# PALO Workbench pilot review
8 November 2026. Scope: a supervised macOS desktop pilot for business users.

PALO Document Workbench 0.2.0 imports PDF and Excel, proposes source-linked entities, attributes, relationships, terms and process rules, and requires individual review before exporting a development blueprint. The native Mac app opens its interface in a local browser.

## Evidence, with boundaries
- Signed application and DMG accepted by Apple notarisation, stapled and Gatekeeper-accepted. This is not App Store publication or a security/compliance certification.
- 39 Python tests passed again on 8 November. Eight onboarding DOM scenarios and the existing review/export regression passed again.
- A complete 4.7 GB model download into a previously empty PALO model directory passed using the frozen backend on the existing macOS 27.2/64 GB qualification Mac. The download was explicitly cancelled at 5%, restarted, fully hash-verified, and followed by a successful real inference probe. This verifies cancellation/retry, not recovery from arbitrary network failures.
- The original compatible model elsewhere on that Mac was left intact; this test explicitly selected download rather than reuse.
- Earlier packaged synthetic PDF+XLSX inference produced 15 candidates. Two descriptions claimed uniqueness unsupported by the source; the developer corrected them before the test export. Exact citations do not establish semantic correctness. One small synthetic workflow is not an accuracy benchmark.
- Earlier full synthetic inference and OCR passed under an outbound-network-denial test harness after model installation. This harness is not a persistent firewall installed by the app.
- Port collision, owned-process cleanup, restart recovery, blocked premature export, model integrity and bundle relocation have retained test evidence.

## Open qualification
A separate clean Mac, macOS 26 execution,16 GB memory behaviour, unexpected network-loss recovery and nontechnical user acceptance remain unqualified. The app configures macOS 26+,Apple Silicon,16 GB RAM and8 GB free for download. Windows and Intel installers are not supplied. Start with the included fabricated example; arrange a supervised pilot before authorised business documents.

## What the three reviewers said
GLM 5.3FlashX, Gemini 3.1FlashLite and Venice Uncensored (Dolphin Mistral 24B) reviewed a supplied technical and functional evidence brief on 8 November. Their actual responses are retained in the local campaign dossier. These are advisory document reviews; none executed the application or certified it.

All identified model errors and export confidentiality as concerns. GLM made first-run qualification a condition; the subsequent complete download/cancellation/retry test addresses only that path on the existing qualification Mac. Clean-machine/minimum-hardware qualification remains open. Gemini warned about local process access and document parsing. Venice supported a controlled pilot while stressing human review.

We rejected or narrowed unsupported reviewer statements:
- Gemini called a same-OS-user limitation cross-user access. No cross-user exploit was demonstrated. Its suggested buffer-overflow scenario is a hypothesis, not an observed finding.
- Venice described the network-denial harness as a functionality risk. The observed harness result was successful; no such failure was shown.
- Suggestions that model agreement, citations or a review click establish correctness were not adopted.
- No autonomous production-readiness or security-by-default assurance follows from the reviews.

## Boundaries users must understand
Local processing avoids sending document content to a remote inference provider in the local workflow after model setup. It does not protect against malicious software running as the same OS user. A company endpoint is a separately configured, explicitly authorised route.
Exported descriptions can retain sensitive business logic. Some deterministic checks exist; there is no anonymisation guarantee. Review what will leave, and follow company policy. PALO does not intercept direct uploads to Claude Code, Codex, IBM Bob or any other service.

The app uses a pinned PALO reference snapshot, controlled dictionary and ontology. PALO Framework supports lifecycle and evidence, PALO-AM is represented by bounded tool-less extraction authority, and implementation checks are informed by PALO-AI. The separate PALO-AI production runtime is not integrated; canonical mandates and certifications are not issued.

## Pilot feedback
Contact info@paloframework.org with your Mac chip, memory, macOS version and a synthetic reproduction of a setup, extraction or review/export problem. Do not send confidential documents, credentials or business exports. The primary pilot questions are whether a business user can get ready, recognise an unsupported interpretation, and understand the export decision.