sign-lang synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From authored knowledge to agent context
SIGN is a knowledge notation and deterministic preparation toolchain, not an agent-turn scheduler. It expresses domain definitions, relationships, inference rules, constraints and provenance for an external agent to interpret. This account covers the repository https://github.com/sign-protocol/sign-lang at f4eb1a82b114615754bbe5c2e6f4255a4ec1eeba, with an applicability cutoff of 2026-11-09. The evidence is inspected implementation and doctrine/design from SRC-1; production and efficiency assertions are reported claims, not supplied execution evidence. The boundary is complete artifact, partial loop: shipped canon sources, compilation and maintenance are included, but the enclosing model consumer, external importer and external drafting implementation are not frozen here. No target execution, inspectable agent run or causal experiment establishes delivery, compliance or benefit (boundary, Source register).
The intended progression is economical orientation followed by requested detail: inject a compact corpus index at each invocation, then let the agent request full documents by ID. SIGN implements preparation of those artifacts; the host must implement their actual delivery. The installed Node.js CLI exposes validate, compile, coverage, drift, diff and import. Fixed JavaScript owns progression within the preparation commands. It does not choose a model, invoke one or execute its proposed actions (RT-CMP-validator, RT-CLM-agent-contract).
Compilation admits structure, then produces different projections
The representative workflow starts when a caller supplies source and output paths. Directory discovery selects .sign files; a directly supplied file has no extension filter. The README's Markdown quick-start example therefore does not establish a Markdown-to-SIGN conversion route in this compiler. Parsing and rules consume structured fields, while authored domain prose remains material for a later human or agent. Source authorship and actual per-content human approval are unresolved (workflow account, MEM-OBJ-canon-source).
Compilation aggregates parse errors, runs rules with strict link checking and aborts before writes on recognized errors. This is a wired pre-write veto over this command's recognized error array, not epistemic acceptance of the corpus. Warnings permit continuation. Standalone validation is weaker: unresolved links can warn, and its zero-document early return can report success while retaining malformed-header errors. That reachable validation case does not bypass compilation's separate error check (RT-RTE-compile-canon, RT-RTE-validate-source).
After that gate, the index and manifest select identified active documents. The index copies source-authored summaries rather than generating them. It omits rule bodies and anchor exclusions, and displays at most four distinct anchor include patterns plus a remainder count. Its nominal 400-token check uses a characters/4 estimate over selected fields, excluding build headers; it is not a tokenizer-independent prompt budget. This is static context preparation, not history compaction, turn eviction or isolated sub-agent work (MEM-OBJ-canon-index, RT-RTE-compile-canon).
The full bundle has a materially different selection rule: it copies each entire raw source file contributing any active document. A file containing both active and inactive documents can therefore put both into the bundle. Changing a document's status alone is not universal withdrawal. Strict links are checked against all parsed IDs before active filtering, so a successful compile also does not guarantee retrieval by a later active-only resolver. Raw authoring, including reasons and provenance where present, can survive bundling; the index cannot substitute for that full text (MEM-OBJ-canon-bundle, RT-RTE-compile-canon).
Successful compilation writes an index, bundle, manifest and checksum. The manifest describes selected identities and build metadata; the checksum identifies assembled bytes, not their author or truth. No later metadata verifier is established. Raw files are reread after checking, requiring stable inputs for correspondence between checked structure and written text. Four sequential writes provide no transaction or rollback: pre-write rejection can preserve previous artifacts, while a later exception can leave a partial update. The caller owns repair and retry; compilation does not expire previous outputs or revoke cached host material (MEM-OBJ-canon-manifest, MEM-OBJ-bundle-checksum, RT-RTE-compile-canon).
Delivery and reliance belong to a separate consumer
At the implemented workflow boundary, delivery means filesystem artifacts, not an accepted agent answer. The README affords an external host role that supplies the entire index in the system prompt and registers get_canon_doc(id) through an unspecified resolver. The whole-index supply is coarse automatic delivery; the full-document return fulfills the agent's explicit request. There is no implemented task-relevance ranking, deployed resolver, full-document budget or host cache policy established here. IDs are locators requiring that resolver, not references the model window can automatically follow (MEM-RTE-canon-delivery, workflow account).
This channel affords knowledge and instructions through model interpretation, not tool-action enforcement. Domain inference and constraints are expressed, but the inspected checker tests Boolean presence of when:, then:, confidence: and src: prefixes. Even empty values can satisfy those presence tests. It does not evaluate premises, derive a domain fact, authenticate a reviewer or accept a proposition against evidence for an intended use (RT-BAP-canon-guidance, EPI-RTE-infer-check, EPI-CLM-governed-inference).
The governance reference requires human review before promoting model-estimated facts to asserted authority. Other specification passages permit human or expert-system assertion and describe a proposed machine-drafted convention. These remain differing source conventions, not one enforced approval mechanism. Provenance text can survive raw retention without becoming authenticated history. Neither an active label, checksum nor compiler pass closes that gap (EPI-CLM-provenance-authority, EPI-OBJ-edge-provenance).
Maintenance produces bounded signals, not a revision loop
Coverage maps parsed anchors to current repository files. Drift intersects those governed files with Git changes; absent deleted files and uncommitted changes are not universally covered. These commands discard parse errors in their loading projection and are not mandatory compilation predecessors. Their outputs can prompt external review, but do not persist a review queue or repair canon (RT-RTE-source-signals).
Diff compares selected textual signatures and anchor sets. Its CLI reconstructs history using current paths and IDs, while the classifier iterates after-documents, limiting deletion detection. A source-level deduction shows that changing only an edge's provenance from inferred to asserted can remain classified additive because edge properties are outside its seven meaning sigils. This is not an observed unauthorized promotion, but it prevents interpreting the classifier as a complete semantic or authority-change gate. Breaking-change revalidation remains a requested downstream obligation (RT-RTE-corpus-diff, epistemic comparison deduction). Import is likewise a handoff: the CLI discovers and executes a separate package; its content transformations, approval and persistence remain uninspected (RT-RTE-import-delegation).
What the supported contribution does—and does not—establish
The strongest supported contribution is machine-enforced structural preparation with a reduced orientation layer and preserved raw detail. For a caller needing that preparation, the checks and projections are inspectable capacities. For a scenario requiring authenticated knowledge, reliable exclusion of inactive text, completed review or compliant agent action, those same outputs do not establish the needed guarantees. Retention is not acceptance followed by lifecycle integration, and transient maintenance reports are not automatic revision (RT-RTE-compile-canon, MEM-OBJ-canon-index, epistemic conclusion).
The expressed inference rule supplies localized inspectable content consistent with theory-builder condition 1's minimum. The records do not establish condition 2's content-dependent agent decisions, condition 3's content-directed criticism and resulting revision or changed reliance, or condition 4's retained criticism shaping another round. Fixed structural checking is not evidence of those operations. No criticism-mediated improvement in future capacity is demonstrated. Reflection, autonomous theory building and self-improvement are independently unestablished at this partial-loop boundary; deterministic command progression alone establishes none of them. Likewise, no qualifying trace-fed durable update and later consumer is established, but unresolved initial provenance and excluded host/importer behavior do not warrant a system-wide trace-learning negative (EPI-CLM-governed-inference, epistemic conclusion, MEM-OBJ-canon-source, MEM-RTE-canon-delivery, profile verification).
A frozen host implementation and traces showing exactly which rule text reaches decisions would change the delivery assessment. Authenticated evidence-consuming approval would change the warrant assessment. Per-document bundle filtering, snapshot inputs and transactional output replacement would change the exclusion and consistency guarantees. Retained criticism, an admitted organizational update and subsequent behavior causally depending on it would be needed to assess exercised self-improvement; outcome comparisons would separately be needed to establish benefit. No new source reads were required for this synthesis.
Limitations
| limitation | affected SRC IDs and record citations | inspected boundary | conclusion prevented | evidence that would resolve it |
|---|---|---|---|---|
| External consumption and run evidence are missing | SRC-1; MEM-RTE-canon-delivery, RT-BAP-canon-guidance | Preparation implementation and README integration sketch, before host/model execution | End-to-end retrieval, complete returned content, compliance, revocation, token savings and causal benefit | Frozen host/resolver implementation plus attributable invocation and outcome traces; controlled comparisons for benefit |
| Initial source admission and external transformations remain unresolved | SRC-1; MEM-OBJ-canon-source, MEM-OBJ-skill-reference, RT-RTE-import-delegation | Caller-supplied source, static drafting reference and importer dispatch only | Actual human approval, authenticated provenance, imported-content quality, or a complete learning classification | Authoring and review records, frozen external implementations and traceable later consumers |
| Selection and persistence guarantees are projection-specific | SRC-1; RT-RTE-compile-canon, MEM-OBJ-canon-bundle | Active identity selection, whole-file raw reread and sequential writes | Universal inactive-content withdrawal, active-only link retrieval, snapshot consistency or atomic recovery | Per-document filtering and resolver contract, stable-input enforcement and transactional persistence/recovery evidence |
| Metadata has no established later verifier | SRC-1; MEM-OBJ-canon-manifest, MEM-OBJ-bundle-checksum | Known metadata producers, no downstream checking route | Enforced artifact consistency, author authenticity or trusted deployment | Implemented consumer checks and attributable use; authentication evidence distinct from byte equality |
| Maintenance comparisons are bounded and do not complete review | SRC-1; RT-RTE-source-signals, RT-RTE-corpus-diff | Current-file/Git intersection and selected textual signatures | Complete deletion/change detection, semantic equivalence, or completed revalidation and repair | Expanded comparison implementation and an evidenced review/admission/repair continuation |
| Optional citation and verification-posture vocabulary is not separately inventoried | SRC-1; MEM-OBJ-canon-source, EPI-OBJ-sign-content, report verification limit | Verification inspected optional @sources citation triples and @currency text in proposals/2026-08-sources-currency-blocks.md, and parser/rules/compiler ordering/raw retention |
Exhaustive provenance/freshness-vocabulary comparison; omission must not imply absence of citation or verify-before-use guidance. Consumer compliance and external citation checking remain unestablished | Separate accepted records for these parts and any implemented checking/consumer routes |
| Partial authority comparison omits strict-link reference criteria | SRC-1; RT-RTE-compile-canon, RT-RTE-validate-source, MEM-OBJ-canon-source, profile verification limit | Retained source identities are reference criteria: strict missing-target errors can veto compilation, while non-strict validation warns | Complete structural validation/enforcement authority comparison; emitted routing and afforded guidance remain valid, without implying domain-truth or action enforcement | Separately classify these retained criteria and their strict/non-strict consumers in a revised profile |