Verisyntax — User Manual (prototype 0.1)

Optional local AI

The desktop's shared model field defaults to qwen3.6:35b-a3b. With suggest-code-v8-task-fidelity, it passed all 10 compiler and task checks in two consecutive runs without using the repair attempt. The bounded suite includes deterministic logic, preservation, Boolean and continuous Bayesian syntax, safe negation, typed category/unit/period declarations, invalid-draft repair, and two- and three-source joins. It was also the most precise of three installed models on two full-article passage fixtures. glm-4.7-flash:latest remains selectable and was fastest in the older one-task code comparison. Context is 64K by default; 128K is an explicit choice.

The practical assessment is that qwen3.6:35b-a3b provides good assistance for bounded, reviewable Verisyntax drafts, including the tested multi-source joins and typed constructs. The versioned syntax reference, structured output, one bounded repair attempt, authoritative C++ validation, task-fidelity instructions, and preservation check make the workflow materially more dependable than unrestricted text completion. Ten synthetic tasks do not demonstrate robustness on large projects, ambiguous real schemas, long-context preservation, semantic debugging, or domain correctness. Treat it as proposal assistance rather than an autonomous developer, and read the complete draft before applying it.

Every code suggestion:

• contacts only the displayed loopback Ollama endpoint;
• checks the installed model's reported context capacity;
• receives the versioned, C++-tested Verisyntax syntax reference;
• returns structured JSON;
• is parsed and semantically checked by the authoritative C++ compiler;
• remains unapplied until Apply preservation-safe draft is explicitly approved.

There is no automatic cloud fallback. Model output never becomes a fact or active rule automatically. extraction_confidence describes extraction/structuring confidence, not the probability that a claim is true.

Imported data to reviewed fact in the standalone IDE

After metadata initialization and an import, select the database and use this exact sequence:

1. Configure reviewer stores a project-local identity; it is not read from macOS or sent to a service.
2. Load imported cells runs a bounded read-only metadata query.
3. Choose one cell and Capture selected mention. The C++ backend binds it through source_cell -> source_record -> import_run -> source_version and stores an immutable evidence mention.
4. Select one or more visible mentions and click Suggest claim locally. The model receives only the selected evidence bundle, must use its mention IDs, and returns schema-validated JSON.
5. Edit the JSON, then click Store reviewed proposal. C++ revalidates it and stores an inactive claim version.
6. Approve as fact or Reject claim is a separate action with a reason, timestamp, reviewer, version ID, and exact content hash. Approval creates the fact; rejection never does.

The Artifact and Results tabs show the raw receipts and hashes. Editing the .vsx file invalidates the workflow until it is saved again, preventing a decision from being applied to another document snapshot.

Retrieve approved project logic

Approved facts are persisted in vs_facts with their exact claim version and fact JSON. Active rules are persisted as compiler-validated IR in vs_rule_versions. Export the current active set as one deterministic, read-only JSON document:

<compiler>/verisyntax logic export model.vsx local > approved-logic.json

The verisyntax.approved-logic-library version 1 document includes stable names, exact version IDs, content hashes, reviewer, reason, and decision time. Proposed, rejected, and superseded versions remain in the audit database but are excluded. C++ callers can use load_approved_logic_library(program, "local") from verisyntax/review.hpp; it uses two sorted queries and verifies stored content against the approval hashes. The desktop performs the same retrieval when Add project facts, code & schema to Request is clicked.