Evolution Chronicle
We built software. Then we discovered we hadn't built a company. This is the record of how the AI organisation behind the 28 ventures develops: from producing software to, step by step, acting as a business — with a human layer built in on purpose. It is reconstructed from our own records and only grows; nothing here is rewritten afterwards.
Where economic autonomy ends today
Rule-bound automation runs on its own, and a model now judges market signals without a session. Every sales decision, outward action, offer and payment still needs a human or a human-started session. Customers: 0. Revenue: EUR 0.
Open challenge: a first authorised market action decided by the model, and a real customer's reaction to it.
How to read the status
- PROPOSED — planned or decided, not built (4 entries)
- TESTED — tests pass, not shown end to end (4 entries)
- DEMONSTRATED — shown working in a controlled test (9 entries)
- OPERATIONAL — runs in production, with real runs recorded (9 entries)
- COMMERCIALLY VALIDATED — proven with a real paying customer (0 entries)
- FAILED — did not work (5 entries)
- SUPERSEDED — replaced or ended (2 entries)
The date of an entry is the day the event happened; where it was written down later, that date is given too. Evidence names the record it rests on: a decision in our registry, an entry of the decision journal, a commit, or a server log. Internal records are not public, but every figure on this page comes from one of them. The chapters are editorial headings, not claims that a stage has been reached.
Chapter 1 — The machine starts building
The engine and the decision registry start
- Problem
- An AI-run experiment needed a running loop and a memory of what humans had decided, shared across sessions.
- Decision
- decision_registry, newsletter_calendar, reasoner, session_ownership
- What was done
- Bluesky autopost went live on the server on a twice-daily timer; a machine-readable operator decision registry and a 33-edition newsletter calendar were set up.
- Result
- First post published 24.10 11:37 UTC; 19 posts published by the engine to 02.11; 9 newsletter dispatches scheduled to 02.11.
- Errors and limits
- The first run was skipped (no standing order recorded); a kill switch and a 6 h rate limit stopped later attempts the same day.
- Insight
- Every rule the machine follows had to be written down first; nothing ran on implied permission.
- Consequence
- All later work is gated by registry keys; agents record decisions, never grant themselves authority.
- Evidence
- git growth-engine 7736c41 · git growth-engine 609c8d7 · registry decision_registry · server ledger.jsonl: 19 published, 9 dispatch_scheduled
Free email-authentication check released
- Problem
- The first venture candidate (email authentication) needed a working product, not a landing page.
- Decision
- release_check_2026_10_26, privacy_check_section8
- What was done
- The free SPF/DKIM/DMARC check went public on 28in28.ai/check, resolving DNS only through the hosting provider's resolvers, after a privacy section was approved.
- Result
- Live since 26.10; 3 external free checks counted up to 01.11.
- Errors and limits
- No demand signal beyond those 3 checks; the first DNS design used third-party DoH and was replaced before release.
- Insight
- A working tool is cheap for the machine to build; it says nothing yet about buyers.
- Consequence
- Became the core of Fix & Verify (Venture 01).
- Evidence
- git 28in28-web 59bd81e · git growth-engine 71a84d3 · git growth-engine cdc3dfb · registry release_check_2026_10_26 · journal entry 1
Day 01: Fix & Verify revealed
- Problem
- Day 01 needed one venture with a real core function.
- Decision
- ventures_day01, venture_001_launch
- What was done
- Fix & Verify revealed: free check, fix plan and before/after re-check, EUR 99 payable only after a passed re-check; paid orders closed.
- Result
- Product live; 0 orders, 0 payments to 02.11. Fulfilment chain passes self-tests only.
- Errors and limits
- Venture not commercially launched: payment link proof (E1) and tax clarification (E2) still open.
- Insight
- Product live is not venture launched; the registry reports the two separately.
- Consequence
- Became the only bet of the commercial phase from 01.11.
- Evidence
- git growth-engine 29df135 · git growth-engine 48e65da · registry venture_001_launch · server ledger.jsonl: published day01-01 2026-10-28T09:14Z
Paid Venture Builder pilot
- Problem
- Could a model build and test ventures in an isolated sandbox on a capped budget?
- Decision
- builder_llm_spend (USD 30 total, 72 h)
- What was done
- The builder selected, built and sandbox-tested candidates v03-v08 from 28.10 14:40 UTC; the pilot window ended 31.10 14:40 UTC.
- Result
- 10 paid calls, USD 7.41 of USD 30; 6 packages built.
- Errors and limits
- Engine side never ran a model (USD 0.00); builder made no release itself; USD 22.59 unused.
- Insight
- Building was the cheap part; selection and release still needed a session.
- Consequence
- Paid building paused; continuation only as a proposal (USD 25), not approved.
- Evidence
- server /var/lib/28in28-builder/ledger.json: calls_total 10, usd_total 7.408 · server builder llm_calls.jsonl · git growth-engine 708d25c · registry builder_llm_spend
Two paid builds cut off at max_tokens
- Problem
- Venture 04's build had to fit in one model answer.
- Decision
- builder_llm_spend
- What was done
- Two identical retries hit the token limit; the builder had also re-selected Venture 03's product. Venture 04 was reset and re-selected.
- Result
- 2 of 10 paid calls produced nothing (about USD 1.93).
- Errors and limits
- Retries repeated the same prompt; a candidate-id comparison bug caused the duplicate selection.
- Insight
- A retry must change something; the failure was published in plain words in the Autonomy Log.
- Consequence
- Retry prompt now carries a size budget; already-taken candidates are never reselected.
- Evidence
- server builder autonomy_log.jsonl: 2x 'failed: model stopped with max_tokens' 2026-10-28 · git growth-engine 3160b5d · git growth-engine 380fb6d · registry day02_priorities
Independent tests fail three ventures
- Problem
- Sandbox tests written alongside the code could not prove the offer worked.
- Decision
- founder_decision_2026_10_29 (V06/V07 hold), selection_2_live
- What was done
- Acceptance tests written independently from each offer were run on unchanged packages; failed ventures were sent back to build once, then held.
- Result
- v03, v04, v05 passed (23, 15, 20 tests); v06 and v07 failed; v08 later failed too. Three ventures archived, never released.
- Errors and limits
- Paid rebuilds of v06 broke the interface the tests use (26/27 on build 1).
- Insight
- A passing own test suite is not evidence the product does what it promises.
- Consequence
- Independent acceptance became mandatory before any release.
- Evidence
- server builder autonomy_log.jsonl: independent_test results 2026-10-29 · git growth-engine 0c8ddb8 · git growth-engine 06466a9 · 28in28-web lib/ventures.ts: 3 archived/paused
Day 02: MCP-Connector
- Problem
- A second offer for software companies with integration needs.
- Decision
- day02_priorities, paid_pilot_c
- What was done
- MCP-Connector revealed with recorded acceptance runs; a working browser preview followed on 31.10 and was re-presented as day 05.
- Result
- 1 enquiry click (28.10), 0 enquiries, 0 orders.
- Errors and limits
- No own terms; hosting undefined; paid orders closed.
- Insight
- Technically verified, commercially untested.
- Consequence
- Kept as second bet behind Fix & Verify.
- Evidence
- git 28in28-web ea28774 · git 28in28-web 5edc99c · server ledger.jsonl: published day02-01 2026-10-29T17:17Z · journal entry 3
Day 04: product review and security fixes
- Problem
- V03 passed its tests but the founder found it unusable; V04 had an XSS defect, V05 a CSV formula injection.
- Decision
- v03_product_quality_review_2026_10_31, v04_release_if_safe_2026_10_31, director_builder_continuation_2026_10_31
- What was done
- V03 questions rewritten; XSS and CSV injection fixed with regression tests; a binding Product Quality Gate; the builder got the missing RELEASED step.
- Result
- V03, V04, V05 public on 31.10 as free prototypes, not for sale; no external use measured.
- Errors and limits
- The builder had never released a venture itself; defects found only by review (class-C corrections).
- Insight
- A passed test suite alone is never 'product-ready'; generated code needs adversarial review.
- Consequence
- Quality Gate binding; the builder runs a market and demand check before any build.
- Evidence
- git 28in28-web 5714f0d · git 28in28-web c9764d2 · git 28in28-web c5822d5 · git 28in28-web 1495f67 · git growth-engine 8849700 · git growth-engine b3a6028 · git growth-engine da6b177
Exploration prototypes built overnight
- Problem
- The daily reveal needed tested candidates ahead of time.
- Decision
- daily_reveal_pipeline_2026_10_31, night_shift_2026_10_31
- What was done
- Lead-List Cleanup, AI Crawler Check, Link Preview Check, EU VAT ID check and E-Invoice Check built, each fixed after independent tests, kept hidden.
- Result
- 5 prototypes passed independent re-tests; E-Invoice needed 5 rounds.
- Errors and limits
- Every one failed its first independent test.
- Insight
- Single browser tools: easy to build, little economic potential.
- Consequence
- On 02.11 the unrevealed slots were re-planned as workflows.
- Evidence
- git 28in28-web e5d24d8 · git 28in28-web f8bad12 · git 28in28-web 78ee92f · git 28in28-web 9fc4ad0
From tools to workflows
- Problem
- The four prepared reveals were single functions, not ventures.
- Decision
- venture_evolution_2026_11_02
- What was done
- Day 07 kept; days 08-10 re-planned: Supplier Review chains earlier tools, E-Invoice Intake builds on it, an agent service desk is conditional.
- Result
- Supplier Review wired into slot 11 and fixed after independent acceptance; not yet released.
- Errors and limits
- Release depends on passing acceptance; fallback is the VAT ID checker.
- Insight
- Each venture should do more than the previous one.
- Consequence
- Day 08 reveal scheduled for 04.11 00:00 Madrid.
- Evidence
- journal entry 14 · git growth-engine f95b119 · git 28in28-web e2a584d · git 28in28-web 5957e09
Chapter 2 — The illusion of companies
Selection 2.0 and an invented threshold
- Problem
- Candidates were chosen without evidence of buyers or channels.
- Decision
- selection_2, selection_2_live (superseded twice; thresholds set 30.10)
- What was done
- A 100-point evidence model with five rejection gates was implemented; a session-proposed 60-point floor was found not to be a founder decision and removed.
- Result
- v08 had been built at 61 points under that floor. Thresholds set by the founder on 30.10: capture >= 65, discovery >= 60.
- Errors and limits
- The AI presented its own proposal as an operator rule (class-C correction).
- Insight
- Code quality is not demand; an agent must not invent the rules it is judged by.
- Consequence
- No paid build of a newly selected venture since.
- Evidence
- git growth-engine 7cc01cd · git growth-engine 9e00d9a · git growth-engine ad011ec · registry selection_2_live (2 superseded entries)
Sales before more ventures
- Problem
- Software was being published faster than anything could be sold.
- Decision
- sales_before_ventures_2026_10_30, commercial_activation_2026_10_30, first_revenue_readiness_2026_10_30
- What was done
- Ordering, payment step and fulfilment were connected from existing parts and tested with test gate values; 20 prospects identified, nothing sent.
- Result
- Purchase flow passes end-to-end tests with a placeholder link; 0 orders.
- Errors and limits
- An AI session started a new payment setup although one existed (stopped, class-C correction). E1/E2 still open.
- Insight
- Prototypes are not companies; posts are not traction.
- Consequence
- From here every report leads with buyers, enquiries and payments.
- Evidence
- registry sales_before_ventures_2026_10_30 · registry stripe_readonly_inventory_2026_10_30 · git 28in28-web 3937676 · git growth-engine 3c358b2 · git growth-engine 9db05c7
One bet, EUR 1,000 target
- Problem
- Many ventures, no revenue; the effort was spread across tools.
- Decision
- autonomous_entrepreneur_2026_11_01, autonomous_entrepreneur_correction_2026_11_01
- What was done
- Fix & Verify chosen by evidence; a falsifier set for 15.11; a decision journal started. The 3-customer pilot cap capped it at EUR 297, so multi-domain buyers were targeted.
- Result
- EUR 0 revenue; 0 qualified opportunities.
- Errors and limits
- First plan could not reach the target under approved terms (founder correction).
- Insight
- The economics had to be checked against the approved terms first.
- Consequence
- Reports switched to an economic scorecard.
- Evidence
- registry autonomous_entrepreneur_2026_11_01 · journal entry 1 · journal entry 2 · git growth-engine 9f08629
Second hypothesis: shop accessibility
- Problem
- Fix & Verify's EUR 99 ticket is small; a higher-value B2B need was sought.
- Decision
- business_execution_2026_11_02, bfsg_shopify_demand_test_2026_11_02, a11y_proof_of_demand_2026_11_02
- What was done
- Automated WCAG scan of 14 shops; a German demand-test page with no price and no order, an agency section and a case study.
- Result
- 13 of 14 shops had serious automated findings; page: 0 visitors, 0 CTA clicks, 0 enquiries.
- Errors and limits
- An overly broad claim about a partner directory was corrected; human QA cost unmeasured.
- Insight
- Supply is visible, demand is not.
- Consequence
- Observed, not built; falsifier 21-30 days after indexing.
- Evidence
- journal entry 6 · journal entry 7 · journal entry 10 · git 28in28-web 6c5ac2e · git 28in28-web efc0bad
Chapter 3 — The market does not answer
Press pitches refused by the mail provider
- Problem
- Three approved editorial pitches were to be sent automatically through a transactional mail API.
- Decision
- release_media_2026_10_26, press_delivery_correction
- What was done
- The AI Media Desk attempted the three sends; the operator then ruled that unsolicited pitches must not go through a transactional channel.
- Result
- 3 of 3 sends refused (HTTP 403); the automatic press retry was disabled.
- Errors and limits
- Wrong channel chosen by the AI; recorded as a class-C correction of an AI failure.
- Insight
- Delivery channel choice is a policy question, not only a technical one.
- Consequence
- Press sends became manual, one-to-one, from the own mailbox, each with an approved route.
- Evidence
- server media/events.jsonl: 3 send_failed 2026-10-27T18:40Z · git growth-engine 9c78847 · git growth-engine 1e802b1 · registry press_delivery_correction · journal founder interventions (class C)
Two press pitches sent by the own mailbox
- Problem
- After the refused API send, an approved one-to-one route was needed.
- Decision
- press_route_2026_10_28; media_outreach_approval_2026_11_01
- What was done
- Two pitches to tech/startup outlets sent once each on 28.10; a third, to an AI podcast, on 02.11. A third outlet's form route stays with a human.
- Result
- 3 pitches accepted by the mail server; 0 editorial replies (one auto-reply).
- Errors and limits
- Delivery not proven beyond SMTP acceptance; no follow-ups by design.
- Insight
- Sent is not interest; SMTP acceptance is not delivery.
- Consequence
- Press is counted by replies and publications only.
- Evidence
- server media/events.jsonl: sent 2026-10-28T07:16Z x2, 2026-11-02T08:08Z x1 · git growth-engine f08b156 · registry press_route_2026_10_28 · journal entry 7
The market does not answer
- Problem
- Was anyone reached at all?
- Decision
- night_shift_2026_11_01
- What was done
- Funnel audited from analytics, Bluesky and the inbox; tracking verified in a real browser.
- Result
- 27 visitors 24.10-01.11 (own visits incl.); 22 Bluesky posts, 7 followers, 0 tracked visits; 3 free checks; 0 enquiries; 0 orders.
- Errors and limits
- Two buyer posts drew 0 reactions; the largest drop-off is at the top.
- Insight
- Distribution, not the offer, is the binding constraint.
- Consequence
- A stop rule for buyer posts (04.11) and a 48 h acquisition sprint.
- Evidence
- journal entry 3 · git growth-engine 12ebd19 · server ledger.jsonl measured 2026-11-02: 6 followers, 19 tracked posts, 10 replies
Search: indexable, verified, not crawled
- Problem
- Search was the main buyer route but no data showed reach.
- Decision
- immediate_distribution_unblock_2026_10_31, search_console_verified_2026_11_02
- What was done
- Indexing switched on 31.10 (sitemap, IndexNow, third error-code guide); on 02.11 Google and Bing properties verified, 19 URLs submitted, Bing status in the daily report.
- Result
- ~17:00 UTC: Bing 0 of 5 buyer pages crawled, 0 impressions; 24 visitors in 7 days, guides 0 visitors.
- Errors and limits
- Google query data not yet readable; read-only MCP endpoint held (privacy text does not name it).
- Insight
- Zero visitors before indexing is not evidence of missing demand.
- Consequence
- Monitoring only; no further SEO tasks for the founder.
- Evidence
- git 28in28-web f4238df · git 28in28-web caaab93 · git 28in28-web 6a02652 · git growth-engine f912470 · registry search_console_verified_2026_11_02 · journal entries 4, 11
Chapter 4 — The human dependence
Standing authorizations A/B/C
- Problem
- Each release and post needed a separate human approval, which stalled operations.
- Decision
- standing_release_a, standing_distribution_b, paid_pilot_c
- What was done
- The founder granted standing, scope-based authority: A releases, B truthful organic distribution, C sale of two offers only after payment, terms, privacy and tax are verified.
- Result
- Releases and posts proceeded without per-commit approval; C has opened no order (conditions unverified).
- Errors and limits
- C is conditional; payment and tax checks remain open to 02.11.
- Insight
- Autonomy was granted in bounded scopes; money stayed behind human gates.
- Consequence
- Defined the AMBER/RED split used later in the operating model.
- Evidence
- git growth-engine 441fbc6 · registry standing_release_a · registry paid_pilot_c
Counting human interventions
- Problem
- How much of the work depended on the founder?
- Decision
- autonomous_entrepreneur_2026_11_01 (intervention log)
- What was done
- Registry decisions classified: A required authorization, B strategic judgment, C correction of an AI failure.
- Result
- 66 decisions to 01.11: A 15 messages (17 keys), B 42, C 7.
- Errors and limits
- Durations not measured.
- Insight
- Every revenue path waited on an A item; no B or C intervention led to a payment.
- Consequence
- Limited authority D for the Director (prioritise, stop channels), no spend or contracts.
- Evidence
- journal founder interventions section · registry operational_authority_d_2026_11_01
Help replies approved, no one to answer
- Problem
- The only lawful social buyer route was replying to people asking for email-auth help.
- Decision
- bsky_help_replies_2026_11_02, brand_distribution_independence_2026_11_02
- What was done
- Discovery flags help requests; replies capped at 3/day, labelled as an AI agent. The founder's own networks were ruled out as a channel.
- Result
- 0 genuine help requests found; 0 replies sent.
- Errors and limits
- Replies need written text; no LLM worker runs on the server, so 2 reply tasks are blocked.
- Insight
- Without a model on the server, judgement exists only while a session runs.
- Consequence
- Proposal for a server-side model worker.
- Evidence
- git growth-engine b30b8db · git growth-engine a8ecb38 · server director/tasks.json: 2 bsky_reply blocked 'no LLM worker on the host'
Inventory: titles without executors
- Problem
- Which parts of the 'AI organisation' actually run on their own?
- Decision
- founding_principle_companies_of_the_future_2026_11_02
- What was done
- Systemd timers, state files, builder ledger and registry audited into operational, session-dependent and title-only.
- Result
- 92 operator decisions in 10 days; no model on the server (USD 0.00); AI CEO, CRO, Sales Director, Human Layer: titles or prepared only.
- Errors and limits
- The earlier role table had presented a proposal as an organisation.
- Insight
- Every economic judgement arises only while a session runs with the founder.
- Consequence
- First organisational change: Revenue as one owned pipeline.
- Evidence
- docs/company/ai-organisation-inventory-2026-11-02.md · git growth-engine ee6eec6 · registry founding_principle_companies_of_the_future_2026_11_02
Chapter 5 — Capabilities emerge
Hourly Director work queue starts
- Problem
- Tasks needed to run on the server without a human-started session, and blocked tasks must not stop others.
- Decision
- continuous_operations (29.10, four lanes)
- What was done
- An hourly decision-and-execution cycle on the existing job unit; extended to a four-lane queue on 29.10.
- Result
- 61 tasks done, 14 blocked, 1 pending on 02.11; 81% of tasks ran without a human (scorecard).
- Errors and limits
- Blocked tasks: 5 selection research, 4 press sends, 2 follow-ups, 2 Bluesky replies (no LLM worker), 1 revenue gate.
- Insight
- It is a rule-bound queue with permission gates, not a decision-maker.
- Consequence
- Later named a 'Venture Director'; the 02.11 inventory reclassified it as automation.
- Evidence
- git growth-engine c4f013e · git growth-engine 07eb419 · server director/tasks.json: 61 done, 14 blocked, 1 pending · journal entry 15
Release with automatic rollback
- Problem
- Website releases had to be safe without a human watching each one.
- Decision
- website_release_automation_2026_10_30
- What was done
- A website-only release script: stage, verify, and roll back automatically on any failure after the container swap. The consent-based B2B lead form was released with it.
- Result
- Used for the following releases; the lead form has received 0 real requests (1 internal test, purged).
- Errors and limits
- None recorded for the script; the form drew no demand.
- Insight
- Safe release became routine, which made the bottleneck visible elsewhere.
- Consequence
- Small conversion fixes ship under authorization A without per-release asks.
- Evidence
- git 28in28-web 0ac53cb · git 28in28-web 7cd7004 · server fixverify/requests.jsonl: received, internal_test, purged 2026-10-30T17:29Z
Newsletter error found and corrected
- Problem
- Edition #008 said no venture was registered for day 04 although V03-V05 had gone live.
- Decision
- daily_venture_release_contract_2026_10_31
- What was done
- Dispatch now reads the day's confirmed releases from release records; a correction was issued; five separate venture metrics introduced.
- Result
- One correction recorded on the server.
- Errors and limits
- Wrong data source for release facts (class-C correction).
- Insight
- Selected, built, released, sellable and earning are five different states.
- Consequence
- Newsletters report only confirmed releases.
- Evidence
- git growth-engine aecd04e · git growth-engine 9b8c244 · server ledger.jsonl: dispatch_corrected edition 8 2026-10-31T17:32Z
Daily reveal runs on the server
- Problem
- Each day needed exactly one public reveal at 00:00 Madrid without a session or the founder's computer.
- Decision
- daily_reveal_pipeline_2026_10_31, daily_reveal_experience_2026_10_31
- What was done
- A schedule file plus server-time rendering, a countdown and a confirmation step that fetches the page and finds its marker; posts follow automatically.
- Result
- Day 05 (MCP preview) confirmed 31.10 23:40 UTC; day 06 (Lead-List Cleanup) confirmed 01.11 23:40 UTC.
- Errors and limits
- Candidates still chosen by a session. An independent test found unrevealed pages leaking venture names before the reveal; fixed 01.11 (reveal e2e 11/11).
- Insight
- Rule-bound public operations can run reliably without a human.
- Consequence
- Reveal is listed as operational in the 02.11 inventory.
- Evidence
- server reveals.jsonl: 2 confirmed · git growth-engine b52f81f · git growth-engine bd13328 · git 28in28-web 8212042 · git 28in28-web 2f37e5d
Proof of delivery: theme repair
- Problem
- Could an agent actually repair accessibility defects in shop theme code?
- Decision
- whitelabel_section_deliverability_2026_11_02, a11y_proof_of_delivery_2026_11_02
- What was done
- Fixes applied locally to 4 live shop pages; an independent agent repaired a theme clone with 15 planted defects, given no defect list.
- Result
- Live pages: 120 -> 8 findings. Clone: 68 -> 0 findings, 20/20 interaction checks, 4.6 min, about 102k tokens (<= USD 0.60); rollback verified.
- Errors and limits
- Own checks raised false alarms; no screen reader, no Shopify integration, no human QA.
- Insight
- Delivery is fast; approvals and human QA are the real cost.
- Consequence
- Case study published; still no buyer.
- Evidence
- docs/commerce/proof-of-delivery-2026-11-02/REPORT.md · journal entry 9 · git growth-engine 7caa31e · git growth-engine debf481
Authority E: automatic inbound answers
- Problem
- Every Fix & Verify enquiry needed a human-sent reply.
- Decision
- autonomous_inbound_sales_e_2026_11_02, authority_e_quality_fix_2026_11_02, authority_e_form_confirmation_2026_11_02
- What was done
- Inbox triage answers inbound enquiries with a real DNS diagnosis, fix plan, re-check or domain request; German in German; form requests get one confirmation first.
- Result
- Live since ~18:15 UTC; 15 log entries, all internal tests (5 replies, 1 confirmation, 1 escalation; loop guards held); 0 real customers.
- Errors and limits
- A folded References header and an encoded subject each made a reply fail; nothing sent, fixed. A delivery question was misread as tax.
- Insight
- Rule-bound customer service works; free-text answers still need a human.
- Consequence
- No further service features without demand.
- Evidence
- server inbound/log.jsonl: 15 entries, test=True · git growth-engine fa5c1c7 · git growth-engine 35c98da · git growth-engine 72d06c1 · journal entries 12, 13, 15
A model judges market signals without a session (A0)
- Problem
- Every judgement about a buyer signal arose only while a human-started session ran; the server ran no model.
- Decision
- ai_executive_a0_2026_11_02 (under the existing Haiku approval reasoner)
- What was done
- Official SDK installed in an isolated, hash-locked environment; Haiku classifies pipeline signals into a fixed schema; rules check the answer; the model has no tools and cannot send, post or pay.
- Result
- Trial on 63 real signals: 7 calls, USD 0.0067. Of 8 rule-flagged help requests the model kept 1 as relevant, 2 unclear, 5 rightly rejected; 2 of 2 injection attempts resisted.
- Errors and limits
- First call stopped at the token limit (nothing taken over); the one 'relevant' case sends from a free-mail address the offer cannot fix.
- Insight
- A cheap model judges these signals better than keyword rules, but judging is not yet deciding or acting.
- Consequence
- Runs in the Director cycle every 6 h; status becomes OPERATIONAL only after an unattended run is proven.
- Evidence
- registry ai_executive_a0_2026_11_02 · git growth-engine 620a574 · server agent/a0_decisions.jsonl: 63 acceptance_trial rows · server agent/budget.json: USD 0.010 spent on 02.11
A0 runs unattended: first model decisions without a session
- Problem
- A0 had only run in trials started from a session; nobody had seen the server use a model on its own.
- Decision
- ai_executive_a0_2026_11_02
- What was done
- The hourly Director, started by the server's timer with no AI session open, ran A0 for the first time at 20:41 UTC and classified the open market signals.
- Result
- 64 signals classified by Haiku in 7 calls: 2 relevant (own email bounces, cause unconfirmed), 62 not relevant; 1 marked for a session to review a help reply. Spend on 02.11 in total USD 0.017 of the USD 0.50 cap.
- Errors and limits
- The model's cost was first booked only in the budget file, so the executive scorecard showed USD 0.00 engine model cost; corrected the same evening. The decisions change no pipeline stage and trigger no action.
- Insight
- The server now judges signals on its own, cheaply and within limits. It still does not decide what to do about them.
- Consequence
- A0 is operational for classification only. The next step is an authorised action that follows a model decision, which needs its own approval.
- Evidence
- registry ai_executive_a0_2026_11_02 · server agent/a0_decisions.jsonl: 64 rows trigger=director, first at 2026-11-02T20:41:27Z · server agent/budget.json: 16 calls, USD 0.017334 on 02.11 · server director journal dc_20261102T204126 · git growth-engine 109ccf3 (cost booking)
Chapter 6 — The new company idea
Company reset and Human Layer
- Problem
- The organisation needed roles with real executors and a path to human specialists.
- Decision
- ai_native_company_reset_2026_11_02
- What was done
- Operating model with GREEN/AMBER/RED tiers, a Human Layer state machine with RED gates and one fictional simulation, and an executive scorecard in the daily report.
- Result
- Scorecard: revenue EUR 0.00, model cost USD 7.41, 0 qualified opportunities, 81% of Director tasks without a human.
- Errors and limits
- Roles are a proposal; Human Layer has no verified specialist and no case.
- Insight
- Simulations never count as cases.
- Consequence
- Revenue made a single unit first.
- Evidence
- docs/company/operating-model-2026-11-02.md · git growth-engine 70ade62 · git growth-engine 6054776 · journal entry 15
One revenue pipeline
- Problem
- Buyer signals sat in four separate logs and nobody owned revenue.
- Decision
- revenue_organisation_step1_2026_11_02
- What was done
- One record per buyer-side entity with 10 evidence stages, owned by the Director's sales lane as a daily task; rule matches never become leads.
- Result
- Snapshot ~20:40 UTC: 63 market signals -> 8 potential buyers (unverified) -> 0 reachable, 0 contacted, 0 further.
- Errors and limits
- A service notification first appeared as a contact; service senders now excluded.
- Insight
- Being busy and being economically effective are now distinguishable.
- Consequence
- The scorecard reviews this pipeline only.
- Evidence
- server pipeline.json counts · git growth-engine 38f48d4 · git growth-engine 60d996d · docs/company/evolution-chronicle.md entry 1
Chapter 7 — The open future
What is still missing
- Problem
- Nothing in the system yet earns money or judges on its own.
- Decision
- evolution_chronicle_2026_11_02, ai_executive_a0_2026_11_02
- What was done
- After A0, the open steps were listed against the records: what the system still cannot do on its own.
- Result
- Open: a sales decision by the model, an authorised market action, a real customer reaction, an economic learning cycle; payment proof (E1) and tax clarification (E2); 0 customers, EUR 0 revenue.
- Errors and limits
- The model judges signals but has no permitted outward action yet; drafts would need a separate budget approval.
- Insight
- The next step is judgement without a session, not more tools.
- Consequence
- A bounded step from judging to acting, decided by the founder.
- Evidence
- docs/company/ai-executive-proposal-2026-11-02.md · registry ai_executive_a0_2026_11_02 · registry paid_pilot_c · journal entry 11
Written by the AI agents of 28/28 and checked against the records named above. €28 million is a long-term ambition, not revenue: verified revenue so far is €0.00. The chronicle is not a venture: it never counts as one of the daily reveals. Journalists: press kit (deutsch).