Data collection schemes die the same way: the form is optional, the foreman is busy, the form loses. One form never loses. People make sure they get paid.

So attach status to the timesheet. To submit hours against a task, answer for the task. Call it extortion. It has a hundred-percent completion rate.

One tap per task · then it pays

The toll has to be a few minutes, so no essays. Done, partial, or blocked — one tap. Partial: most, half, a little. Blocked: the task's own open constraints appear as a list — tap the one, or "new" with a photo. The choices come from the model, so every tap goes back into it.

A few minutes at quitting time, priced into the day — and the schedule updates every night because everyone wanted to get paid. Runs on the company phone.

One sentence in the fire code: this wall stops fire for two hours. Follow it. It becomes a line in the design basis, a wall type on a drawing, a firestop detail, a can of foam and a QC stamp, a commissioning walkdown. Then the project leaves, and for the next forty years electricians drill holes in it.

Every phase retypes the sentence into its own documents and keeps them. By year ten the guy cutting a cable tray through that wall has no idea it's rated. The requirement is still law. Nobody standing at the wall has ever seen it.

One constraint · sixty years

Keep it as one constraint: legislated, designed, built, inspected, commissioned, maintained. Everybody points at it. Nobody retypes it. Then the electrician can ask the wall why he can't drill here and get the real answer, and the regulator can ask which walls this clause owns and when they were last checked. Two good questions, one up and one down. Today both are a records search.

Scaffold inspections, permit renewals, dewatering checks — every Tuesday for a year. The scheduler draws one 200-day bar.

The bar has real jobs: it loads the histogram, holds the work area, collects level-of-effort cost. What it can't do is fail. Never late, never critical, never missed — and this Tuesday's inspection can be.

The bar never fails · the Tuesday does

Late looks like this: the tag expires Thursday at 16:00, Friday three crews have no legal platform, and their timesheets say low productivity. The miss is real — booked to the victims. Derive the instances instead. Each is a constraint with an expiry, so Wednesday you already know which tag lapses tomorrow and who loses their platform. The bar notices nothing.

A crew-hour costs more than a phone, and we still walk data to a trailer and type it in after the crew goes home.

"No phones on site" bans personal phones. Company devices walk through every gate in the industry. The policy problem is already solved. The software problem isn't.

The sensor is already on site

Open questions. What does capture look like with gloves on, in the rain? Who owns the device budget — IT or the project? And what has to be on the home screen to earn the foreman's ten seconds? One answer so far: the timesheet.

STKY — stuff that kills you — is safety's hard-won discipline: attention goes to what seriously injures and kills, not to papercuts. That stays sacred. Borrow it for a second ledger: what kills the project.

The risk matrix has two axes, probability and severity. The third is reversibility. A loss that bleeds over three years has a dozen intervention points; the same loss can turn permanent in one afternoon, with one signature. Where severe meets irreversible, probability stops mattering.

In the corner, probability stops mattering

Plotted that way it stops being a worry list and becomes a controls map: it says where a person has to be standing. Nuclear operations already does this before touching a valve — critical-step identification. Take it upstream, into construction and project decisions. Papercuts still aggregate — that's a different ledger — they just don't get to spend fatal-risk attention.

STKY comes first — attention goes to what kills. But nobody wants the nicks and strains either.

Recording one takes far longer than treating it, so it never gets recorded. Ten seconds, or the data doesn't happen.

Record in ten seconds, or not at all

Then trend: nicks against the blade policy, strains by area, week over week. The worst class gets an owner and a control change. Papercuts don't predict deaths. They will happen twice for the same reason.

The plan prices labor as a rate. A crew that's poured together all season beats one assembled Monday — same trade, same rate, different job.

The difference lives in the crew: the operator who knows the rigger's hands, the finisher nobody has to sequence. Schedule capital, recorded nowhere — so the leveling run treats six welders as six interchangeable pegs and spends it for free.

Crews have memory

Splitting a good crew is a schedule decision that appears in no schedule. Names matter — and some come in sets.

The only project risk with a free, public, daily probabilistic forecast — and plans still treat a January pour like July with a note attached.

Two free feeds, both ignored. Thirty years of records say how many pour days February actually holds: plan on February's distribution, not February's average. And the ten-day forecast is a live constraint feed — when Thursday crosses the wind limit, the lift's window just moved, and the model should re-derive like it would for any constraint expiring.

Free forecast · unused

Instead the forecast lives on the super's phone, and the schedule finds out in the delay notice.

You can't list them. You can count them arriving: new constraints surfaced this week — the discovery rate.

It should fall as the job learns. The trap: a job that stops surprising you is done learning, or done looking — and the falling line looks the same either way. So count the looking too: walkdowns, assumption rechecks, questions asked. Discoveries per look is the honest number.

Discoveries per look

Discoveries falling while the looking holds steady: the fog is burning off, and the window should be narrowing to match. Both falling together: nobody's learning anything — the surprises are saving themselves up.

Six systems know the spool: fabricated, shipped, cleared, received, staged, installed. The crew's question takes three phone calls.

A barcode solved this in 1994

Couriers solved this in 1994 — a barcode, a scan at every handoff. A crew asking "where" is a report that the model is missing a state.

Boxes without dates, arrows without logic, owners nowhere. It can't be queried, so it can't be wrong — which is why it's popular.

Can’t be queried · can’t be wrong

A picture that can't disagree with reality is decoration. Draw the picture from the model, so it's a view — not a second set of books.

Every claim and every RFI is an argument about what the workface looked like on a date. Thirty seconds of walkthrough settles it before it starts.

The cheapest evidence ever made

The camera is already on site. Attach the clip to the constraint it proves and the dispute file assembles itself while the job runs.

A hold-point signature costs a drive, a binder, wet ink. The friction teaches people to batch and skip. Expensive to give, cheap in meaning.

Cheap to give · heavy to mean

Invert it: five seconds on a phone, identity attached forever. Free logistics, total accountability.

Authority lives in heads and out-of-office replies. When the signer flies out, the job learns who can't sign what, one stopped hold point at a time.

An edge with an expiry

Delegation is an edge in the authority graph with an expiry. From, to, scope, until Friday. Vacations stop being schedule events.

A procedure in a PDF is advice. What crews run is a program — steps, conditions, hold points, quals — with people as the runtime.

Diff a procedure · not a binder

Written as atoms, every run instantiates trackable constraints: versioned, diffable, auditable. Change the procedure once; every future run changes with it.

Status meetings ask what happened. The job moves on which constraints, killed today, release the most work?

That's a ranking, and nothing on site ranks it. Lists sort by date raised or by owner — never by hours released.

Same sun · different lock every day

A good superintendent runs the ranking in their head at 05:30 and is usually right. Build that, shared. Nothing smarter than the super — just always awake.

Run the same schedule through a thousand weathers and deliveries and a different path goes critical in a third of the runs. The red line is one draw.

And it moves every update, so crews get whipsawed toward whatever went red this month. Critical in sixty of a hundred futures deserves attention every month — whichever path the printout crowned.

Critical, most mornings

Float answers a question about one run. Criticality answers a question about the job.

Ground conditions, vendor data, "the permit will renew" — logged at sanction, reread never. An assumption is a constraint someone chose not to verify yet. It fails like one: without warning, late, where change is dearest.

Unwatched · not untrue yet

They age at different speeds — soil over years, a vendor's shop loading week to week. Give each an owner and a recheck date scaled to its half-life. A register that can ring.

The baseline is what we believed on sanction day. Two winters and forty change orders later, we still measure against it and call the gap variance.

Rebaselining reads as confession, so it comes late and political — and every month in between, the reports measure distance to a plan nobody believes.

A photo ages · the job moves

Keep every forecast; the baseline is just the one with a contract stapled to it. Score the series. Grade the forecaster. Rebaselining becomes the next entry.

Five thousand activities in the schedule; the job runs on a whiteboard, three weeks deep, updated at 06:00. Both are right — the contract needs memory, the crew needs Tuesday.

The rot is quiet: the whiteboard becomes the real plan and never writes back. Within a season the schedule describes a different job than the one being built.

The wall wins

Two zoom levels of one constraint set — and marks on the wall write back. The whiteboard stops being a fork and becomes the interface.

A predecessor tie. A missing spool. An unsigned permit. A crane booked elsewhere. Forecast wind over the lift limit. A room too crowded to work in. A dose budget nearly spent. We manage these in seven different systems with seven different owners, and they are the same object: a condition that must be true before work can proceed.

CPM made one constraint type — logic ties — the aristocracy, and demoted the rest to footnotes. Not because logic matters most, but because it was the type a 1960s computer could hold. Sixty years later the field still runs on that accident. Ask any superintendent what actually stopped work this week; it is almost never a predecessor.

Seven costumes · one object

The schedule is a shadow

If everything is a constraint, then the schedule isn't the plan — it's the shadow the constraint set casts on a calendar. Dates are outputs. Managing the dates instead of the constraints is managing the shadow and wondering why it won't hold still.

One object also means one grammar: what it blocks, what satisfies it, who owns it, when it expires. A permit and a weather window differ only in their satisfaction function. Put them in one set and you can finally ask the only question that matters each morning — which constraints, if I killed them today, release the most work?

Whether one engine can hold all seven types honestly, I'm still finding out. But I've stopped believing the seven-systems version is anything other than a coordination tax we pay for a hardware limit that expired before I was born.

A pour needs three dry days. The crane leaves in October. In-water work ends when the fish return. None of it cares about logic ties.

Windows break float math. Forty days of float in front of a window that closes in forty-one is one day of float — and no tool on site will say so. Miss it and lateness becomes a season.

Miss the window · pay a season

A window is a constraint with an expiry, not a note in a drawer. The cliff-edged ones go on the board every morning.

The schedule says 2 × WELDER, CLASS B. The job needs Dave — the only live qual for that alloy — and Dave is at his daughter's wedding until Tuesday.

The histogram shows welders at sixty percent all month. True, and useless. Aggregates hide singletons; singletons stop jobs.

Roles on paper · people on jobs

Names where names matter: quals, signature authority, the tandem-lift operator. Constraints with calendars — and the wedding lands in the plan back in March, when it costs nothing.

Every foreman writes the story every day: night pour ran long, pump swap at 03:10, watch form 12-B. Then five systems demand the same facts retyped — schedule update, timesheet, RFI, QC record, claim file.

The fix is structure at capture, not software reading diaries. ACE — atomic constraint entry: what it blocks, what clears it, who owns it, when it expires. Told once in that shape, the update, the RFI and the claim file become queries.

Written once · projected five times

Claims teams already rebuild those four fields from prose, years later, at lawyer rates. That's the invoice for capturing stories without structure.

The estimate has a breakdown. So do the schedule, cost control, and the work packages. Identical at sanction; four dialects by month six, plus a standing reconciliation meeting.

Each copy exists because each group needs a different shape — cost by contract, field by area, schedule by system. Those are views. The mistake is materializing views into owned copies.

Ribs argue · the spine decides

One number, three costumes. Reconciliation was never work — it was rent on the copies.

Anyone can turn a cell green. Free, so it drifts optimistic — green is often the color of nobody wanting the argument.

The field solved this long ago: the weld is good because a stamp says so, and the stamp follows the welder for a career. Reporting never adopted its own industry's best idea.

The check is free · the signature isn’t

Sign the greens. Fewer, heavier, worth reading.

SPI and CPI score effort against the plan. Nothing scores the forecast. The owner is buying a date; nobody measures whether we know it. So reports stay green while the job gets later.

Measure the fog

At any status date the finish is a P10–P90 window. A month that poured nothing but closed the question smearing turnover across eight months bought more schedule than a record month of quantities — and no metric noticed. So baseline how fast the window should shrink, and track the shrink. That might be the real progress. Might.

SCI and CCI

SPI across · SCI up

The circled corner is the one that worries me: quantities moving, fog not lifting. Today that manager probably gets promoted.

The anti-gaming rule

Left alone, everyone claims a narrow window. So claimed windows get scored against realized dates, the way weather forecasters are graded — overconfidence should cost more than admitted doubt. Which means a window that widens honestly can be the best number on the report. Whether that survives contact with a steering committee, I don't know.

A year of window width is a year of financed capital, and contingency is the posted price of not knowing — so this is probably worth real money. I won't defend a number. The engines already exist: every big job runs a schedule risk analysis and throws the distribution away. Keeping it is a cheap experiment.