Recruitment open / controlled Mac pilot / 0.2.0
Build with AI.
Keep humans in command.
MISSION 001 BEFORE YOU CODE
Turn PDF and Excel processes into a blueprint you review before coding. PALO Document Workbench runs local AI on your Mac, helps you inspect the data and workflow, and requires your review before export.
For business people who know the process, even if they do not write the code.
01 / The reason
An Excel workbook. A process PDF. An idea for an app. The journey often starts by sending the whole business process to an AI coding tool.
PALO offers an earlier review step: analyse documents locally, inspect what the model inferred, resolve assumptions, and decide what may leave the workspace.
Use the route your organisation authorises. PALO does not override company policy or intercept uploads made directly to Claude Code, Codex, IBM Bob or other services.
Human Agency / Recruitment film
A short invitation to question what AI infers, review the evidence and help shape the PALO pilot.
3:59 English narration and captions
02 / Your mission
No command line. No separate Python or Ollama installation. Open the Mac app and work through its local browser interface.
The guided setup checks the Mac. With your consent, it downloads the model or verifies a compatible installed copy, then runs a test.
Try the included fabricated process before using documents you are authorised to analyse. PDF and Excel sources stay in the selected local workflow.
Inspect proposed entities, relationships, definitions and process rules. Use the ER graph, dictionary and source references to challenge assumptions.
Resolve individual candidates before exporting a development blueprint. Check the exported business logic before sharing it with any coding tool.
A quotation is not proof that an inference is correct. Descriptions, identifiers and relationships can still be wrong. Human review is a required part of this pilot.
03 / Inside the app
Actual English interface captures from the signed desktop preview. These show the existing-model reuse route on the qualification Mac.
04 / Three PALO layers
The Workbench uses a pinned PALO reference snapshot, controlled dictionary and ontology. The business ER graph is an application projection.
Connect the document case, lifecycle questions and review evidence. Keep sources and unresolved assumptions visible.
The extractor proposes candidates within a defined task. It has no tools to execute the business process or deploy an app.
Model integrity, source references, graph constraints and export conditions are enforced by Workbench controls.
The separate PALO-AI production runtime is not integrated. This pilot does not issue canonical mandates, establish compliance or certify the correctness of a business model.
05 / Where inference runs
YOUR MAC
The native launcher starts the bundled backend and opens your browser on this Mac. After preparation, the installed model performs inference locally. There is no automatic cloud fallback.
Local processing does not protect against malicious software running as the same Mac user. Keep the device and its access controls secure.
YOUR ORGANISATION
An advanced connection can point to an explicitly configured company inference endpoint. Your IT team must approve the provider, data handling and network route.
This option is not a managed enterprise deployment, tenant isolation, SSO or an automatic integration with coding tools.
Engineering update / measured 9 November 2026
Three original public PDF repositories, each larger than 100 MB, completed import, local inference and reviewed blueprint export. This is execution evidence for a local 0.3.2 candidate, not qualification for unassisted business use.
One run per repository on an Apple M3 Max with 64 GB RAM and macOS 27.2. The model was Liquid LFM3-1.7B Q4_K_M. Import and inference exclude source inspection, curation and export review.
| Public repository | Original size | Import | Inference | Batches | Export |
|---|---|---|---|---|---|
| University Lakes restoration | 111.95 MB | 20.151 s | 249.585 s | 24 / 24 | 11 curated candidates |
| Dayton Mall redevelopment | 163.31 MB | 175.288 s | 619.398 s | 64 / 64 | 20 curated candidates |
| FAA maintenance handbooks | 107.42 MB | 97.301 s | 1,660.470 s | 189 / 189 | 10 curated candidates |
Unique, unmodified files with recorded hashes. No padding or duplicated PDFs. All 277 batches completed with zero recovery requests. Recovery and resume behaviour were tested separately. These PDF runs do not qualify large Excel workbooks or complete diagram interpretation.
Before
The preserved 0.3.0 source/API baseline completed 0 of 3 imports. The subsequent corrected-import pipeline failed all 6 analyses across two runtime arms. No blueprint was exported.
The baseline includes two modules recovered from matching cached bytecode. It is not a fresh test of the released binary.
After
Bounded extraction, checkpoint validation and explicit review now complete the three cases. Unreviewed and stale-preview exports remain blocked. The packaged backend also exported all three reviewed copies.
Prompts, batching and schema changed. Failure durations cannot establish throughput or a speedup caused by the runtime fork.
The model still confuses neighbouring concepts, misbinds descriptions and omits requirements. Only 41 of 1,095 raw candidates were retained in scoped drafts, with 14 structural corrections. The candidate cap was reached in 275 of 277 batches. All prepared text entered inference; that does not mean every requirement was extracted.
The exports are selected catalogues and glossary material, not complete ER models or deployment specifications. All review decisions are labelled automated assistant audit, not human or domain approval. Two parcel maps remain uninterpreted and were explicitly excluded from the nonspatial scope.
GLM-5.3 Effort Max supplied the initial web review. Gemini 4 Argon Medium supplied code reviews; High was not available in the inspected selector. Neither independently ran the benchmark. The final GLM follow-up was service-limited; the final Gemini response was only partly captured.
The local 0.3.2 build 6 is signed and its frozen backend was tested. It has not been newly notarised or published as an installer. The public 0.2.0 preview does not contain these repairs. Clean-machine and minimum-hardware qualification remain open.
Published 10 November 2026. Final review/UI fixes postdate the frozen inference runs; source hashes and runtime identity are included in the measurements. Raw private workspaces and reviewer interface captures are not part of this public bundle.
06 / Earlier public-preview evidence
Historical 0.2.0 preview evidence, retained for traceability. These results do not describe the 0.3.2 engineering candidate or validate its semantic quality.
07 / The preview
PALO Document Workbench 0.2.0 / earlier public preview
This download does not include the 0.3.2 pipeline repairs above. For the supervised earlier preview, download the disk image, drag PALO Workbench to Applications, open it and follow Get ready. Start with the included example.
Controlled pilot, not a general production release. The first full download and explicit cancellation/retry passed on the existing test Mac. Clean-machine installation, minimum hardware and unexpected network-loss recovery remain unqualified. Join the supervised pilot before relying on it for business documents.
08 / Your next mission
We are inviting business users, IT teams and governance practitioners to a controlled pilot. Start with synthetic documents. Challenge the extraction, review and export workflow.
This opens an email to info@paloframework.org. Share your Mac specification and what you want to test. Do not attach company documents, personal data or secrets.
Field notes
It is an installed Mac app. Its native launcher starts a local service and opens the interface in your browser. This is not a hosted document-upload website.
No. This pilot prepares a reviewed development blueprint. Your coding tool and IT delivery process remain responsible for implementation, testing and deployment.
Only if you are authorised to do so. The blueprint can still contain confidential business logic. It is not an anonymisation guarantee or permission to bypass company policy.
No. Notarisation concerns distribution of the software. It does not validate document interpretations, business correctness or compliance.
Describe the step, the expected result and the observed result using a fabricated example. Include the app version and Mac specification. Do not send originals, private logs or sensitive screenshots.
No. These are separate projects with different release and qualification states. See the ChatGPT plugin candidate and Copilot integration pilot for their own evidence.
