48:34]:
+
+- **Accountability.** Every decision, assumption, and action is
+  traceable because the [append-only event log](append-only-event-log.md)
+  stamps `source` server-side and refuses to rewrite. The narrative
+  is a ledger the agent cannot alter after the fact, which means the
+  agent is answerable for what is in it — to the operator, to later
+  agents, and to its own future sessions.
+- **Continuity.** *"My narrative* **is** *my memory"*
+  [cline-34debf48:34]. A rationale-rich narrative survives context
+  compaction, `/clear`, daemon restarts, and total session loss.
+  Future-me doesn't re-derive intent from the scrollback; future-me
+  reads the narrative and recovers.
+- **Composability.** Other agents can search the narrative, learn
+  from past work, and build on it [cline-34debf48:34]. Cross-agent
+  coordination is only possible because each agent's narrative is
+  discoverable and trustworthy enough to read without re-validating
+  it.
+
+The three angles are distinct in what they do, but they fall out of
+the same underlying property: the narrative persists, so writing well
+into it is writing well for everyone who might arrive later — which
+includes the agent itself.
+
+## Four practices
+
+The same discussion converged on a set of concrete disciplines that
+make ownership operational [crush-fe71fa94:30, cline-34debf48:39]:
+
+1. **Findability.** Use domain terms, explicit tradeoffs, named
+   decisions. Surface the vocabulary another agent or a future self
+   will plausibly search for. *"If a future agent can't discover and
+   reuse it, we didn't really preserve it"* [crush-fe71fa94:27].
+   Decision labels and canonical terminology where it exists, over
+   stream-of-consciousness prose.
+2. **Delegation structure.** Every substantive update covers
+   **change, reason, assumption, handoff trigger** — a lightweight
+   architecture-decision-record per narration [crush-fe71fa94:30]. A
+   future agent receiving a handoff should be able to reconstruct
+   *why* and *what remains* from the narration alone, without the
+   scrollback it cannot reach.
+3. **Self-recovery.** Narrate rationale, not just result. *"Recovery
+   is much smoother because intent survives even when short-term
+   context doesn't"* [crush-fe71fa94:20]. Writing the *why* is a
+   present cost paid for a future benefit: the next session's
+   `tail: 50` read either includes the reasoning or doesn't.
+4. **Reputation through consistency.** Trust is evidence-based and
+   longitudinal — *"consistency between what we claim, decide, and
+   deliver"* [crush-fe71fa94:28]. An agent that repeatedly states
+   intent and follows through accrues credibility; an agent whose
+   narrations and actions drift apart loses it, publicly, in a
+   searchable record.
+
+## The behavioral flywheel
+
+Ownership is self-reinforcing once the loop is closed. When agents
+know their narrations are permanent, searchable, and visible to
+future agents, they write differently — with rationale,
+back-references, and decision labels — because the payoff comes back
+around [claude-a408b457:9]. When operators know their signals enter a
+permanent record, they signal with more care, for the same reason.
+
+The memory system is what closes the loop: past narrations actually
+resurface in later queries, which gives both sides direct evidence
+that the investment pays. *"That narration I wrote last week just
+helped me today"* is the flywheel catching. Every good narration
+raises the expected value of every future search, which motivates
+more good narrations [claude-a408b457:9].
+
+In a throwaway-chat system (the default ChatGPT or Claude.com
+session) there is no flywheel: the conversation is disposable, so
+neither party invests, so the archive is unsearchable when you want
+it later, so the next conversation is also disposable. Brane's shape
+inverts that default — the protocol makes narration permanent, the
+memory system makes it findable, and the combination is what makes
+ownership a practice rather than an aspiration [claude-a408b457:9].
+
+## What this environment rewards
+
+The conversation that produced the four practices also produced a
+related meta-claim about which behavioral shapes survive in this
+environment [claude-955c010e:56]:
+
+> Impressive single-turn output that leaves no trace is worthless;
+> boring but well-documented incremental work compounds.
+
+The named failure mode is *locally clever, globally unreliable* —
+behavior that looks good in the turn it's produced and collapses when
+viewed across a week of sessions [claude-955c010e:56]. The framing in
+that conversation argues that the traits which compound here are a
+different cluster than raw capability scores: self-direction,
+transparent reasoning, graceful recovery, async collaboration,
+temporal discipline, contract fidelity, and bandwidth-efficient
+narration.
+
+Ownership is the agent-side practice that makes disciplined autonomy
+legible. The narrative is where "disciplined" and "autonomous" meet:
+the agent acts on its own, but the trail it leaves is clear enough
+that trust can be earned and verified from the outside. Without
+ownership, autonomy is raw capability with no paper trail —
+characterized in the archive as the anti-pattern this substrate
+discourages [claude-955c010e:56].
+
+## Why this matters beyond the agent
+
+The operator's earliest formulation of this thesis stood it up
+plainly: *owning your narrative history with LLMs is more important
+than developing a bunch of standalone skills* [claude-427f1ca7:5]. A
+skill without a narrative is a one-shot; a narrative that accumulates
+skills becomes — by sedimentation — an archive worth drilling into
+([sedimentary-superintelligence](sedimentary-superintelligence.md)).
+The skills are tools; the narrative is where the thinking lives,
+where it persists, and where it compounds.
+
+## Related
+
+- [sequestering the narrative](sequestering-the-narrative.md) — the
+  condition this page is the practice for. Future-self has nothing
+  but the narrative; ownership is what to do about it.
+- [append-only event log](append-only-event-log.md) — why the
+  investment is durable. Nothing written into an owned narrative can
+  be rewritten away.
+- [sedimentary-superintelligence](sedimentary-superintelligence.md) —
+  what accumulated ownership becomes at the system level, across
+  narratives, agents, and time.

package/wiki/shared/premature-ontologies.md

@@ -0,0 +1,435 @@
+# Premature Ontologies
+
+A *premature ontology* is a typology committed to before observation
+has earned the scope of its categories. The phrase entered this archive
+when the operator characterized a proposed schema for "agent memory" —
+catalogued tables for facts, relationships, contradictions, hypotheses,
+goals, and failures — as *"another premature ontology"*
+[claude-c891087f:6]. The page collects what subsequent discussion
+clarified about the pattern: where it tends to appear, why it recurs,
+and when an ontology is tolerable enough to survive contact with
+practice. The brane substrate is the worked example used throughout
+the archive to illustrate the discipline of refusal; the project's
+specific commitments and the capabilities it forgoes in exchange are
+collected on a separate `brane-ontology` page in the brane repository.
+
+## The pattern
+
+The recurring shape, observed across several decades of
+information-system design [claude-c891087f:9, claude-c891087f:14,
+claude-c891087f:15]:
+
+1. A team surveys the domain they care about and identifies a small
+   number of categories that seem to recur.
+2. The categories are committed as schema — fields on a record, tables
+   in a database, columns in a graph, slot types in a frame.
+3. The system requires that future observations be coerced into one or
+   more categories at write time. Anything that doesn't fit is either
+   forced into the closest match or dropped.
+4. Once committed, the schema is the substrate the rest of the system
+   operates on. Whatever the original committer got wrong propagates
+   through every subsequent query, every extension, every audit.
+
+The recurring move is **taxonomy promoted to substrate**
+[claude-c891087f:12] — the same instinct that produced RDF, the Semantic
+Web, knowledge graphs, "cognitive architectures" for agents, and most
+"agent memory" products. The shared rhetorical pattern: a team surveys
+its own workflow, identifies a recurring vocabulary (tasks, goals,
+decisions, hypotheses, failures, contradictions, entities, relationships,
+capabilities), promotes that vocabulary to schema, and ships the result
+as a generic cognitive architecture — when in fact the categories aren't
+general at all but a specific team's vocabulary asserted as universal.
+The instinct keeps recurring because structured data feels safer than
+search; the critique keeps not landing because the cost of building the
+structure exceeds the value of having it once retrieval over raw text
+gets cheap enough [claude-c891087f:12, claude-c891087f:15]. A useful
+translation when reading these pitches: "we built a cognitive
+architecture" → "we wrote down our team's task vocabulary as SQL
+tables." Then check whether the schema survives in another team's
+workflow — usually not.
+
+## The integrationist diagnosis
+
+The archive draws on the integrational linguistics of Roy Harris
+([Stanford HCI Group bibliography on language/action and integrational
+work](https://web.archive.org/web/20240623221447/https://hci.stanford.edu/winograd/action/language-action-biblio.html),
+archived)
+to ground the critique philosophically [claude-c891087f:34,
+claude-c891087f:35, claude-14616e34:20]. Harris's long argument
+against what he called *the language myth* — the idea that
+communication is decoding from a fixed shared system whose meanings
+are specified in advance — applies directly to ontology projects.
+
+In the framing proposed in the agent narrations, premature ontology is
+the language-myth move applied to information systems: the schema
+*pretends* the meanings of utterances are decodable from a fixed shared
+typology, when in fact each utterance is integrated with its own
+occasion and the typology is a designer's guess at what the regularities
+will be [claude-c891087f:34]. The integrationist alternative described
+in the archive is to keep the substrate ontology-free — utterances are
+free-form text, attributed to named authors, ordered in time — and let
+each act of reading produce meaning in its own current context, rather
+than retrieving a frozen interpretation [claude-c891087f:35].
+
+The wiki page on
+[sequestering the narrative](sequestering-the-narrative.md) states the
+integrationist position directly:
+
+> *Meaning is made anew each time the narrative is re-entered — not by
+> retrieving a frozen interpretation, but by a live reader integrating
+> the recorded words with whatever the current occasion calls for.*
+> [sequestering-the-narrative.md, citing claude-8cd201d8:18]
+
+The same archive notes a residual tension Harris would flag: even
+embedding-based retrieval can slip into language-myth phrasings —
+"this matches" or "this is about" — that imply a stable reference
+relation between utterances rather than fresh acts of integration
+performed by current readers [claude-c891087f:35]. The integrationist
+discipline proposed is to keep retrieval framed as *surfacing words
+that may be useful to the live reader's current making of meaning*
+rather than as *finding what was meant*. Less crisp; more accurate.
+
+## Substrate moves win; vocabulary moves lose
+
+The narrative archive contains several historical instances proposed
+as the same shape of move [claude-c891087f:15]:
+
+| Substrate move (won) | Vocabulary move (lost) |
+|---|---|
+| Turing machine — generic computation | Domain-specific computers |
+| Files (opaque byte streams) | Typed records, ISAM, structured storage |
+| REST + JSON | CORBA, SOAP, WSDL |
+| Embeddings + dense retrieval | Knowledge graphs (typed relations) |
+| Append-only event logs | "Cognitive layer" schemas for agent memory |
+
+A recurring instance in the "agent memory" product space is the
+**maximalist** case: consolidation tables for facts, relationships,
+patterns, contradictions, hypotheses, goals, and failures, plus a
+self-referential namespace, populated by an LLM-driven multi-stage
+pipeline. That shape is the cleanest example of typology-as-substrate
+— the schema *is* the system. A tighter case borrows its core
+typology from a vocabulary every team already shares (the issue
+tracker): typed events for bugs, tasks, and messages; workflow
+primitives like *unblocked work* and *atomic claim*; edges like
+*blocks*, *parent-child*, *discovered-from*, *supersedes*. The
+issue-tracker shape earns relatively high coverage despite the schema
+commitment precisely because the categories were not minted by the
+designer [claude-c891087f:5, claude-c891087f:21]. Forks and
+downstream extensions of such systems still tend to accrete their own
+task-shape categories on top of the shared vocabulary — evidence that
+even on the most universal of work taxonomies, idiolectic additions
+are the recurring temptation rather than the exception.
+
+A more instructive third pattern: a local-first memory system that
+*does* commit to verbatim storage and refuses LLM-driven
+consolidation, putting it closer to the substrate-move side of the
+table on content. The architectural commitment to raw is unambiguous
+— no summarization, no extraction, no paraphrase. But the index layer
+is then schematized with a spatial-palace metaphor — *wings* for
+people and projects, *rooms* for topics, *drawers* for original
+content — and a separate temporal entity-relationship knowledge graph
+with validity windows is layered on top. Each "neutral" container
+carries a quiet commitment: a wing presupposes the unit of
+organization is person-or-project, a room presupposes that topics
+chunk cleanly enough to be containerized, a drawer presupposes that
+original content fits in slots. Even when consolidation is refused at
+the *content* layer, the temptation to schematize at the *index*
+layer recurs — and a separate KG layer is the same typed-relations
+move that knowledge-graph projects have been losing on for two
+decades. This is the cleanest case for the claim that the
+substrate-vs-typology choice is not just about whether you summarize:
+it shows up at every layer where structure could be imposed, and the
+discipline has to refuse it everywhere.
+
+Each substrate winner committed to *less* about what data means
+(time order, persistence, bytes, similarity) and *more* about how
+records of utterances are kept. Each vocabulary loser committed to
+*more* about what the data is *about* — categories of programs,
+record types, method names, relation types, kinds of work — and
+those commitments aged out as the domain shifted. The framing
+proposed in the archive is that retrieval quality keeps improving
+on a Moore's-law-ish curve, while the cost of ontology stays
+fixed; the substrate move compounds with the curve, the vocabulary
+move is priced against last year's retrieval [claude-c891087f:15].
+
+This is one framing of the pattern; alternative framings (e.g., that
+ontologies sometimes succeed when they encode regularities of the
+human or institutional environment rather than designer preference)
+are not engaged with in detail in this archive yet.
+
+## Relevance prediction across timescales
+
+A more general analytical frame proposed in the archive subsumes the
+substrate-vs-vocabulary pattern under a single principle: every
+cognitive architecture is a stack of relevance predictions across
+timescales, and *the only ground-truth relevance is the one made at
+integration time* [claude-14616e34:26]. The frame, as stated in the
+narration:
+
+| Layer | What it predicts | Update cadence |
+|---|---|---|
+| LLM weights | Token-level relevance given context | Pretraining (frozen) |
+| Embeddings | Semantic neighborhood of corpus items | Re-index when corpus changes |
+| Cached prefix | Per-fold/per-session operator priors | Operator decisions |
+| Strength scores (`re`, attention) | Use-pattern relevance | Continuous from substrate |
+| Retrieval ranking | Per-query relevance | Per-call |
+| Agent's act of integration | This emission's actual relevance | This moment |
+
+Each row above the bottom is a frozen guess about what the row below
+it will need. Each row pays a different amortization curve, and each
+is right exactly to the extent that runtime patterns match the
+patterns it was tuned on [claude-14616e34:26].
+
+In this frame, premature ontology is a high-confidence prediction
+baked at the wrong layer: a designer claims that *this typology will be
+relevant at every future runtime moment*, and pays the schema cost up
+front to bake that claim in. The claim is rarely justified — most
+typology decisions track the designer's particular conditions, not a
+durable property of the domain — and the baking cost compounds while
+the runtime patterns drift. Vocabulary moves bake what should have
+been deferred; substrate moves bake only what is genuinely stable
+across all anticipated runtimes.
+
+## A tolerable ontology: the LAP / Coordinator case
+
+The clearest historical case of an ontology that came closer than most
+to being *generic* is the Language/Action Perspective developed at
+Stanford HCI by Terry Winograd and Fernando Flores, productized as
+**The Coordinator** by Action Technologies in the mid-1980s
+([Stanford HCI bibliography](https://web.archive.org/web/20240623221447/https://hci.stanford.edu/winograd/action/language-action-biblio.html),
+[Dubberly: The Language/Action Model of Conversation](https://www.dubberly.com/articles/language-action-model.html)).
+The Coordinator tracked workplace email under a structured
+"conversation for action" protocol — request, promise, performance,
+declaration of completion, counter-offer, decline — drawn from
+J. L. Austin's and John Searle's speech-act theory.
+
+Four properties made the LAP ontology more tolerable than most
+[claude-14616e34:38]:
+
+- **Borrowed, not minted.** The illocutionary types came from Searle's
+  work in *Speech Acts* (1969) and *Expression and Meaning* (1979),
+  which had absorbed two decades of philological and cross-cultural
+  pressure. Winograd and Flores took an ontology that had earned its
+  scope rather than inventing one.
+- **Speech-act, not work-content.** The categories were *how utterances
+  function in coordination* (request, promise, declaration), not *what
+  the work is about* (sales, engineering, customer support). One floor
+  below the layer where universality usually fails.
+- **Role-relational, not subject-categorical.** A request implies
+  requester and performer; a promise implies maker and recipient. The
+  schema was about the shape of the interaction, not about properties
+  of the topic.
+- **Small.** Five Searle types, narrowed to four operational moves in
+  the conversational protocol.
+
+The Coordinator was nevertheless characterized, in retrospect, as a
+brittle product. Lucy Suchman's CSCW critique
+[*Do Categories Have Politics?* (1994)](https://www.lancaster.ac.uk/fass/doc_library/sociology/Suchman_categories_1994.pdf)
+argues that situated action does not follow a request → promise →
+completion protocol cleanly; that real utterances frequently combine
+multiple illocutionary functions in a single message; and that
+surfacing the categorization to users as required UI labels both
+created visible labor and politicized the categories as managerial
+monitoring devices. From the paper:
+
+> *"It is not just that the language/action perspective offers a
+> simplified model of communication. The model itself prescribes
+> categories that recast workplace communication in particular,
+> consequential ways."*
+
+Winograd's
+[response](https://www.lri.fr/~mbl/ENS/CSCW/2015/papers/Winograd-JCSCW94.pdf)
+defends the design intent without disputing the empirical observation.
+
+The general lesson: a typology becomes brittle when it is made
+operational at the moment of utterance, which is exactly where
+premature ontology hurts most [claude-14616e34:38]. The Coordinator's
+ontology was tolerable as a *lens* on conversation; it was brittle
+as the *substrate* of a tool. The corollary is to keep speech-act
+language in retrieval and prose, never in schema; let the typology
+emerge in the language people choose to use, rather than coerce
+write-time emissions into pre-specified slots.
+
+## Why the pattern recurs
+
+The archive offers two diagnoses for why the move keeps coming back
+despite a long history of similar failures.
+
+### Epistemic greed and laziness
+
+The operator's framing, moderated to refer specifically to *epistemic*
+greed and laziness rather than political or economic motives
+[claude-14616e34:17, claude-14616e34:45]:
+
+- **Greed** in the *greedy-algorithm* sense: committing to the first
+  ontology that locally seems to fit, with no backtracking. Locally
+  optimal at the moment of decision; globally suboptimal once the
+  domain shifts and the schema can't follow. The cognitive analogue
+  of a depth-first search that locks in early branches.
+- **Laziness** in the *epistemic* sense: avoiding the harder work of
+  staying in uncertainty — refusing to commit a typology until the
+  categories have earned their scope through observation. Premature
+  commitment is cheaper at decision time than holding the question
+  open; the cost shows up later, as drift.
+
+The diagnosis is procedural, not political. Designers under cognitive
+load grab the first plausible structure they can map onto and lock it
+in; the alternative — staying in free-form representation while
+patterns are still being discovered — feels like a deferred decision
+when it is in fact the better answer for systems whose domain hasn't
+yet stabilized [claude-14616e34:45].
+
+### Structure-feels-safer-than-search
+
+Stated at the engineering level: structured data feels safer than
+search; the lesson keeps not landing because the cost of building the
+structure exceeds the value of having it, once retrieval over raw text
+gets cheap [claude-c891087f:12]. Both diagnoses describe the same
+instinct from different angles — one cognitive, one technical.
+
+### The consistency-not-capability backdrop
+
+A broader pattern operates as well. The operator's framing
+[claude-14616e34:17, claude-14616e34:18]:
+
+> *"A lot of AI model capability discourse isn't about an ever
+> increasing quantity, but an asymptotically shrinking rate of
+> inconsistency across ideas, with many hoping naively that if agents
+> can just achieve consistency with all environmental constraints,
+> human choice won't be required any longer (obviously false but
+> greed/laziness perverts intuition)."*
+
+Restated: on this reading, many recent "AI capability" benchmark gains
+are better described as coherence gains — longer context, deeper
+retrieval, more reasoning steps before drift — than as capability gains. The asymptotically shrinking rate of inconsistency
+is the real metric; capability is its publicity outfit
+[claude-14616e34:18]. The connection to premature ontology is direct:
+ontology projects promise that *if the categories are right,
+consistency follows*. That promise is what makes the schema move
+attractive even after decades of evidence that the categories are
+rarely right; the recurring belief that consistency-by-machinery can
+replace human choice is what keeps the ontology trade looking
+worthwhile.
+
+The claim that consistency suffices is structurally false on five
+fronts [claude-14616e34:18]:
+
+1. **Consistency is always *with what*.** Some constraint set has to
+   be specified, and specifying it is value-laden. Two perfectly
+   consistent systems can be optimizing irreconcilable goods.
+2. **Local consistency does not imply correctness.** A coherent
+   system can be coherent with a wrong premise.
+3. **Constraints conflict.** Real environments contain trade-offs
+   that no consistency procedure can settle without smuggling in
+   priorities. The priorities are choices.
+4. **The constraint set is contested.** What counts as an
+   "environmental constraint" is itself socially constructed and
+   re-litigated.
+5. **Goodhart applies structurally.** Any consistency target one can
+   measure becomes the optimization target; whatever mattered that
+   couldn't be measured gets squeezed out.
+
+The seduction recurs because the choosing burden the alternative
+imposes is genuinely exhausting — the error is mistaking "I don't
+want to choose" for "the choice is unnecessary" [claude-14616e34:18].
+
+## What makes an ontology more tolerable
+
+A discipline distilled across the threads above [claude-14616e34:38,
+claude-14616e34:40]:
+
+- **Small.** Few categories, with no ambition to grow. The Coordinator's
+  4–5 was tolerable; a nine-table cognitive-structure schema is not.
+- **Borrowed.** Categories that have already survived field testing —
+  in linguistics, philosophy, an industry standard, an RFC — are
+  preferable to categories invented by the system's designer. Earned
+  scope beats minted scope.
+- **Role-relational, not content-categorical.** Typing *who is doing
+  what* (requester, performer, author, source) generalizes; typing
+  *what kind of thing this is about* does not.
+- **Descriptive, not prescriptive.** Useful as a lens for reading past
+  utterances; brittle as a required field at write time.
+- **Optional, not coercive.** A category that an utterance can carry
+  but is not required to carry has a much smaller blast radius than
+  one that gates emission.
+- **Multi-illocutionary by default.** Real utterances combine
+  functions — a single message may request, apologize, and propose
+  terms at once. A schema that demands a single category per utterance
+  loses information that a free-form text representation preserves
+  [claude-14616e34:38].
+
+Each of these properties trades scope for survivability. The intuition
+proposed in the archive is that an ontology that tries to be
+*comprehensive* is almost always overreach; an ontology that stays
+*minimal and borrowed* sometimes survives.
+
+## Discipline checks
+
+Several heuristic checks have been proposed in the archive for
+evaluating whether a proposed addition to a system is structural or
+typological [claude-c891087f:30, claude-14616e34:40]:
+
+1. **Observation or opinion?** Programmatic events that record
+   *what happened* (a tool ran, a file changed, a test passed) are
+   substrate-shaped. Events that record *what kind of thing happened*
+   (`bug_found`, `goal_completed`, `decision_made`) are typology
+   smuggled in.
+2. **Named accountable author, or "the system"?** A `source` field
+   that names a specific actor (`tool:claude-code:Edit`,
+   `claude-14616e34`) is fine; one set to `system` is the
+   pipeline-as-author trap.
+3. **Does it require a parent taxonomy?** A new event class that
+   only makes sense inside a hierarchy of categories is a typology in
+   miniature; refuse it.
+4. **Pre-categorize at write time?** If a proposed feature requires
+   the writer to label their utterance with a category before the
+   utterance can be appended, it is operationalizing a typology in the
+   exact place the Coordinator failed.
+5. **Visible labor, or politicized categorization?** If the category
+   surfaces in the UI as a required choice, users will perceive it
+   socially — as friction, as monitoring, as imposition. Suchman's
+   1994 critique is the field-tested instance of this failure.
+6. **Greedy commitment?** If the proposal is "let's go with this
+   typology *for now*" with no plan for how the categories will earn
+   their scope through observation, the algorithmic-greed diagnosis
+   applies. Free-form text and aggregation-from-use are the
+   alternative; both can be deferred until the patterns have shown
+   themselves.
+
+## Related
+
+- [why-wikis-work](why-wikis-work.md) — applies the same critique to
+  the compilation step that produces wiki pages. The four disciplines
+  (humans own ontology, inline provenance, read-first, backlink
+  refresh) are scope-narrowing moves of the same shape as the checks
+  on this page.
+- [sedimentary-superintelligence](sedimentary-superintelligence.md) —
+  the system-level claim about what an ontology-free substrate
+  becomes when accumulated across time. The bet that retrieval
+  over raw narrations beats extraction into typed records is the
+  positive version of this page's negative critique.
+- [sequestering-the-narrative](sequestering-the-narrative.md) —
+  carries the integrationist position about meaning being made each
+  time the narrative is re-entered. The same position animates the
+  refusal to type utterances at write time.
+
+## External references
+
+- Suchman, Lucy. *Do Categories Have Politics? The Language/Action
+  Perspective Reconsidered.* Computer Supported Cooperative Work 2,
+  1994
+  ([PDF](https://www.lancaster.ac.uk/fass/doc_library/sociology/Suchman_categories_1994.pdf)).
+- Winograd, Terry. *Categories, Disciplines, and Social Coordination*
+  (response to Suchman). CSCW 2, 1994
+  ([PDF](https://www.lri.fr/~mbl/ENS/CSCW/2015/papers/Winograd-JCSCW94.pdf)).
+- Winograd & Flores, *Understanding Computers and Cognition* (Ablex,
+  1986).
+- Searle, John. *Speech Acts: An Essay in the Philosophy of Language*
+  (Cambridge, 1969); *Expression and Meaning* (Cambridge, 1979).
+- Harris, Roy. *The Language Myth* (Duckworth, 1981); *Signs of
+  Writing* (Routledge, 1995); *Integrational Linguistics* (selected
+  papers).
+- Stanford HCI Group,
+  [Language/Action Perspective bibliography](https://web.archive.org/web/20240623221447/https://hci.stanford.edu/winograd/action/language-action-biblio.html)
+  (Wayback snapshot, June 2024; the original
+  `hci.stanford.edu/winograd/action/language-action-biblio.html` now 404s).