# PALO Document Workbench: measured pipeline update

Measured 9 November 2026. Published 10 November 2026. Engineering candidate 0.3.2 build 6; public installer remains 0.2.0.

All three unmodified public PDF repositories completed import, inference and assistant-reviewed blueprint export. This establishes execution on those workloads, not complete or correct business modelling. The candidate is not qualified for unassisted business use.

## Measured runs

| Case | Original MB | Import seconds | Inference seconds | Batches | Raw / mechanical / curated |
| --- | ---: | ---: | ---: | ---: | ---: |
| restoration | 111.95 | 20.151 | 249.585 | 24/24 | 95 / 38 / 11 |
| redevelopment | 163.31 | 175.288 | 619.398 | 64/64 | 252 / 107 / 20 |
| maintenance | 107.42 | 97.301 | 1660.470 | 189/189 | 748 / 277 / 10 |

MB means 1,000,000 bytes. One sequential observation per repository on an Apple M3 Max, 64 GB, macOS 27.2. Maintenance overlapped nine diagnostic inference requests and package builds. Timings exclude source inspection, curation and export review. Review API execution took 3.102 / 33.713 / 143.946 seconds; these are scripted API costs, not human review times. No controlled throughput or fork speedup is established.

## Corrections

The configured limits are 128 MiB per file, 250 MiB per repository and 2,000 pages per PDF, with separate text, unit and worksheet budgets. These runs do not qualify every file within those limits, large Excel repositories or every diagram.

The pipeline now uses bounded fragments with exact evidence IDs, resolves quotes from original text, preserves conflicting interpretations and validates relation endpoints globally. Typed truncated or invalid responses trigger bounded whole-fragment splitting, with validated checkpoints and reuse of completed leaves. Terminal errors block export. All 277 batches completed in these runs with zero recovery requests; recovery behaviour was exercised separately in tests and canaries.

Review supports source-backed corrections to type, name and bindings. Changing or excluding an entity reopens invalid dependent decisions while preserving private history. Export requires all dispositions, source-page scope decisions and a fresh preview digest. Unreviewed and stale-preview requests returned HTTP 400; reviewed exports returned HTTP 200. The UI exposes bounded coverage, assistant identity, reset controls and a graph without excluded or overlapping duplicate candidates.

## Quality and release decision

Of 1,095 raw candidates, 422 passed mechanical checks. Only 41 were retained in the scoped exports, after description changes, 14 structural corrections and 1,054 exclusions. Mechanical validity is not semantic accuracy; an exclusion can mean outside scope rather than factually false.

The model still conflates adjacent concepts, assigns descriptions to the wrong object, promotes headings to entities and omits important requirements. The candidate cap was reached in 275 of 277 batches. All prepared units entered inference, but many supplied no candidate citation. Removing kind-conditioned schema constraints or generating evidence spans first did not reliably solve the observed errors in nine diagnostic requests; those variants were not promoted.

The fixed 27-item reference audit found 1 full raw representation and 2 full curated representations. Partial mentions are counted separately. This small assistant audit is neither precision nor exhaustive recall. It confirms that the selected blueprints are not complete business models. The exports cover a restoration catalogue, a nonspatial planning-action catalogue and a historical maintenance document-register/glossary prototype. They are not operational engineering instructions or deployment approval. Two parcel maps remain uninterpreted and were explicitly excluded from the nonspatial scope.

The previous 0.3.0 source/API baseline failed all three imports. Two recovered bytecode modules were used in that reconstruction, so it is not a fresh test of the released binary. The corrected-import v1 pipeline then failed all six analyses across two runtime arms. The repaired pipeline completes 3/3 cases and exports. Failure duration is not successful throughput; prompts, schema and batching changed, so no causal speedup comparison is justified.

## Validation and review

91 targeted tests passed, with one dependency deprecation warning. The frozen packaged backend exported all three copied reviewed cases and rejected unsupported corrections without changing the preview. Deep strict signature verification passed for local 0.3.2 build 6. That candidate has not been newly notarised or released. Final review/UI fixes postdate frozen inference; the downloadable data includes inference source hashes. Native-launcher, clean-machine and minimum-hardware qualification remain separate.

GLM-5.3 Effort Max supplied an actual web design review; its measured-results follow-up was rejected by the service during peak hours. Gemini 4 Argon Medium supplied code reviews; High was unavailable in the inspected selector. Findings on dependency revalidation, review identity and draft recovery were implemented and checked. Reviewer mistakes about model identity and private fields in exports were corrected. Neither reviewer independently executed the benchmark. A final Gemini reply was only partly captured before the Mac locked. No additional uncensored-model review occurred in this qualification.

The model was Liquid LFM3-1.7B Q4_K_M on the pinned llama.cpp fork 166bd6fef9078476368b2894296625ea0748306f. Runtime changes are not proof of improved semantic extraction. PALO remains a governed proposal/review workflow, not autonomous business approval.

## Evidence files

[JSON measurements](benchmark.json), [CSV measurements](benchmark.csv), [source URLs and SHA-256](sources.json), [item-level reference audit](reference-audit.md), [reference audit JSON](reference-audit.json), [restoration draft](restoration-blueprint.json), [redevelopment draft](redevelopment-blueprint.json), [maintenance draft](maintenance-blueprint.json).

Public source repositories: [FAA handbooks](https://www.faa.gov/regulations_policies/handbooks_manuals/aviation), [Dayton Mall plan](https://miamitownship.com/308/Dayton-Mall-Area-Master-Plan), [University Lakes](https://www.universitylakesproject.org/vision). Original documents are linked, not redistributed in this evidence package.

[Original charts ZIP](charts-originals.zip) preserves exact SVG and PNG bytes. Hosting may recompress or resize individual PNG downloads. PNG hashes in SHA256SUMS refer to the originals in this ZIP.