LOCAL AI / PRACTICAL TASKS
Compare your selected computer and settings with retained task runs. Download a public sample, run its original verifier and use version-scoped troubleshooting when it fails.
Invoice extraction and cited PDF search have narrow, dated Mac evidence. Three coding attempts failed the unchanged public fixture. The final prompt supplied implementation guidance; coding remains unqualified and the request budget is exhausted. Existing corrected-note kits remain available.
OWN INPUT / PILOT SOFTWARE
The retained task cards below keep their dated public sample evidence. The new own-input extension adds local source review and JSON/CSV export; fresh own-input model qualification and reader recovery remain unestablished.
These are declared settings. This page does not scan or test your computer. Defaults display the retained Mac setup so you can inspect its evidence; change every value that differs.
Exact CPU memory fit: insufficient information. Low/high allocation, headroom and capacity reserves are unavailable for this CPU task configuration in the retained shared estimator. Process RSS in a task receipt is not a capacity requirement. No fit or speed verdict is inferred from the file size or a task pass.
The existing shared estimator covers other supported execution settings. Open hardware/model planning · Local Lab estimator and sources. Changing context or model requires recalculation there and task requalification here.
Turn a small text invoice into four typed fields, preserving a missing bank account as null.
One public synthetic sample passed on 2026-10-28 UTC; 4.181 s answer time and 5.912 s startup. This is dated shared-machine observation, not a general quality or speed estimate.
Download retained working sample kit ↓ · Check the public sample locally
ZIP 64,591 bytes. SHA-256 51a339ef686c7f064327d6396d6d8ca45361e42044cfa47042b8b13f62ae9eaf
Find an explicit loan policy in two small public synthetic PDFs and retain its source page and quote.
One public synthetic sample passed on 2026-10-28 UTC; 80.827 s answer time and 3.576 s startup. This is dated shared-machine observation, not a general quality or speed estimate.
Download retained working sample kit ↓ · Check the public sample locally
ZIP 140,225 bytes. SHA-256 fdc4a0ade2c8880cd6480d345ede0cff780aad95a81ff4b194db5fe678b09b3f
One public Python fixture and seven unchanged tests; no general coding ability claim.
Three actual model attempts failed. The .2 baseline rejected neither boolean; the .3 correction also overbilled exact minutes. On October 31 UTC, the .4 supplied-guidance attempt failed True and False rejection; independent replay also failed. The three-request budget is exhausted. No further automatic model tests.
Download coding fixture and retained failures (unqualified) ↓ · Inspect coding manifest
ZIP 27,049 bytes. SHA-256 bac9cff4f68362f0aba3cae1553b73228e76aec44a3e11826398400826cdd094
The .2 baseline, .3 failure, and .4 failure are separate observed runs. The .4 prompt explicitly supplied exact type and integer-ceiling guidance after reviewing failures. This is not an independent coding competence test. The original source and seven tests remain unchanged.
A separate authored control passed a replay, but is not model output or a reproduced model repair. The ZIP retains the model responses and labeled projections.
These existing Fix Finder records retain their own exact task and settings. A matching runtime is a reference, not a reproduced repair for your task. Keep the original failure; apply only a supported correction; recalculate changed memory settings; rerun the unchanged verifier. No log text or commands are uploaded or executed by this page.
Return to the invoice verifier → · Browse existing Fix Finder diagnostics
Journey registry 20261102.1 retains invoice and PDF task runs dated October 28 UTC and coding failures dated October 30–31 UTC. No measurement date is refreshed by packaging, a metadata check or replaying an output. New runtime metadata cannot inherit a task pass.
Versioned task registry (JSON) · Exact pinned metadata and package health. The health report covers source and package integrity only; it does not qualify this machine or a current runtime.
Setup links contain public configuration fields only. No accounts, uploads or per-visitor AI calls. No telemetry is added. A download is not evidence that the reader ran the task.