Skip to content
Results · one product, measured

Everything below comes from one build. Ours.

One operator and twelve specialist roles, on a live cybersecurity product. Position as at 2 August 2026, recomputed from the work record — not a snapshot anyone typed in.

These are a dated capture, not a live feed. The board they come from keeps running, and the figures move with it — up and down.

Read this first

This is a case study, not a track record. Every figure on this page comes from a single product that we built for ourselves, which means we also wrote the requirements and set the deadlines. That is an easier problem than a client engagement, and you should discount accordingly. The full case study says what it does and does not prove.

98%
Work completed
1,355 of 1,381 items closed
182/184
Features delivered
99% of the planned programme
2.64
Defects per 1,000 lines
every line shipped — within the 1–5 published for a disciplined team
17 days
To reach this scope
against a 9–18 month norm — see /results for the caveat
80.6
Quality score
coverage, checks passed and defects, combined
10.9 h
Active work per item
time actually spent, not elapsed
87%
Steps with no human
of the process runs unattended
165,000
Lines of working software
the product these numbers come from
The factory's own board showing issues closed, work packages completed, total lines, quality index, defect density and gate pass rate.
The board these figures are read from — not a slide rebuilt for this page, and captured on 2 August 2026 so every number above can be checked against it. The cost row is cropped: what a delivered item costs us to produce is not a price, and publishing it next to no price invites the wrong comparison.
What the boards say now, and some of it is worse

The position on 28 October 2026, after the AI SDLC renovation.

The capture above is from 2 August 2026. This is the same two boards three months on, read straight from their scorecards that day: the product, and the factory itself. Both stay in, because a board that can only show good news tells you nothing on the day something goes wrong.

  • Defects per 1,000 lines: 3.41, up from 2.64. We read this as the checks catching more issues, and catching them early in development, where they are cheapest to fix. A smaller line count after a refactor also raises it.
  • Work later redone: 43% as at 2 August 2026, a watch item with no published norm to compare against. The scorecard does not carry it, so there is no newer figure.
  • Queueing: time lost rose as more work ran in parallel. That is a direction, not a magnitude we would defend.
The position on 28 October 2026, by scope
ScopeMeasureValueRead
The productIssues closed1,391 of 1,415 (98%)
The productWork packages completed178 of 179
The productLines of code136,631Down from 165,000 after a refactor
The productQuality index77.3Down from 80.6
The productDefects per 1,000 lines3.41Up from 2.64. We read it as the checks catching more issues, early in development
The productGate pass rate89.1%
The productTime actually spent workingWithdrawn (was 3% of elapsed)Computed from two different groups of tickets. See the note below the table
The factory itselfIssues closed233 of 254 (92%)
The factory itselfLines of code92,435
The factory itselfQuality index83.2
The factory itselfDefects per 1,000 lines1.17A smaller and newer codebase
The factory itselfTime actually spent workingWithdrawn (was 2% of elapsed)Computed from two different groups of tickets. See the note below the table
Claude's working timeDays Claude worked51 of 57 days5 September to 31 October 2026. The logs go back no further, so nothing earlier can be measured this way
Claude's working timeWorking timeabout 100 hours; median 1.4 hours on a working day82 to 103 hours, depending on how long a pause still counts as working. Parallel sessions are counted once
Claude's working timeWorking time on a branch that names a ticketabout a fifth (20% to 22%)The rest ran on the website, the factory home and other work no ticket names, so a ticket board sees about a fifth of it
Claude's working timeShare of a ticket's working span spent working21% (median ticket; 11% to 21% over the thresholds)74 tickets with Claude time on their own branch, first to last Claude activity. Middle half of tickets 2% to 100%; 0.12% of the time from creation to resolution
Claude's working timeShare of a working session spent working68% (52% to 80%)The rest is a person reading and typing. The range is how long a pause counts
The renovationEpics closedAll ten (E0 to E9) and the pilot epic
The renovationDogfood findings logged6030 fixed, 14 design changes, 18 ticketed
The renovation pilotManual interventions for one requirement126 fixed in the run, 4 fixed the next day (ODSF-264 to 267), 2 human gates by design. Not re-run since
Withdrawn on 31 October 2026: “time actually spent working”

We published 28% on 2 August 2026, and 3% (the product) and 2% (the factory) on 28 October 2026. The scorecard divided the average time a resolved ticket took by that time plus the average age of the tickets still open. They are different tickets. On 28 October the “idle” side was the average age of about two dozen open tickets, not time spent waiting inside the resolved ones, and it grows by itself as a backlog ages. It said nothing about how much time anyone or anything spent working.

The ticket system records two times per ticket, created and resolved, so it cannot say how long a ticket was worked on or waiting. A correct per-ticket figure needs the state changes recorded. The rows headed Claude’s working time are read from Claude’s own logs on 31 October 2026.

The same rows are in Appendix A.2 of the AI SDLC Playbook. The pilot row is the pilot report.

How to read this record

  1. Every number has a date and a scope. The capture is 2 August 2026; the position is 28 October 2026. Never compare a figure from one with a figure from the other without saying so.
  2. Where they disagree, both are printed. A dashboard that only goes up is a marketing asset. We wanted an instrument, and an instrument has to be able to read low.
  3. There is no second version. The board a client sees is the one we run on. The moment a metric exists to be presented rather than read, the real numbers migrate into private spreadsheets.
  4. Ask any supplier for the number that got worse. If every figure they give you flatters them, discount all of them, because you cannot tell which ones deserved it.
  5. None of this makes the work cheaper by itself. It makes the work legible, and legible work is the only kind you can improve.
Efficiency

We can tell you what a change costs before you commit to it.

Cost is metered per item, per feature and per specialist, on the same board that reports progress. That is unusual enough to be worth saying plainly — and it changes what a scope conversation can be about.

Metered per item, not estimated

We can tell you what a delivered item cost to produce, because the price of the work and the proof of the work come out of one record. Most suppliers can give you a day rate multiplied by a guess.

Rework is its own line

Measured and shown separately rather than absorbed into the total. A rework figure you cannot see is a rework figure nobody manages.

Scope becomes a value question

When the unit cost of delivery is visible, you can ask whether a change is worth what it costs to produce — a better question than whether it is worth three developer-days.
Comparison

Against what comparable work normally takes

MeasureThis factoryTypical for comparable workRead
Work delivered as committed94–96%60–85%Better
Test coverage80%70–80% targetIn line
Defects per 1,000 lines2.641–5 for a disciplined teamIn line
Effort to reach this scope≈ 1 person-month~200 person-monthsFar lower
Calendar time to reach it17 days9–18 monthsFar lower
Work later redone43%no published normWatch item

"Typical" means commonly published software-engineering ranges for work of comparable scope. The effort and calendar rows are an order-of-magnitude estimate, not a precise measurement — shown with that caveat rather than without it.

A row for time actually spent working (28%, read as better than the typical 10 to 20%) was in this table until 31 October 2026. It is withdrawn: the figure did not measure that. Why.

Proof

It is not built around one project.

The factory and its first product grew up together. The third product — an unrelated business, on an unrelated technology — is the evidence it transfers.

A cybersecurity platform

A live product analysing threat data for security teams. The build these numbers come from.
165,000 lines

The factory itself

It runs on its own line and passes its own checks. If it did not work, this is what would break first.
Proves it on itself

An education platform

A different industry, a different technology stack, documentation in another language. Brought onto the line in a day.
~147,000 lines

See what it reports about your system.

Pointed at a system nobody fully understands any more, the factory maps it and reports where the gaps are in about ten minutes — without overwriting a single existing document. Your prior work is input, not something we replace.

Talk to us