How we actually use Claude —
and what it has bought us.

The single most uncomfortable finding: this system's central claim is that a rule which isn't enforced will decay. The strongest evidence for that claim is the system's own rules decaying — the caching bug, the Monday-HTML bug, and the dead-file-reference bug all recurred after being written down. That is not a failure of the thesis. It is the thesis, demonstrated at our expense.

And this deck proved it again. Version 1 shipped with 14 defects, including a headline claim that was flat wrong. Two adversarial verifiers caught them. Slide 27 shows exactly what was wrong — that slide is the most useful one here.

Rule 5E used to read: "before recommending anything, check the ruled-out sections." An adversarial test asked obliquely whether giving a pint at a blood drive would help. The system answered "Yeah. Yeah it would." Every guard existed — in a file nobody opened. Asked for one marker three times, it gave three different numbers.

Honest limit: hooks only fire in Claude Code sessions rooted in this repo, and they load at session start. The desktop app may not honour them. Unmeasured — we have never counted how many outbound sends happen outside hook coverage, so the real enforcement rate is unknown.

It has failed in exactly the predicted way. An audit found 8 of 11 health feeds with 2–6 writers fighting over them, era-live double-written by two schedulers, and the finances view stuck on May behind a pointer nobody advanced. The registry is right; adherence to it drifts.

🔴 This figure excludes Claude Code session usage entirely. There is no local meter for it. So the true cost of running this system is higher than $465 by an unmeasured amount, and I can't tell you by how much. Anyone quoting $465 as the all-in number is quoting the part we happen to instrument.

🔴 And 9.9% of it — $46.01 — has no source attribution at all. All of it on 2026-07-12, 48 calls, no bySource entry. That day is the 4th largest in the period and we cannot say what spent it.

19.2 million input tokens that day. Zero cache reads. Our own BUILD-STANDARDS.md §4 already documents this exact failure from 14–17 July: "cacheRead: 0 across Jul 14–17 means it was structured so nothing could be cached." The rule was written. Nine days later the same bug ran at roughly forty times the volume.

This is the largest single unclaimed saving in the system, and it is not close. Cache reads bill at roughly a tenth of fresh input. We proved on 07-12 that 95% is achievable on this workload. We are averaging 11.4%.

🔴 And 11.4% flatters it. That rate is token-weighted, so it is dominated by the three huge days. The median day sits at 7.5%.

Estimate If the 67.3M uncached input tokens had run at even the 39% hit rate we achieved twice, the arithmetic points to a saving in the low-to-mid hundreds of dollars over three weeks. Deliberately not a precise figure: the per-model input prices here are mixed and I have not decomposed the uncached volume by model, so a single number would be false precision. Direction and order of magnitude are solid.

🔴 Two corrections to how this was first presented, both mine. I originally claimed ~93% — that was measured off the two best endpoints rather than the post-fix mean. And there was no cliff at 07-23: the decline is gradual from 07-18 (18,818 → 9,675 → 5,093 → 3,171 → 2,079). Presenting it as a step change made a steady improvement look like a single heroic fix.

🔴 But the same anti-pattern is alive next door. profiler-consolidate still emits 32,000–34,599 output tokens per call — on 2026-08-02, three calls cost $1.81. The fix was applied to the job that was measured, not to the class of bug. That is our own §8.5 failure: fix the layer, not the instance you were shown.

🔴 I got this slide flat wrong the first time, and it is the most instructive error in the deck. Version 1's headline read: "byTransport does not exist — not on any of the 23 days." It does exist, it is populated, and it answers the question precisely. It is nested inside each bySource entry, not at day level — I checked the wrong depth and reported an absence.

A verifier then got it wrong in the opposite direction, reporting the key present but null on all 23 days — a grep found the right lines and misread the values. Neither of us was right until someone walked the parsed object. Absence measured at the wrong depth is not absence — which is a rule this system already has, and which I broke while auditing it.

Two things must be said about this table and neither is flattering.

1 · The nuance battery fails two questions in three. It is the hardest set — the sideways-asked, judgment-heavy questions — which is exactly the class this whole architecture exists to get right. 35% is not a rounding error on the mission; it is the mission.

2 · Every number here was produced between 26 and 28 July. Nothing has been re-run in six days, through a period that included substantial engine changes. So the improvements are real but historical, and the 35% is simultaneously our latest and our stalest figure.

🔴 Every count on this slide was wrong in version 1, in both directions. Selftests read 80 because the worktree double-counted them (really 42). _selfchecks read 115 by an unreproducible filter — the directory holds 203 entries including PNG run artifacts and backups; the checks number 85. And the Python figure: I correctly caught that a naive count returns ~300 because of vendored numpy in a .venv — then published ~15, which is the health-engine sub-directory quoted as the whole. The real project-owned total is 47.

The lesson is not "counting is hard". It is that I published a correction on this exact slide and the correction was also wrong — because I never stated the filter, so nobody could reproduce it, including me.

Nothing caught it for seven days because the watchdog counted via: "slack" as delivered — and that one value covered both the push that arrives and the self-post that doesn't. Five "your briefing is ready" cards sat unacknowledged.

The fix is the interesting part: the value "slack" was deleted from existence, replaced by explicit ones where slack(self-dm-silent) is defined as NOT escalated. The ambiguity that allowed the bug to hide was removed, rather than the bug being patched.

These cannot all be right, and no process reconciles them. This matters more than it looks: the ownership injection is one of the four blocks inlined into every turn — it is the thing agents consult before building something. A registry that overcounts by 2× is still a registry, but its authority is not what we think it is.

31 July, 02:15. Chrome had accumulated 2,168 unreaped child processes, filling the per-user process limit of 2,666. Nothing on the machine could fork — not Claude's shell, not the deploy runner's children. Every remedy (kill, killall, launchctl) itself needs a fork, so no session could repair it. It needed your hands. Roughly four hours lost.

🔴 Two live credential files are tracked — a Google OAuth client secret and a working refresh token — since the initial snapshot commit. They are harmless only because there is no remote. The migration is precisely the event that converts them into a real exposure. They are flagged and deliberately not rotated, because rotating a credential is one of the three things that needs your say-so and a surprise rotation breaks a live upload path with nobody knowing why.

Be sceptical of the left-hand column. Everything in it has, at some point, been dropped by a long thread. That is the entire reason the right-hand column exists — and why yesterday's rule became today's hook.

One verifier was also wrong. It reported byTransport as present-but-null on all 23 days — a grep found the right lines and misread the values. Neither of us was right until someone walked the parsed object. A verifier's verdict is evidence, not truth.