Verisyntax — User Manual (prototype 0.1)

Practical article evidence project over time

Use this workflow when you want to collect articles over time, extract claims with a local model, approve facts, and later examine their logical relationships.

A complete fictional exercise project is included:

<example-project>
├── article_analysis.vsx
├── <article-storage>/
│   ├── article_2026-08-01.txt
│   └── article_2026-08-15.txt
├── data/
└── reviewed_logic.vsx

Copy the folder to your own workspace, or let Start Center create the same base structure.

Create the project with the guided article workflow

1. Open Start and choose New article evidence project.
2. Select an empty project folder. The IDE creates article_analysis.vsx, <article-storage>/, and data/, opens the model, and shows Article workflow. It imports and runs nothing, and refuses to overwrite an existing .vsx project.

The project also receives a validated <managed-storage> that binds all project .vsx files and marks the primary model. Reopening the project automatically opens that model instead of the starter demo. Use Update VSX project binding… after adding another analysis or rule model.

3. Click Prepare database…. The app checks the model and asks for approval before initializing local SQLite metadata. You do not need to remember the Check → database → metadata sequence.
4. Save the articles as ordinary UTF-8 .txt files in <article-storage>/. Preserve the natural prose; do not rewrite it into one logical passage per physical line. Click Choose article files… and select several files if useful. The app shows one batch preview, skips SHA-256 hashes already in import history, and requests one combined approval. Every file still receives a separate source version and import run. The first sample article contains:

On 1 August 2026, the North Harbour board said that 20 new charging points were fully funded.
The board expected installation to finish before 31 December 2026.

1. Choose Create for first file, then append remaining files in a new project or Append every new file in an existing one. Review hashes, destination, and the line → line mapping, then click Commit batch.
2. Close and reopen the resumable guide with Article workflow…. It reports whether metadata and imports already exist. Optionally use Preview rows to inspect the searchable line view.

Extract and approve a claim

1. Open Claim review, click Load imported text articles, and choose the article's hash-bound text snapshot.
2. Keep the recommended local article model and 64K, then click Find logic passages locally. The first model stage reads the full raw text and proposes verbatim, non-overlapping passages with types such as claim, causal, quantitative, or temporal.
3. Verisyntax derives every byte offset from the original file. It rejects the whole result if the model paraphrases, invents, or reorders an excerpt. Accepted passage drafts are stored only as source-hashed mentions, never as facts or active rules.
4. Select one passage or a small coherent evidence group under Evidence mentions, then click Suggest claim locally. This second model stage proposes one structured claim from exactly the selected evidence. Repeat this step for additional independent claims.
5. Review the editable JSON, especially subject, relation, object/value, attribution, period, evidence, and extraction_confidence. The latter measures structuring confidence, not truth probability.
6. Click Store reviewed proposal. It remains inactive.
7. Click Approve as fact or Reject claim, enter a reason, and confirm the version ID and content hash. Approval is the step that creates an active fact.
8. Click Open claim library to inspect every version, status, evidence hash, and active_fact read-only.

The two model stages are deliberate: verbatim passage selection can be checked against the source before claim structuring interprets it. A model may later propose a rule from approved predicates, but parser/type validation and separate user approval are still required. The LLM never calculates posterior probabilities.

The four right-side functions

• Local AI · Script help creates an editable program proposal. Choose a local model and normally 64K context. Suggest a Request… turns your goal and desired output into an editable, schema-aware Request without contacting Ollama. Use Add project facts, code & schema to Request for questions that depend on project facts or rules. It hash-checks every registered .vsx file and loads active facts/rules from the project's metadata databases, but does not read ordinary domain rows. Suggesting contacts Ollama only after approval. Apply preservation-safe draft is enabled only when C++ validation confirms that the draft is valid and has not removed existing facts, rules, or data declarations. It does not approve new facts or rules.
• Claim review follows text snapshot → verbatim mention → editable claim → stored proposal → approve/reject. Check scope, attribution, time, and evidence. Extraction confidence is not truth probability.
• History & provenance is read-only and exposes import hashes/runs, their records/cells, inference, native execution, and evidence mentions.
• Rule review stores reviewer identity and exact-hash decisions. Confirm Rule(s) opens a readable library with each exact version, status, Verisyntax source, safety/stratification, model provenance, and verbatim evidence mentions. Versions awaiting review are selected by default so several can be approved or rejected in one workflow. You must still enter a reason and confirm every displayed content hash; selection never activates a rule automatically. A proposal must use known predicates and pass Verisyntax validation. Local AI never activates a rule or calculates probabilities.

Each section also contains expandable How to use this panel help in the app. Hide the complete side with Hide right panel and restore it with Show right panel; the choice is remembered locally.

Add a later article

The second sample article reports a different number for the same project:

On 15 August 2026, the contractor said that only 12 North Harbour charging points were funded.
The contractor said that the remaining eight points required a later budget decision.

1. Open Article workflow…, choose Choose article files…, and use Append every new file. Several new articles can be selected in one batch.
2. Review the new file hash and preview. Commit adds rows and a new text snapshot without rewriting the first article.
3. Reload text snapshots, run Find logic passages locally, and repeat claim review for the statement about 12 funded points.
4. Compare sources, timestamps, and versions in Open claim library. Conflicting source statements may coexist; Verisyntax preserves both evidence chains.

For every later article, repeat this small loop: save the raw text, import with append, let the local model propose verbatim logic passages, review claim proposals, and record the decision.

Ask local AI to help formulate a question

Open Start → Question for local AI, choose Analyze my current data, and click Suggest a Request…. When you accept that dialog, the app itself adds validated project facts, code, and schema before your goal. You do not need to click Add project facts, code & schema to Request afterwards; if you do use it to refresh context, the edited goal is preserved. Enter:

Return the complete current program with a dataset that lists every imported
article passage as text. Use only the declared evidence database and
article_lines source. Do not invent article contents or facts.

The model receives every registered project program and its declared facts and rules, not table rows. Other project files are context only; the draft targets the current file. A dataset is run through the IDE dataset selector and must not end with ask dataset_name; ask is only for a declared logic-predicate query or probabilistic variable. The draft must pass Check, and applying remains blocked if an existing fact, rule, or data declaration is missing. For a logic draft, you can therefore refer directly to already declared project facts:

Create a complete logic-only Verisyntax program from these human-reviewed facts:
the board article reports twenty funded points for north_harbour, and the
contractor article reports twelve. Add a funding_disagreement rule and ask
whether north_harbour has a disagreement. Do not add new factual claims.

The included, compiler-tested reviewed_logic.vsx is:

reports_funded_points(board_article, north_harbour, twenty)
reports_funded_points(contractor_article, north_harbour, twelve)

funding_disagreement(Project)
  if reports_funded_points(board_article, Project, twenty)
  and reports_funded_points(contractor_article, Project, twelve)

ask funding_disagreement(north_harbour)

Run it through Check → Compile & build → Program queries → Run. The result is TRUE, and the proof shows both reviewed facts and the derived rule. This is a deliberate human-reviewed program step; metadata facts never become logic predicates automatically.

SQLite is the simplest recommended choice for one person working locally on one Mac. Choose PostgreSQL when several processes or users need to share the database, or when operations and volume require a database server. The metadata schema keeps sources, mentions, claim versions, reviews, and facts separate for auditability; this is safer than flattening the entire evidence chain into one table, although article passages may share the article_lines table.