Skip to content

History

Revisions

  • Move to DiGi-Organization: update owner links and workspace paths The DiGi repositories were transferred from ZiolkowskiJakub to DiGi-Organization (client projects to DigiProject-Organization) and the local workspace folder was renamed from DigiProject to DiGi. - Owner replaced in links, issue references and sync-wiki.yml (old URLs keep redirecting). - Workspace paths DigiProject/... -> DiGi/... in docs and comments. - Regenerated .agents and the README guideline block, which also carries the pending Coding - YOLO.md guideline updates. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Oct 14, 2026
    9909cfc
  • ConsoleApp guide, Home and detector training: retire the regressor narrative Step 5 is detector-only dating: the reference page's options, preflight, exit codes and gotchas now describe the first-detection rule, the detection-only feature guard and #28 clearing; the Home step list names the dating rule; the model.pt rationale becomes the county rewrite requirement (#23) rather than a paired regressor. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Oct 9, 2026
    07cd63f
  • Running a County: document detector-only dating, clearing and the tray recipe Dating is now first-confident-detection@0.5, not the regressor: the guard requires the detection group alone, the contract check reads what the predictor states, never-detected buildings have their stored year cleared (#28), the tray task forces the six step flags, and the shipped identity table carries the dating rule beside the detector SHA-256. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Oct 9, 2026
    127c993
  • train9_fresh and the 2026-10-08 regressor are the shipped pair (DiGi.GIS.YOLO.UI#12, DiGi.GIS.ML#13) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Oct 8, 2026
    e9a93da
  • Training the Detector: fresh model and continued training, with the data to keep (ZiolkowskiJakub/DiGi.GIS.YOLO.UI#12) New guide page: what a --train run is, prerequisites, training a fresh model from yolo26x.pt (new or existing dataset), continuing an existing model (continue vs resume, the data required to continue, appending new labels), interrupted runs and automatic resume, the gate and the staged roll-out, and the traps seen in train9. Complete option files for each case. Linked from Home and Running a County section 12. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Oct 7, 2026
    f34637e
  • Running a County: training dataset, label check and detector evaluation (#13) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 29, 2026
    286ff03
  • Running a County: step 3 and the verification table now cover the three year-built columns (predicted/user/calculated) and state the calculated fallback rule (DiGi.GIS.PostgreSQL#94)

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 24, 2026
    8258e83
  • Feature-contract refusal: Years/Radiuses are now checked against the model, narrowed projections are refused (exit 1) (ZiolkowskiJakub/DiGi.GIS.ML#6)

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 23, 2026
    5c91afa
  • Correct the recipe against a real run, and record the deployed layout (#7) #7 §4 required Running-a-County to be re-read against what a run actually does rather than amended in passing. Three runs over county 104106 - one from the deployed tray application, two completing after the ModelPath fix - contradicted five things this wiki asserted. Building Data updated is 11076 on a combined run, not 5538. The published sample could only have come from a run with one of the two write steps on: the counter accumulates once for the detection columns and once for the predicted year built column, so a run with both counts every building twice. Nothing but watching a real run would have shown this. "The committed template ships with all three write flags off" was true of the combined template only. Score.json ships UpdateYearBuiltData and UpdatePredictedYearBuilt both true, Detections.json ships UpdateDetections true, and §6 hands both to the reader as recovery templates. The runner's own folder is not the same place on a deployed machine - it is extensions\DiGi.GIS.YOLO.UI.ConsoleApp inside the tray application - and CopyUserFiles shadows the committed options file there with the build machine's own, so the file beside a deployed runner is not the default it looks like. A tray run has no options file, so the members absent from the dialog come from the class defaults rather than from a file. That makes those defaults part of the tray contract, which is why a null ModelPath broke a run rather than inconveniencing one. Trap 7 did not say what the tray application resolves a relative scratch path against: its own folder, which on a deployed host is the deployment directory, so a bare "scratch" exports tens of thousands of images into the installed application and whatever syncs it. New trap 10 records that a machine holding both a deployment and a checkout runs two tray applications with the same icon - which cost two runs here before it was noticed. Exit code 2 now names the detector weights alongside the interpreter and the .mlnet model, all three refused before a single image is exported. Home gains the two run surfaces and the opt-in extensions hosting, which no page mentioned. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 8, 2026
    3500e89
  • Fix #9: the committed options defaults now carry their clean names

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 8, 2026
    6495290
  • Running a County: the tray prompt settles scope, and every tray run writes The tray path described a dialog that no longer exists. Predict year built now asks for the counties, the scratch directory, the interpreter, max concurrent requests and the scratch cleanup - five controls - and runs the full six step flow every time, so there is no read-only tray run any more. The dry run and the two split recovery passes are console app work, with the templates below. Fourteen members are settled in the options file rather than the prompt, and the table says which and why. Confidence and ModelPath join Years and Radiuses as things that change what the regressor is handed without failing (ZiolkowskiJakub/DiGi.GIS.ML#6). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 4, 2026
    556f743
  • Say what the Predict year built prompt settles and what still needs an options file

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 4, 2026
    bc9bf7e
  • Rewrite the recipe around a single run, and document the cleanup Running-a-County presented the detections and scoring passes as two runs whose ordering was explicit by design. They are one run now: the scoring pass rebuilt its building list from results.bbrf on disk rather than from the detection columns already stored, so a scratch directory cleaned between them skipped the county reporting exactly what a county with no detections reports. Section 2 carries that reasoning, the order table drops to three steps, and the split passes move to their own section as the manual recovery path - with the warning that their templates set CleanScratchDirectory false, without which they cannot hand off at all. Section 5 states what the shipped template actually does: it carries every write flag off, which on a fresh county is refused by the feature coverage guard, so it is not the configuration that populates one. The exit code 5 row, trap 8 and the Resume warning now say who gets cleaned - a county that completed, and a cancelled one, but not a county that failed. Using-the-ConsoleApp gains the option, the Resume interaction, and a gotcha for the missing results file now being a failure rather than an empty county. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 4, 2026
    c8b5a26
  • Fix broken Dependencies links in wiki syntax The double-bracket wiki link needs the pipe and URL inside the closing brackets ([[Name|url]]); [[Name]|url] rendered as literal broken text. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 4, 2026
    f0e742f
  • docs: add the standing run recipe and correct the ConsoleApp guide Running-a-County.md is the operator entry point: why the run order exists, the county readiness check, the two console passes, verification and exit-code triage. Corrections to Using-the-ConsoleApp.md that the feature-coverage guard invalidated: a read-only dry run is the second pass, not the first, because scoring a county whose detection columns were never written is now refused with exit code 4; Years is a Range<int>, not a List<int>; Radiuses was undocumented. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 4, 2026
    2c46600
  • docs: add wiki landing page and ConsoleApp usage guide

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 2, 2026
    d48f28f
  • Initial Home page

    @ZiolkowskiJakub ZiolkowskiJakub committed Sep 2, 2026
    37fffe7