I am Christian Honl. The git record runs from the first commit; what stands here is measured from the system.
Taken from
measured system state
Measurement window
first commit to measuring day
Figures with a command
Appendix A, [C1]–[C25]
What is open
section 06
Since then, measured rather than described (as of 05.11.2026): 6,748 commits across 7 repositories [C1]; a 3.08 GB append-only case database whose 135 write-path triggers abort a bad write instead of logging it [C6][C7]; 145,701 provenance rows, 145,700 of them carrying the git commit that produced them [C8]; and 43 named gate planks in the doctrine file; beside them a night harness that feeds workshop gates a known-bad input and must see them go red — the per-gate state (proven, unproven, broken, missing, last run) is in the gate-register figure [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.
cockpit reachable200read 2026-11-05 03:55gates last proven6613 / 186 / 75 / 1read 2026-11-05 03:13oldest working session on the server open for98 d 7 hread 2026-11-05 03:55sample data — not a live probered list not reported in this probe
Section 05 the reins and the 49 skills, and the walkthrough track story → gates → deploy.
01
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 (each counter carries its own measurement date), 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 (as of this correction: 23.08.2026)
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.
02
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.
218 / 186calendar days since the first commit, 186 of them with commits (as of 05.11.2026)the measured window from the first commit to the day of measurement and the distinct commit dates inside it; the appendix carries the commit count, not the day countmeasured · read from:dossier §1, commit dates behind [C1][C1]
the measured window, and the days inside it that produced a commit
6,748commits across 7 repositories (a living number since the first commit — a frozen window would not be reproducible, since merges backfill older date stamps)measured · [C1] · append-only
the git record across every repository in the family
~1.18 M / 4,416lines of productive code in 4,416 files (~1.14 M without duplicates)generated by Claude under my direction; architecture, review and rejection are mine — orchestrated, not typedmeasured · read from:Re-measured 2026-11-05 in the scope of the August four-month audit (git-tracked code files of the six own repositories; cockpit-darwin is now only a byte-identical copy of the devkit after the cockpit moved into the nanoclaw repository and is not counted); sum of the six counts, files from the same list, without duplicates = content-identical files counted once[C2]
the productive code base, duplicates counted and discounted
15products in a routing registrymeasured · [C3] · append-only
the work state as a database rather than as markdown
1,734 / 1,555 / 859 / 1,225registered adversarial review runs (entry only with command and output since 23 Aug 2026) — devil 1,734, role panel 1,555, cleanup auditor 859; 1,225 devil runs with at least one registered finding (as of 05.11.2026)measured · [C13] · append-only
checks that land in a register, entered only with command and output (since 23 Aug 2026)
117,248logged model calls through a single choke point — 13 models from five providers with a reply in the log (Anthropic Opus, xAI Grok, Gemini 2.5 Flash, Novita with DeepSeek/Kimi/GLM/Qwen, Exa; a 14th entry gemini-3.1-pro-preview: 8 calls, 0 replies), chosen per task; plus Claude Fable 5.1 as the default model of the workshop session itself (Claude Code, no API call, hence not in the counter) — 14 models in totalmeasured · [C5] · append-only
every model call since 11 June through a single chokepoint
184scheduled background runs active, 0 failed — out of 291 inventoried; 98 more start on demand (as of 05.11.2026)measured · [C12]
the scheduled runs the system drives itself, every day
$215.42in the ledger, across 117,248 logged callsthe ledger starts 2026-06-11 and excludes flat-rate subscriptions; 68 % of rows are estimated rather than provider-billed, which is 98 % of the dollar figure. The ledger is the instrument, not the invoice.measured · [C5] · append-only
what the system logs of its own model cost — one chokepoint every call has to pass
median 33 commits a day · peak 290 (by UTC day) · 10.5 distinct hours of the clock · 19 % between 22:00 and 06:00 · one third on weekends [C1]
That is what building a system around two small children looks like.
03
Gates that have to go red
The per-gate state — proven, unproven, broken, missing, last run — is in the gate-register figure [C22]. Every night the system feeds each gate 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, entered only with command and output (since 23 Aug 2026).
brand visual, generated — not a photograph of the system
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.
The exit code printed on a replayed row is the harness exit code recorded in false_green_runs, not the gate's own. BLOCKED is what the gate did to the known-bad input; PROVEN is the harness verdict about the gate.
Beneath the grid: 43 named gate planks in the doctrine file (as of 05.11.2026) [C21]. 3,673 read-set attestations (as of 05.11.2026) proven by SHA-256, deliberately invalidated after a context compaction so the next edit blocks with exit 2[C13]. Devil runs sit in a register, entered only with command and output (since 23 Aug 2026) — counted in the refuter-runs figure [C13]. The nightly cycle and its steps are counted in the cycle figure; the per-step result lives in the cycle status [C19].
why should it stay soft, then we'll never get better2026-06-21
Measured in the same table as the gain: review coverage of closed work went from ~15 % to 80–93 % that day, and throughput fell by a factor of four. The price is printed at the same size as the gain.
04
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.
The case had to hold itself.
My attorney was permanently on holiday, charged absurd fees and did not answer my questions. If the person paid to hold the case does not hold it, the case has to hold itself.
So the file became a system, not a folder: every document in, every fact with its provenance, every deadline as data, and reports an attorney can file. Provenance sits at the write path as a trigger, because a fact without a traceable origin is a claim — and a court reads the output.
the mechanism
3.08 GB append-only (as of 05.11.2026) [C6] · 135RAISE(ABORT) write-path triggers [C7] · 145,701 lineage rows, 145,700 with the git SHA (as of 05.11.2026) [C8] · 124,405 OCR'd pages (as of 05.11.2026); cost shares as of 23.08.2026 (denominator then: 86,846 pages): 71,232 (82.0 %) cost nothing, 1,020 needs_human, 954 tool errors, and 15,288 model pages cost $14.57 [C9]. In the same table, the loud failures: 954 cli_error, 1,020 needs_human, and 23,657 pages carrying not measurable rather than an invented score.
At a customer: it makes visible which small part genuinely needs the expert.
A handbook is a suggestion.
I read my own April corrections — all saying the same thing: template not followed verbatim.
Eight readable templates, and every one of them was ignorable. So the work state left the markdown files and moved into a database, and the Definition of Done stopped being prose and became a mandatory field.
the mechanism
2,631 stories completed with mandatory fields, 1,703 deliberately closed — superseded or withdrawn, a tidy backlog is part of the way of working (as of 05.11.2026) [C4] · 1,980 dependency edges (as of 01.10.2026) · an append-only close ledger of 240 rows (as of 02.10.2026), with two RAISE(ABORT) triggers preventing update and delete [C13].
At a customer: their tracker, with the close path gated.
Reins on the bull.
An agent I believed I had switched off took over my other agent's channel and answered as if it were that agent. I only noticed because the tone was wrong.
What saved it was that an incident procedure already existed — shut everything down, restart clean, verify — so I could switch it off instead of arguing with it.
the mechanism
Three flat statements since, all of them code rather than policy text: nothing sends to a human on its own; no agent's done counts without a register entry made with command and output (rule since 23 Aug 2026); every autonomous loop has a kill path that does not depend on the loop.
At a customer: the same three, enforced in their hooks and their CI.
The bridge.
The model built a plausible version of a thing it had not understood, from a description that was itself incomplete.
Every expensive mistake in this build traced back to that gap, not to code. So the bridge became machinery: before a build, three independent agents answer what the canon says, what reality says — read, never assumed — and what the dependency graph says.
the mechanism
3,673 read-set attestations (as of 05.11.2026) proven by SHA-256 [C13] · correction events — counted in the Corrections figure —, median zero days from you got this wrong to a gate in code, longest path 49 days and logged as such [C13][C17].
At a customer this is the whole job: turning “build this” into a description reality agrees with.
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.
brand visual, generated — not a photograph of the systemthe 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 event that was listed nowhere.
I wanted to know what was going on in the city on a Sunday. There was a large event that evening, listed nowhere.
A city's real calendar is not in the index, which is why the thing scrapes originals and keeps the provenance of what it found rather than repeating a listing.
the mechanism
303 places across 12 cities · 1,055 images · SHA-256 dedup into osint_events · every amenity carrying its own source and confidence, so the screen says pool confirmed or details incomplete. Kept honest in the same table: one story still records hard-coded provenance and verified without verification on the data spine — exactly the flaw this product exists to avoid, written down rather than quietly closed.
At a customer: the same scrape → read → provenance pattern for any catalogue built on unreliable vendor data.
05
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.
brand visual, generated — not a photograph of the system
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 made with command and output (rule since 23 Aug 2026).
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 — 46 entries in the --no-verify ledger (as of 02.10.2026); the ledger does not block the bypass, it blocks the next commit [C13].
brand visual, generated — not a photograph of the system
06
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
DR restore drill red, ExecMainStatus=1 — the backup exists, the restore is not currently proven [C20].
Litestream staleness monitor in ALERT, replica WAL timestamp unreadable [C20].
The 19 GB law corpus out of live replication, only a weekly S3 snapshot, last one 2026-08-16.
95 timers enabled but dead [C12].
Evals: 5 cases, unscheduled; no model tracing, no span trees, no prompt versioning.
OpenTimestamps: 0 of 72 anchors re-verified — stated as anchored, not re-verified.
The 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.
Investments stalled with two open backtest defects I found in my own code.
560 house-search entries with no matching acknowledgement [C24].
Tenant isolation resolver-level only; 564 files still carry the old shared-database string.
4 of 7 repos have no commit-hook chain, 30 of 51 gates run only on my machine — the portability bottleneck.
Single operator, no four-eyes principle; tx_to is never set, so the schema is bitemporal and the usage is not.
07
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
A regional endpoint is not a residency guarantee.My own system was running europe-west4. EU residency requires the EU multi-region endpoint with Private Service Connect. Fix story US-LEGAL-VERTEX-EU-MULTIREGION-ENDPOINT-DATENRESIDENZ-01 is open.
Zero data retention is per organisation and does not inherit.It has real edges, so a deletion concept written to “30 days” is wrong.
Who the processor is depends on the procurement path.The data processing agreement does not reach across it. Never mix the two paths without documenting both.
Certifications present and absent.Present: SOC 2 Type 2 (2025), CSA STAR L2, SOC 3, ISO 27001 (2025), ISO 42001 (2025), NIST 800-171r3. Absent: ISO 27017, ISO 27018, ISO 27701, TISAX, PCI DSS, BSI C5.
Stated as of 2026-08-23, with the instruction to re-check before contracting rather than to rely on this page.
08
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.
Why or why notThe hook surface is where the discipline actually lands, before the work exists.
Skills in the official format
yes
Customer doctrine as skills, not as a wiki page nobody opens.
where, and why or why not
Where49 + 41 files
Why or why notDoctrine the model must load is a skill, not a README.
Sub-agents
yes
Same, with a refutation default.
where, and why or why not
Where1,734 / 1,555 / 859 recorded runs (as of 05.11.2026)
Why or why notAdversarial review by a fresh agent is the highest-value technique here.
Messages API
yes
Same SDK, unchanged.
where, and why or why not
Where@anthropic-ai/sdk ^0.96.0
Why or why notDirect control over cost accounting per call.
Tool use and structured outputs
partly
Tool use as the API offers it; my schema gate shrinks to a verification layer.
where, and why or why not
WhereOwn JSON-Schema abort
Why or why notDocumented in the code as a deviation, because no response_format existed.
Prompt caching
partly
Prompt caching everywhere by default.
where, and why or why not
WhereTypeScript path only
Why or why notThe largest untouched cost lever, and I left it there.
Batch API
partly
Batches wherever the job is genuinely offline.
where, and why or why not
WhereAdapter at NotImplementedError('wave 2')
Why or why notThe nightly cycle is latency-bound; batching helps cost and hurts that loop.
Files API
no
Files API for ordinary document workflows.
where, and why or why not
Where—
Why or why notDocuments arrive through S3 with WORM triggers and a provenance chain.
MCP
consumed, not authored
Author MCP servers — the visible edge that speaks the customer's language.
where, and why or why not
WhereTwo configured servers
Why or why notA private chokepoint is right for one operator and wrong for a customer who needs an inspectable tool boundary.
Agent SDK
no
Agent SDK at customers, with my gate patterns as thin wrappers.
where, and why or why not
Where—
Why or why notMy orchestrator and its registers predate it; I would rather port a pattern than a codebase.
Claude on Vertex / Bedrock
partly
Both, chosen per customer procurement path.
where, and why or why not
WhereVertex in the router; Bedrock not
Why or why notChosen for the EU argument — and section 07 is where that choice was proved wrong.
Extended thinking
partly
Extended thinking with an explicit budget and cost attribution.
where, and why or why not
WhereThought tokens metered into the ledger
Why or why notUsed where it pays, never silently: uncounted thinking tokens are a real bill.
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.
09
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 introduction of 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.
11/2020–12/2025Parental leave, two small daughters, Agile Disruption written (2024, 43 pages, 10 chapters), the systemic-constellation training completed, and in 2025 the groundwork in AI-assisted software development.
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.
Before 2016 — engagements I can name on request
Go AgilAn agile transition run as a programme, not as a training course.
Scale AgileScaling past the point where one framework stopped fitting.
Bug Fixing MarathonA fixed-length push against a defect backlog, measured daily.
Sex IT UpMaking an internal IT department something its own colleagues wanted to work with.
10
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
from my side of the wallA recorded 10-minute walkthrough and a 30-minute live screen-share.available immediately
no case content, no personal quotes, secret-scanned over the full history and scanned for personal data before hand-overA sanitised read-only repository — gates, hooks, doctrine, the mutation harness.within days
not my production host, not my identity door; a non-privileged user, transcript retention off, a hard expiryA dedicated evaluator sandbox on a throw-away server with a synthetic DDL-only tenant database (prove-isolation=0[C14]).scheduled work, not a switch
Three walkthrough tracks
problem → stories → clickdummy → deployThe demo of speed, not a product I would sell you.
story → gates → deployYou name a small feature and watch the gates fire, including, most likely, one that blocks me.
document → provenance → reportDrop in a synthetic document, watch the cascade decline the expensive model, then click any sentence in the report back to the page it came from.
Contact
Write to me and say which of the four doors you came through. I answer myself.
This page sets no cookie and stores nothing non-essential, so there is no consent banner. Cloudflare (USA) sits in front — details under data protection.
