Skip to content

ProjectLens: use it on your own schedules and updates - #27

Merged
MatthewPaver merged 20 commits into
mainfrom
feat/lift-and-shift
Nov 1, 2026
Merged

MatthewPaver merged 20 commits into
mainfrom
feat/lift-and-shift

Conversation

@MatthewPaver

@MatthewPaver MatthewPaver commented Oct 30, 2026 •

Copy link
Copy Markdown
Owner

What this does

ProjectLens can now be used on other people's real schedules and progress updates, not only the built-in example.

  • One engine: the schedule and change-assurance pages share one engine (docs/engine/projectlens-engine.js), and the GMPP pipeline calls it too, instead of keeping a Python copy of the rules.
  • More formats: Primavera P6 XER and XML, and MS Project XML. Word narratives are read in the browser. Windows-1252 exports and multi-project XERs are handled.
  • Renamed schedules: renamed or re-numbered updates can be compared after the user confirms they're the same project (matched on activity codes), instead of being refused.
  • Narrative check: it reads stability claims and dates from the written update, tells report dates, last year's date and milestone dates apart from the current end date, and separates "check the published date" from contradictions.
  • Why it was flagged (round 5): each finding shows the sentence, the matched phrase and both dates.
  • Command line and monthly batch: node bin/projectlens.mjs compare --previous a.xer --current b.xer --narrative update.txt and batch --dir monthly/ (CSV history). Exit code 1 means a contradiction was found.
  • Guide: docs/USE_ON_YOUR_OWN_SCHEDULES.md.

A simplicity pass removed the duplicate Python rules and dead helpers: about 320 fewer lines with identical results. CI now sets up Node 22 for the pipeline.

How it was checked

  • npm test (142), PYTHON=python3.12 make test (53) and the browser tests pass.
  • A fresh clone works by following the guide: compare and batch on the worked example.
  • npm run bench scores five sets of real UK Government Major Projects Portfolio year-on-year pairs (Open Government Licence; sources in bench/real-cases/SOURCES.md). It fails if any set drops below its floor.

Honest accuracy

Set Score Contradictions caught Explained moves left alone Note
1 32/32 8/8 14/14 Seen during development
2 29/35 10/13 11/12 Seen during development
3 28/35 8/13 10/10 Seen during development
4 28/30 7/8 8/9 Unseen when first scored (22/30), then used to guide round 5
5 29/40 3/10 17/18 Unseen when scored. This is the headline number.

The narrative check rarely raises a false alarm (17 of 18 explained moves and 8 of 8 unmoved projects left alone in set 5), but on unseen narratives it catches only 3 of 10 contradictions. The rules fit the cases they were built from much better than new ones, so treat it as a prompt to read, not a detector. Schedule comparison itself (dates, float, constraints, links) is reliable on the real P6 files tested.

Limits

  • MPP binary files and PDF narratives aren't supported; export to XML, or paste the text.
  • The narrative check is rule-based.
  • P6 already compares schedule versions, and Deltek tools offer far broader diagnostics. Deltek also has narrative assessment in PM Compass/Cobra. ProjectLens's addition is a local check of the written update against the schedule's dates, with a human decision record. Treat it as a review aid.

🤖 Generated with Claude Code

MatthewPaver and others added 20 commits October 29, 2026 18:36
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…d work)

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…guide for your own schedules

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
…finish dates

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Adds three sets of real GMPP year-on-year pairs (Open Government Licence,
sources in bench/real-cases/SOURCES.md). npm run bench scores the narrative
check per label, lists every miss, and fails if a set drops below its floor.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y shell

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d finding sentences (round 4)

Programme-level on-track claims count again, month and quarter end-date claims
feed the published-date check, and the CLI, GMPP page and change assurance
share the same finding sentences. A follow-up pass restored every explained
and no-move result. Real-case bench: set 2 27/35 (contradictions 8/13, was 2),
set 3 22/35 (contradictions 5/13, was 2), no new false alarms. Implemented by
Codex; verified with npm test (108), make test (53), browser tests and the bench.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ntences

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ore on unseen set 4

The GMPP pipeline now calls the shared JavaScript engine instead of keeping
its own Python copy of the claim rules (CI provisions Node 22), and dead
helpers are removed: about 323 fewer lines of code with identical results.
Adds the unseen real-case set 4 to the benchmark (22/30) and freezes floors at
32, 27, 22, 22. Done by Codex; verified with npm test (108), make test (53)
and npm run bench.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each contradiction or published-date check now carries the sentence, the
matched phrase and both dates, shown in the CLI, the change-assurance page and
the GMPP page. Bench: sets 1-4 at 32, 29, 28, 28 with no fall in explained or
no-move; unseen set 5 scores 29/40 (contradictions 3/10). Built by Codex;
verified with npm test (142), make test (53), browser tests and the bench.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@MatthewPaver
MatthewPaver merged commit f9c2812 into main Nov 1, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant