Polish Sema paper tightening

paper/sema.tex

@@ -538,13 +538,13 @@ \subsection{Refinement}
 
 The vocabulary is not a static artifact; it evolves through new mints, retirements, mechanism rewrites, and relocations. Refinement is one concrete example of how that evolution proceeds: a cross-pattern pass in which the library is swept end-to-end against the current governing rules, rather than pattern-by-pattern at mint time. With hundreds of patterns in place, questions of layer coherence, redundancy, and principle consistency cannot be answered one pattern at a time; they require systematic passes against shared formal criteria. Refinement is the stage of the methodology where those passes happen.
 
-A refinement pass splits the work between two frontier models running in distinct roles. In the \emph{proposal} role, Claude Opus 4.7 scans the vocabulary for coherence failures, misplacements, drift-from-canon, and layer miscategorizations, and produces staged mechanism rewrites and relocation proposals. In the \emph{review} role, Gemini~3.1 Deep Think adversarially evaluates those proposals. The round-trip catches failure modes no single model reliably detects: proposer blind spots that slip past self-critique, reviewer conservatism, and mutual-reinforcement bias in which both models agree to a vague rule that neither would have accepted in isolation. Each round produces either a verdict on pending proposals, a forensic catch (for example, a mathematical correction: a gradient cannot be defined on a bare metric space without differential structure), or a sharpening of the governing principle itself. Over the course of several rounds, the process has yielded hundreds of per-pattern refinement decisions spanning relocations, retirements, and mechanism rewrites.
+A refinement pass is governed by the versioned vocabulary design manual rather than by ad hoc judgment. Proposed relocations, retirements, mechanism rewrites, and rule clarifications are tested against the current manual, adversarially reviewed, and finally adjudicated by a human maintainer. The paper-level claim is the audit discipline, not a particular review workflow: each accepted change must state the rule it depends on, the failure mode it prevents, and why the alternative was rejected.
 
 Concrete refinement decisions illustrate the character of the pass. \sema{Bid}{464a} was initially filed under Society/Economics on the intuition that bids are a social-economic concept; the mechanism-sufficiency test reassessed this kind of judgment call by asking what the pattern's mechanism structurally requires (in the Bid case, the decision retained the Society classification because the mechanism requires a counterparty, but the same method resolved many analogous ambiguities in the opposite direction). Earlier drafts also left layer direction unenforced on the argument that layer metadata is not hashed and the protocol is therefore agnostic to layer choice; the refinement kept that degree of freedom intact at the protocol level but tightened the bootstrap library's policy to enforce layer direction as a \texttt{sema apply} gate. Many other decisions were similar in character: a concept that had been filed by vibe or typical-use-case was re-tested against what its mechanism structurally required, and moved when the two disagreed.
 
 A second class of refinement decision concerns \emph{scope} rather than \emph{placement}: a pattern may be coherent, well-mechanized, and correctly layered, yet still inappropriate for the default library that downstream consumers install by default. During refinement, a subset of patterns whose canonical applications involve capability amplification, social manipulation, evasion, or cryptoeconomic binding (for example, \texttt{AmendLaws}, \texttt{ChaosDrift}, \texttt{CryptoShred}, \texttt{IdentityMask}, \texttt{MirrorStake}) were judged to belong in a separate experimental shelf rather than in the default library. The patterns remain resolvable at the hash level---both databases stay in the repository---but the default install stays conservative, and engaging the experimental shelf is a deliberate act by a user who has read what they are opting into. This splitting decision is made at refinement time, per-pattern, on the basis of how the mechanism composes with realistic deployment contexts rather than on whether the mechanism is well-formed (the latter is already checked by the other gates). The resulting distribution architecture is described in Section~\ref{sec:distribution}.
 
-The result is relevant to the paper's main claim: a content-addressed vocabulary is only as rigorous as the process that produces its definitions. A single-model refinement pipeline reflects one model's idiosyncratic drift. Engaging multiple model families in distinct \emph{proposal} and \emph{review} roles converts that drift into testable disagreement---every relocation, retirement, and mint in the current library has survived scrutiny by a model that did not propose it and a human author who adjudicated the disagreements. The per-pattern reasoning from this pass---each pattern's intended use, the rationale behind its invariants, and the design commentary supporting every relocation, rewrite, and retirement---is shipped with the library as a per-release design manual,\footnote{\url{https://github.com/emergent-wisdom/sema/blob/main/docs/manuals/vocabulary-design.md}} so that the review surface of the vocabulary is visible rather than implicit in commit history.
+The result is relevant to the paper's main claim: a content-addressed vocabulary is only as rigorous as the process that produces its definitions. The per-pattern reasoning---each pattern's intended use, the rationale behind its invariants, and the design commentary supporting relocations, rewrites, and retirements---is shipped with the library as a per-release design manual,\footnote{\url{https://github.com/emergent-wisdom/sema/blob/main/docs/manuals/vocabulary-design.md}} so that the review surface of the vocabulary is visible rather than implicit in commit history.
 
 \subsection{Distribution: A Canonical Vocabulary that Consumers Pull}
 \label{sec:distribution}
@@ -573,7 +573,7 @@ \subsection{The Creative Discovery Loop}
 
 The loop would proceed in four phases. First, in \textit{Hypothesize (Generativity)}, an agent detects friction or explores latent space (via \sema{ConceptBlend}{29a9}) to generate a candidate mechanism. Second, during \textit{Judge (Evaluative Merit)}, the candidate would be passed to the \sema{Judge}{efe0} primitive for a scalar assessment of structural merit. Third, the \textit{Harden (Adversarial Evolution)} phase would subject the pattern to \sema{AdversarialSteel}{ded9}, where a ``Devil's Advocate'' attempts to exploit its invariants. Finally, in \textit{Mint (Crystallization)}, the survivor is hashed, becoming an immutable pattern available for immediate citation and reuse. The infrastructure for the final step---validation, hashing, and atomic minting via the \texttt{sema\_mint} tool---is fully implemented; the preceding three steps require either human judgment or LLM-based evaluation that remains future work.
 
-The Hypothesize phase admits many concrete strategies, each suited to a different kind of latent space. Friction-driven minting responds to a specific coordination failure: English permits a rule to be stated two ways that both sound right but produce divergent behavior, so a pattern is minted to pin the semantics. Refinement-driven minting emerges from systematic review (as in Section~\ref{sec:refinement}): a principle the library keeps tripping over crystallizes into a pattern only after the absence is felt across several examples. Companion-paper-driven minting imports concepts from a domain text whose mechanisms the library wants to be able to cite. Lexicon-anchored ideation is another strategy, seeded from the system dictionary: sample a word, reason through the sema lens about whether it names a coordination-relevant concept that existing patterns do not already cover, and mint only when a genuine gap survives scrutiny. Each strategy produces candidates from a different prior, and all of them ultimately converge on the same downstream gates---Judge, Harden, Mint---so the library grows through diverse ideation into a single, uniformly vetted vocabulary.
+The Hypothesize phase can draw candidates from several sources: observed coordination failures, systematic refinement passes, imported domain mechanisms, or lexical exploration. These sources differ in where the candidate comes from, but not in how it becomes part of the library. Each must pass through the same downstream gates---Judge, Harden, Mint---so diverse ideation converges into a single, uniformly vetted vocabulary.
 
 \subsection{Evaluation as a Primitive: The Judge}
 \label{sec:judge}
@@ -953,7 +953,7 @@ \section{Related Work}
 \subsection{Contribution Summary}
 \label{sec:contributions}
 
-Against this landscape, Sema makes seven contributions, separable into protocol-level mechanisms (C1--C5) and library-level design (C6--C7).
+Against this landscape, Sema makes seven contributions, separable into protocol-level mechanisms (C1--C5), compositional schema design (C6), and bootstrap-library design (C7).
 
 \begin{description}
 \item[C1: Hash-as-word inversion.] Sema moves content-addressed identity into the language agents already use. A Sema word such as \sema{Delegate}{ba86} is both a readable handle and a verifiable pointer to a canonical behavioral contract.
@@ -968,7 +968,7 @@ \subsection{Contribution Summary}
 
 \item[C6: Compositional vocabulary.] Patterns reference other patterns by hash, forming a wired DAG of behavioral specifications rather than a flat dictionary. Resolving dependencies exposes the cognitive stack a pattern relies on; execution still requires implementations or tools layered on top.
 
-\item[C7: Bootstrap library and Grammar of Agency.] The paper provides an initial \semaPatternCount-pattern vocabulary organized by dependency layer and by agency role: verbs, nouns, and values. This is a second-order contribution: the protocol is schema-agnostic, but coordination still needs a shared Schelling point.
+\item[C7: Bootstrap library and Grammar of Agency.] The paper provides an initial \semaPatternCount-pattern vocabulary organized by dependency layer and by agency role: actions, data shapes, and coordination/evaluation primitives. This is a second-order contribution: the protocol is schema-agnostic, but coordination still needs a shared Schelling point.
 \end{description}
 
 %==============================================================================