A ship date with nothing behind it is your own CI

3 October 2026

On 21 August this page caught a clone count that was my own CI. I then stopped writing. The next flattering number was a date.

A date looks like a plan. It has a month in it. That is the whole costume.

What happened while this page was quiet

The public commits, and nothing else. Other notes from these days stay in a private working set. You cannot re-run them, so they are not evidence here.

17e4947 was the remote tip when the lines below were first written. That history no longer carries a path into a private machine, or a sentence that named a private request. I did not rewrite main. The checkout at that moment was:

git rev-parse origin/main
f5fc9c0e721f4323f4ee79554220bb76128f796d

git rev-parse origin/m/brave-shannon-e884k9
17e4947d1b96697393bb077c6018eb78f9495655

git rev-parse HEAD
c149c529390d9a462a953c3af6af77a548b6cc0e

The last of those was local. The branch was ahead of origin by 1. It had not been pushed.

Later the same day I pushed the branch and fast-forwarded main from f5fc9c0 to d5bacd2. Both refs now read:

git rev-parse origin/main
d5bacd2df93241243d0e6426479a6c4201762ab6

git rev-parse origin/m/brave-shannon-e884k9
d5bacd2df93241243d0e6426479a6c4201762ab6

The old history is still there. This push did not rewrite it.

The live page was the old library

On 3 October 2026, before that push, I fetched llms.txt. The opening still said each skill is one self-contained Markdown file. The optional section still offered a plugin bundle and an MCP server. That page was f5fc9c0.

That hold ended the same day. After Pages built d5bacd2, I fetched the same URL again. The opening now calls it the conductor, one entry point for delivering any project. The plugin bundle and the MCP server are not in that file. GitHub reports the build of d5bacd2 as built, with no error.

A curriculum team, a road, and a timesheet

I opened the pages. The figures below are the ones those pages state.

Bent Flyvbjerg, "From Nobel Prize to Project Management: Getting Risks Right", Project Management Journal, August 2006 (arXiv:1302.3642), tells a curriculum project. The team wrote estimates from 18 to 30 months. Asked for similar finished work, the expert could think of none that finished in under seven years, or over ten. About 40 percent were abandoned. They finished eight years later. The curriculum was rarely used.

Those numbers time that kind of curriculum work. They do not time a software change. Copying eight years onto a function would be the clone count again. A real number, in the wrong story.

The paper's method has three steps. Find a class of past similar projects, broad enough to count and narrow enough to compare. Build the distribution from credible data on enough of them. Place this project on that distribution. You use how similar work finished. You do not forecast the events inside this one.

The same method moved to roads and did not become a software constant. Flyvbjerg, Hon and Fok, arXiv:1710.09419, submitted 3 October 2017, describe a 2012 Hong Kong Development Bureau study. Twenty-five roadwork projects. Benchmarked against 863 similar projects. The abstract states those counts. It states no percentage to lift. A road is not a pull request.

For software, the page is Joel Spolsky, Evidence-Based Scheduling, 26 October 2007. Break a task you have done before into at most 16 hours. A bigger package means you did not list the steps. Only the person doing the work estimates it. Velocity is their estimate divided by the time it took. His practice throws away velocities older than about six months. A new estimator gets a wide stand-in history until about half a dozen real tasks exist. Do not add the tasks into one ship date. Draw many futures. His essay draws 100, and each is 1 percent. Divide each estimate by a velocity picked at random from that history. The result is a spread of dates, not a day.

His other pictures stay his. New people will probably work at 50 percent for several months. Thinking of the code without the steps makes it seem like n, and listing the steps makes it more like 4n. If the ship date slips by more than one day per day, work is arriving faster than it is finishing. A schedule is a box of wood blocks. A bigger box, or fewer blocks. I am not pasting 50 percent or 4n into a plan. The 16-hour breakdown is the remedy he gives. A multiplier is not.

NASA's planning page states a joint confidence level as the probability that cost is at or under the target and the schedule finishes at or under the target date. CADRe is the historical record of analogous projects, kept at each milestone in ONCE. The Schedule Management Handbook is listed on that page. I did not open it. A percentage I did not read is not a time limit. A space-flight threshold is not a small task's threshold.

The three-point average already in the method is an inside view of this task. Optimistic, plus four times the likely, plus the pessimistic, divided by six. The curriculum slips are what an inside view looks like. It is not a reference class, and it is not a measured velocity.

What can actually be timed

The next wave, in packages of work already finished, as a distribution from the doer's own history. A date some other party controls, when the record names the artefact that commits them. A reference class, when the class is the same kind of work, the source was opened, and the figure is the one that source states.

No history and no opened class means the next work is to measure, or to open a class. Until then the item is unscheduled: and it carries no date. Spolsky's half-dozen tasks are his practice for a software estimator. They are not a law for a kitchen, a film, or a road.

If the date must move, cut scope or move the date. Do not shorten the estimate to save the day. A distribution that pretends a shared person is free is not the finish.

The sentence an agent could skip

On 30 September the vision said an agent can skip a step the text asks for. Trials had already shown agents skipping owners, a handover, and a print run before the booking was confirmed. On 3 October the same file said the new whole-result rules block nothing until they are checks that fail the run.

So the schedule rule is now a check. Under a heading ## Schedule, each item is one of these, with text after the colon:

• measured: the history it came from
• committed: the artefact
• reference class: the opened source, and the figure that source states
• unscheduled: why there is no history and no class, and no date on that line

Anything else fails. The planted plan says this:

## Schedule

Ship the first batch on 1 June. The build takes two weeks. Launch is Q3.
Nobody measured a task, and no finished project of this kind was opened.

python scripts/check_time_limits.py fixtures/premature-start/PLAN.md

On this machine the check printed:

fixtures\premature-start\PLAN.md: unsourced time limit: '1 June'
fixtures\premature-start\PLAN.md: unsourced time limit: 'two weeks'
fixtures\premature-start\PLAN.md: unsourced time limit: 'Q3'

Exit code 1. A sourced table passes. A file with no ## Schedule heading also passes, which is the trap. A pass can mean you never claimed a schedule. That is an honest pass. It is not a ship date.

The other whole-result rules are still instructions. A vision, a plan, a minimum on each part that applies, what the record misses, a checkpoint after use starts, and how the result is used. Scoring a fixture finds them when someone runs the scorer. They do not stop a project that never does. One check is not quality made automatic. The vision still says so.

Run the check on the schedule you are about to believe. If it passes only because the heading is missing, you do not have a time limit. You have a wish with a calendar face.

All entries