Who is speaking

I am Christian Honl. My first chat with Claude was on 14 January 2026; from 2 April the git record runs.

Taken from

measured system state

Measurement window

from 2026-04-02 to the measuring day

Figures with a command

Appendix A, [C1]–[C25]

What is open

section 06

Since then, measured rather than described (as of 29.09.2026): 6,517 commits across 7 repositories [C1]; a 2.46 GB append-only case database whose 118 write-path triggers abort a bad write instead of logging it [C6][C7]; 135,014 provenance rows, 135,013 of them carrying the git commit that produced them [C8]; and 56 named gates, 23 of which are fed a known-bad input every night and required to go red — 4,721 proven, 186 unproven, 71 broken, last run 10.09.2026 at 01:59 [C21][C22].

Two of my monitors are red today. They are on this page, in section 06 — the restore drill has been red since 1 August [C20].

I did not hand-type this code. I orchestrated Claude; my work was architecture, judgment and rejection. The customer was a real legal case — self-acquired, unpaid — and the deadline was a court. Every figure on this page carries either the read-only command that produces it or the document it was read from — and says which. So do the figures that are red.

In the end the system has to deliver. Promises are not an option.

How to read the numbers on this page

Four rules, and then the numbers. Every figure carries, in the margin, either the read-only command that produces it or the document it was read from — and says which. Every figure carries its own date (core counters last measured 2026-09-10), and the counters are append-only, so they read low by the time you run them. Where a figure was corrected during verification, the old value stays on the page with a line through it. And what is broken says so, in the same type size as what works.

• 58 gates with daily proof → 56 named / 25 proven gates
• 116 screens → 92 screens
• 842 test files → ~636 test files
• 85 enrichment timers → 29–41 enrichment timers
• 13 models → 12 models
• six LifeRadar frontends → one screen with six views

A 49-claim internal audit came out 28 holds, 20 overstated, 1 refuted — I would rather hand you that ratio than a brochure.

The build, measured

This is the whole measured window on one table, so nobody has to assemble it from prose. Two rows carry a qualification in the row itself rather than in a footnote, because without it the number would be read as something it is not.

Gates that have to go red

186 unproven, 71 broken, 4,721 proven — last run 10 September 2026, 01:59 [C22]. Every night the system feeds each of 23 gates an input it knows is bad and requires the gate to refuse it. A gate that stops catching its own known-bad case is reported, not assumed. Self-reported review is not review, so each run lands in a register the agent that ran it cannot write.

This replays the recorded verdicts: 23 of 25 from the nightly run of 2026-08-23 02:03, two from earlier runs — the date of each run is printed on its own row. It is a replay, not a live gate; the command that reads the register is below.

Why each thing exists, in the order the pain arrived

Six things, each introduced by the moment it became necessary rather than by what it does. I would rather you learned the judgment than the feature list.

A session must not depend on the device in my hand.

I was pushing my child's stroller with one hand and holding the phone with the other so a session would not die.

So the sessions moved to the server, persistent behind an identity plane at the edge; a dropped connection costs a reconnect, not an afternoon. The hard part was not the terminal but the rendering — making server output read cleanly on a phone.

the mechanism

4,731 events / 310,622 transcript rows · xterm 5.5.0, WebGL addon 0.18.0, canvas 0.7.0, fit 0.10.0 pinned exactly · 16 logged research runs on this one stack · an upgrade to xterm 6 evaluated and deliberately rejected. In the same table: it runs without a systemd unit — no auto-restart.

At a customer: their identity plane and their infrastructure; I bring the retention rule and the inherited policy caps.

The reins

What got me into this was the hype around autonomous agents. Building taught me the opposite: to get the best out of an autonomous model you need tight reins, not a loose hand. Riding the bull is the point; the reins are what make it possible.

The standing limit: single operator, no four-eyes principle, no central access audit log.

• Nothing sends to a human on its own — a hard-coded allowlist, not a review comment.
• Every autonomous loop has a kill path that does not depend on the loop.
• No done counts without a register entry the agent cannot write.
• External tool output is treated as untrusted, and never executed.
• Secrets only via .env, with a scanner in every commit.
• Bypasses are counted, not forbidden — 43 of 43 --no-verify bypasses resolved; the ledger does not block the bypass, it blocks the next commit [C13].

What is open, and what failed

This section has the same type size, the same rows and the same commands as section 02, and it sits above my CV in reading order. It is live: it is fed by the same reading as the strip in the masthead, so if the restore drill goes green this page says so, and until then it says this.

sample data — not a live probe

• 2026-08-01DR restore drill red, ExecMainStatus=1 — the backup exists, the restore is not currently proven [C20].
• 2026-08-22Litestream staleness monitor in ALERT, replica WAL timestamp unreadable [C20].
• 2026-08-22The 19 GB law corpus out of live replication, only a weekly S3 snapshot, last one 2026-08-16.
• 2026-08-2395 timers enabled but dead [C12].
• 2026-08-23Evals: 5 cases, unscheduled; no model tracing, no span trees, no prompt versioning.
• 2026-08-23OpenTimestamps: 0 of 72 anchors re-verified — stated as anchored, not re-verified.
• 2026-06-18The content and video pipeline resting; the renderer failed 4 of 4 nights while reporting exit 0 — a silent failure inside the pipeline that claims to fail loudly, and the one claim the internal verification pass flatly refuted, kept here for that reason.
• 2026-06-18Investments stalled with two open backtest defects I found in my own code.
• 2026-08-23560 house-search entries with no matching acknowledgement [C24].
• 2026-08-23Tenant isolation resolver-level only; 564 files still carry the old shared-database string.
• 2026-08-234 of 7 repos have no commit-hook chain, 30 of 51 gates run only on my machine — the portability bottleneck.
• 2026-08-23Single operator, no four-eyes principle; tx_to is never set, so the schema is bitemporal and the usage is not.

Where the data goes

This is usually the first question, so it is on the page rather than in a call. On 23 August my own research engine ran against my own configuration and found four things. I would rather you read them here than find them later.

What the system does

• Purpose-typed delivery with default-deny at the delivery boundary.
• Code names instead of third-party names in every output.
• A personal-data purge on the deliver path.
• IBAN redaction across 21 importers.
• Deletion-marking instead of DELETE.
• A per-source licence registry.
• An Article 17 erasure log in WORM storage.
• Article 25 privacy-by-design evidence generated from live state.
• An SSRF-guarded fetch.
• HMAC-signed SSO that fails safe without its secret.
• Cloudflare tunnels instead of open ports.
• The hard-coded rule that no email and no message ever leaves without a human.

The four findings of 2026-08-23

Stated as of 2026-08-23, with the instruction to re-check before contracting rather than to rely on this page.

At a customer, I deliver with your stack

The honest per-component account: what I use here, what I do not, and what I would reach for on your side of the table. I built my own where the official piece did not exist yet or did not go deep enough, and I adopt the official piece at the visible edge, as a thin wrapper, never as a rebuild.

Deviations I would undo today

• Ad-hoc HTTP routing instead of MCP tool contracts — my tool boundary is private and undocumented.
• My own structured-output hard gate instead of API tool use — it should shrink to a verification layer.
• My own prompt-role panels instead of Agent SDK sub-agents — the persona design is worth keeping, the runtime underneath it probably is not mine to maintain.
• Prompt caching only on the TypeScript path — the router leaves the largest cost lever unused.
• The Batch API adapter left at “wave 2” — the nightly steps that are not latency-bound belong in batches.

The two decades the build record cannot show

Everything above is one person and a model. This is the other twenty years, and it is the one part of this page you cannot verify from a terminal: it carries no commands, only references on request.

• 2018–2020Deutsche Bahn RIS Mobil, 45+ people — Scrum Master of two teams toward self-organisation, hired and built a third team and a DevOps team, agile process adapted to high-security requirements, the app in use across Germany.
• 2017–2018Deutsche Bahn RIS Fahrzeug, 20+ people, three teams, product management built up.
• 2015–2016Tourism go-live, 70+ people, live after eight years of development.
• 2014–2015Public services, 100+ people, eight agile teams heading for twelve where LeSS hit its limits; prepared the scaling toward SAFe and built the visual centre.
• 2009–2013Commerzbank, 50+ people — started as a senior developer, found the V-model too far from the customer, built four agile teams as Scrum Master while leading the frontend overhaul.
• 2005–2009Dresdner Kleinwort Wasserstein / Dresdner Bank — introduced the web limit-check service that validates risk limits at deal closing.
• 2002–2005Dual study programme, thesis on evolutionary algorithms, grade 1.0, which is where the interest in search over unbounded solution spaces began.
• 2021–2026Parental leave, two small daughters, Agile Disruption written (2024, 41 pages, 10 chapters), the systemic-constellation training completed.
• since 2016Managing director of AgilConsultants.

The book → system bridge

• idiot index → question 3 of the seven-question decision filter
• every requirement carries a name that defends it → 49 dated quote sites in the doctrine file, 56 of 62 distilled principles verbatim
• delete until you have to restore 10 % → a 26.8 % backlog cut with an append-only close ledger and a --reopen path
• automate last → a gate goes SOFT, then proven by mutation, then HARD, then scheduled — never straight to automation

The time building alone with a model was not a retreat from teams; it was the fastest way to learn the material I now want to bring back into one.

Check it yourself, and what I will not hand over

I would rather you touched the system than read about it. The refusal comes first, because it is the part that took judgment: my own security review failed my first draft of this section — a plain terminal login on the production host would have been a shell as the database-owning user. So: no shell on the production host, ever.

Three offers, in ascending cost, each with its limit named first

Three walkthrough tracks

Contact

Write to me and say which of the four doors you came through. I answer myself.

[email protected]