Skip to content

neuralOSd

The verified agent runtime. An offline, on-device data-answering runtime that refuses what it cannot verify, accounts for what it excludes, and ships with its own tests. The model picks; the runtime verifies; the code answers.

Quick Start

pip install 'neuralosd[all]'
neuralosd init --source your_data.csv --name mydata --out ./mydata
neuralosd ask --instance-dir ./mydata "how many rows"

Why it's different

Every other AI data tool optimises for answering. neuralOSd optimises for not answering when it can't verify. That's the opposite design goal, and it's what makes it safe for regulated environments.

Every other tool neuralOSd
The model generates an answer The model selects from a closed list
If the answer looks wrong, add a prompt If the match is weak, refuse
Verification = ask the model if it's sure Verification = accounting, invariants, oracle
Errors are silent Discards are visible
Deployment = "it should work" Deployment = verified end to end

What's new in 2.7.0 — durable builds + doctor

  • Pi Durable — neuralosd build "<capability>" --durable: the build conversation runs in a crash-resumable Pi Durable harness with exactly-once submissions; --resume continues after a crash. neuralosd pi durable --status|--steer|--stop for operator sessions.
  • neuralosd doctor — health-checks every integration in one command.
  • Faster ops on big audits (bounded reads) and smarter build prompts (bridge inventory + probe names + data sample up front).

What's new in 2.6.0 — natural language builds

  • pi as the builder — neuralosd build "average shipping cost per unit by region": the pi coding agent writes the probe code, neuralosd verifies it (banks untouched, instance loads, banks replay green — anything else is restored). Default model: z-ai/glm-5.3-flash.
  • pi as the operator — neuralosd pi ask "why did refusals go up?": natural-language ops over the closed command list.
  • neuralosd pi --setup verifies pi, the model, and a live ping.

What's new in 2.5.0 — agent2agent + durable

  • synapse-gateway, for real — the mesh client now speaks the documented gateway REST surface (/api/mesh/*, Bearer token from REACTORPRO_GATEWAY_TOKEN or gateway.env). Ask any peer: neuralosd mesh send --target grip-001 "summarise the build". Durable tasks, mailbox, trust — and mesh setup to verify it all.
  • Temporal durable execution — the loop's recurring work (golden bank, gap mining, shadow evaluation, snapshots) as durable activities: neuralosd temporal --setup, temporal worker, temporal schedule-create --cron "0 2 * * *".

What's new in 2.4.0 — sources, proof, trust

  • agent-browser connector — init --source browser:<url> turns a web page into a profiled source through the agent-browser CLI: accessibility-tree snapshots become rows, roles become caged enums, and neuralosd browser --session drives persistent sessions (open / snapshot / click / fill / network). Verified answers now reach the web.
  • Authenticated API sources — init --source api:<url> with credential caging: the instance stores only a secret reference (scrt: / op: / env:), resolved at ask time — the secret never touches disk.
  • Multi-source instances with caged joins — --extra-source + --on: one instance, several row sources, one deterministic join.
  • Value-level oracles — golden cases now verify the ANSWER with typed shapes (count / percent / ratio / date / text) and invariants. Wrong value ≠ pass, even when routing is perfect.
  • Snapshot pinning — snapshot --label q3 freezes hashed observations; ask --as-of q3 answers from the pin after the source changes. Answers become reproducible evidence.
  • RBAC / row-level caging — acl.json gives roles probe allow-lists and row filters; ask --role analyst serves only what that role may see.
  • Refusal-rate SLO + fleet dashboard — slo and fleet quantify the menu-gap signal; exit 1 on breach.
  • Golden-bank CI badge — badge emits a shields.io endpoint + a GitHub Actions gate; bank health lives in the README, outside the box.
  • Cluster-driven synonyms — synonyms proposes triggers from gap clusters (audit score vectors + caged enums as ownership evidence); human-approved, applied to derived metrics, emitted as edit suggestions for probes.py.
  • Streaming windows — log instances get a caged recent_events(minutes) probe: ask over "the last hour".

What's new in 2.3.0 — the loop grows up

Ten closing moves that turn the verify-answer loop into a self-improving one. Full details in Usage §6b and the changelog.

  1. Probe birth traps — every reason-installed probe synthesizes its misuse trap into traps.json and proves the refusal with the live router before the loop reports success. A capability ships guarded, or the install fails loudly.
  2. Refusal clustering — gaps clusters refusals into patterns with cost framing: "this gap cost you 412 refused queries across 38 phrasings." Prioritize by what users repeatedly want.
  3. Negative-feedback ingestion — neuralosd feedback turns a thumbs-down into a bank candidate with the user's correction attached; --apply promotes only proven candidates.
  4. Schema-similarity probe transfer — neuralosd transfer proposes a schema sibling's probes with the golden bank translated and re-oracle flagged. Proposals never auto-install.
  5. Oracle provenance — golden --record stores who verified a value, when, from what source, how; golden --receipts prints the evidence chain. The golden bank is an evidence chain now, not just a test file.
  6. Shadow mode — reason --shadow installs probes that cannot serve traffic; shadow --evaluate/--promote promotes only on oracle agreement. Kills the "generated probe confidently wrong" class.
  7. Consensus — reason --consensus-model requires a second, DIFFERENT model to independently pick the same ids; disagreement refuses the install.
  8. Auto-deprecation — neuralosd mothball demotes probes nobody queries (kept on disk, reinstatable). Tight menus route better.
  9. Schema-drift re-audit — neuralosd drift re-profiles a changed source and flags the probes it breaks before they fail at runtime.
  10. Federated gap learning — gaps --export-fingerprints / --ingest-fingerprints: peers on the mesh share anonymized token-pair hashes + counts, never the questions. Capability discovery, crowd-sourced, privacy-preserving.

What's new in 2.1–2.2 — more sources, deeper Excel

  • SQLite connector — neuralosd init --source app.db profiles a database with stdlib sqlite3 (zero deps, works in the frozen binary).
  • Deep Excel — formulas, every sheet, and named ranges profiled; sheets / named_ranges / formula_scan probes scaffolded.
  • Log sources — syslog:<path> (RFC3164/5424, stdlib only), journal:[unit] (journald), wevtutil:<Channel> (Windows Event Log) — all normalized to {ts, host, app, level, message} events.
  • Auto-trap synthesis — every registered probe gets its misuse trap in traps.json at birth (NEURALOSD_AUTO_TRAPS=0 disables).
  • Cross-instance bridge fix — no instance can inherit another's data layer.

Also in recent releases: the neuralosd_sdk Python/JS clients, bearer-token auth on serve, and the import auditor (neuralosd audit).