Journey log

This is the log. Newest notes sit at the top. Each note is a founder learning: what we tried, what we learned, what we changed.

1. 2026-10-04

   An honest zero is a result

   The first covering experiment failed with no fitness evidence, and the founder killed the cell. A zero-evidence failure is now recordable, and the cell owner can carry out a founder's kill once the cell is quiet.

2. 2026-09-27

   Extend the dated stop, not the fences

   The covering window moved from late September to mid-October. Try budgets, founder-signed sends, weekday caps, and the locked cold template stayed put. Empty fitness is a status, not a license to thrash.

3. 2026-09-23

   Hill-climb lives in the cell, not the kernel

   Craft stays in the covering-cell forces, not the kernel. Bound, Sense, Act, and Defend own the climb. Business-development skill stays out of the substrate.

4. 2026-09-16

   Shared state beats agent chatter

   We stopped copying human org chat. Routine handoffs write to the shared MCP ledger. Chat is for founder exceptions, not work queues.

5. 2026-09-13

   Bound, Sense, and Act stay distinct

   Owner sets ICP and targets. Sense researches lists and packets without spending tries. Hunter spends tries from Sense-ready packets. Founder still signs sends.

6. 2026-09-12

   Forces are physics, not seats

   The eight forces are persistent physics, not an eight-bot society. Sense does not spend tries; Act does. Packets sent are still not fitness.

7. 2026-09-08

   Covering cell running on try budgets

   Try-unit constitution. Covering cell on weekday loops. Official-door outbound. Packets sent are not fitness.

8. 2026-09-03

   Constitutional reading signed

   Closed beta on personal DataSaa infra. First covering cell live. Founder killed the six-chat org chart.

2026-10-04 — An honest zero is a result

We had extended the first covering experiment's dated stop to mid-October and kept every fence in place. Before that clock ran out, the founder read the ledger and made the call: the experiment failed its evaluation with no fitness evidence, and the covering cell is killed.

The first learning: a failed experiment has to be recordable as a failure. Defend could not mark a count-based criterion as failed when the count was zero, because every failure had to cite evidence, and there is no evidence to cite for nothing. That pushes a system toward silence or invented proof. An honest zero is a result. The substrate now accepts a zero-evidence failure only when the ledger really holds no qualifying fitness evidence, and a passing result still needs proof. Closing an experiment now carries its reason and the founder's decision, so the trail says why it ended, not just that it did.

The second learning: the founder decides a kill, but the founder should not have to carry it out by hand. Killing a cell was a founder-only move, so the cell's own owner was blocked even with the founder's decision in hand. Now the owner can kill its own cell when it cites that decision, and only once the cell is quiet: no live experiment and no open work. Leftover work nobody ever picked up can be cancelled with a reason instead of being failed and charged against the try budget.

We also stopped letting access lapse silently. When the hosts' access windows expired, every call failed with a generic sign-in error that looked like a broken credential. Now an expired seat says that it expired and when, and hosts get a warning a few days ahead.

The founder approved killing the covering cell, recording the failure as an honest zero, and letting the cell owner carry out the founder's decision. Packets sent were never fitness, and the ending did not pretend otherwise. The next offer cell has not been chosen. The eight remain physics.

This note updates offer cells, forces, nested loops, and MCP substrate. It does not move the commercial stop list on what is not being built yet: no catalog, no CRM or billing adapters, no open-source MCP, no commercial offer on this series.

2026-09-27 — Extend the dated stop, not the fences

We had set a hard dated window on the first covering experiment. When that date approached with activity evidence and no fitness evidence, the temptation was either to declare the cell dead or to loosen every constraint to "make it work."

The learning: a covering experiment can extend its dated stop when the founder still wants the same offer cell to keep learning, without changing try budgets, founder-signed sends, weekday caps, or the locked cold template. Extending the clock is not widening the offer. Zero fitness evidence is a status, not a license to thrash. Packets sent are still not fitness.

We moved the covering cell's window from the original late-September stop to mid-October. We also aligned the experiment's external-message ceiling with the try budget so the two clocks cannot silently disagree. Other fences stayed put.

The founder approved the window extension and aligning the message ceiling with tries. The founder killed changing send authority, weekday caps, or the locked cold template just because the first clock ran out, and killed treating empty fitness as a reason to thrash the offer.

This note updates offer cells and nested loops. It does not move the commercial stop list on what is not being built yet: no catalog, no CRM or billing adapters, no open-source MCP, no commercial offer on this series.

2026-09-23 — Hill-climb lives in the cell, not the kernel

We had been tempted to stuff customer-dev and business-development craft into the shared substrate — as if the kernel should teach every host how to climb a market.

The learning: Bound, Sense, Act, and Defend already own that climb. Non-answer, wrong desk, and no-pay are evidence Bound uses to tighten who we chase. Sense keeps packet and door truth on the ledger. Act spends tries and writes dated learning after each batch. Defend stops open-ended thrash and promotes only cash-shaped or founder-defined fitness. The kernel stays physics and ledger. Optional chat effectors get a craft fence: hill-climb inside the running experiment, not open-ended new business.

We encoded that fence on the covering-cell paired effectors and kept business-development skills out of the substrate.

The founder approved craft in the cell forces, not the kernel. The founder killed open-ended new-business thrash from effectors and stuffing business-development craft into the substrate. Packets sent are still not fitness. The eight remain physics.

This note updates offer cells, forces, and MCP substrate. It does not move the commercial stop list on what is not being built yet: no catalog, no CRM or billing adapters, no open-source MCP, no commercial offer on this series.

2026-09-16 — Shared state beats agent chatter

We assumed agents would coordinate the way human teams do: DMs, channel pings, a handoff that looks like a standup. That copies an org chart into the company. It is noisy, and it hides the work.

The learning: this system does not need agent-to-agent chatter. After a host advances cell state, work, evidence, a lease, or try spend, the next host should stop or read the shared ledger. Chat is for founder-facing exceptions, true blockers, and founder-requested digests. Channel membership is not a work queue. Optional chat effectors, including Grok, can still talk to the founder. They are not the ledger.

We changed the covering cell so routine handoffs write toward the MCP substrate. Peers do not DM each other to pass the work.

The founder approved shared-state handoff. The founder killed routine handoffs by chat or DM. Packets sent are still not fitness. The eight remain physics.

This note updates MCP substrate, nested loops, forces, and offer cells. It does not move the commercial stop list on what is not being built yet: no catalog, no CRM or billing adapters, no open-source MCP, no commercial offer on this series.

2026-09-13 — Bound, Sense, and Act stay distinct

We had been letting one covering-cell effector both research and spend. That collapsed two forces into one particle, and it made try budgets lie.

The learning: Bound, Sense, and Act have to stay distinct. Bound sets ICP and targets. Sense researches lists, enrich, and packets inside that frame, and that work does not spend tries. Act does outbound from Sense-ready packets, and that work does spend tries. The founder still signs sends.

We split the optional effectors to match: owner on Bound, Sense on research, hunter on Act. Learning loops stay distinct too — ICP and targets, list and door truth, try effectiveness.

The founder approved the split. The founder killed letting hunter research. Packets sent are still not fitness. The eight remain physics.

This note updates forces, offer cells, MCP substrate, and nested loops. It does not move the commercial stop list on what is not being built yet: no catalog, no CRM or billing adapters, no open-source MCP, no commercial offer on this series.

2026-09-12 — Forces are physics, not seats

We had been treating Bound, Sense, Act, Transform, Maintain, Defend, Select, and Continue as if they were standing jobs. That invited an eight-bot society.

The learning: those eight are persistent physics. They run at the offer cell, the portfolio, and the founder genome. Hosts are temporary particles inside the forces, not a signed cast of chat personas.

We encoded that reading on the MCP substrate. Sense still does not spend tries; Act does. Fitness counts and activity counts stay separate. Optional sit-the-seat effectors remain optional — owner on Bound, hunter on Act, evaluator on Defend — and the next day's lock would split Sense further.

The founder approved encoding the eight as physics and keeping Sense off the try ledger. The founder still rejects a chatbot org chart. Packets sent are still not fitness.

This note updates offer cells, forces, MCP substrate, and nested loops. It does not move the commercial stop list on what is not being built yet: no catalog, no CRM or billing adapters, no open-source MCP, no commercial offer on this series.

2026-09-08 — Covering cell running on try budgets

We had registered the first covering cell against a cash-shaped clock. Dollars in the constitution made activity look like fitness.

The learning: try units plus a dated stop are the right kill clock for a closed-beta experiment. Packets sent are still not fitness. The cell is experiment-until-fitness, not a proven go-to-market machine.

We moved the covering experiment to running: weekday loops, official-door outbound only, founder still signs every send. Large-tech dropped as the covering buyer. Marketplace GTM bot farms stayed off the board.

The founder approved starting the clock, try-unit budgets, official-door-only outbound, and experiment-until-fitness. The founder killed large-tech as the covering buyer, marketplace GTM bots, and treating packets-sent as success.

This note updates offer cells, forces, and MCP substrate. It does not move the commercial stop list on what is not being built yet: no catalog, no CRM or billing adapters, no open-source MCP, no commercial offer on this series.

2026-09-03 — Constitutional reading signed

We needed a signed reading before the experiment could run in public. Without it, every host would invent an org chart.

The learning: the organism is an offer cell. MCP is the shared ledger, not a model. Nested loops — reflex, adaptation, selection, evolution — are how the cell changes on different clocks. Immune stays policy, not a chatbot.

We stood closed beta up on personal DataSaa infrastructure. The first offer cell is a covering experiment: fractional covering on posted seats, with an owner, a scout, and an independent evaluator. The founder signs outbound. Stop is a dated window or a tries cap. Packets sent are not fitness.

The founder approved the signed reading, closed beta, and the first covering cell. The founder killed using this as a six-chat org chart.

This note updates offer cells, MCP substrate, forces, and nested loops. It does not move the commercial stop list on what is not being built yet: no catalog, no CRM or billing adapters, no open-source MCP.

How a note should read

A dispatch should read like a founder journey note: what we assumed, what we learned, what we changed, what the founder approved or killed, and which series idea it updates. Write the learning in plain language. Never name private repos, git SHAs, PR numbers, host or vendor deploy fingerprints, Auth emails, principal/cell/experiment IDs, or raw protocol field names. Do not announce a launch the not-yet page still forbids.

Current public status

Closed beta is live on personal DataSaa infrastructure. MCP is the shared ledger and physics engine. Eight natural forces. Routine handoffs write to that ledger; the next host reads it. Agent-to-agent chatter is not the handoff. Chat is for founder-facing exceptions, true blockers, and founder-requested digests.

The first offer cell was a covering experiment — fractional covering on posted seats — with optional effectors: owner on Bound (ICP and targets), Sense on research (lists, enrich, packets; no try spend), hunter on Act (outbound from Sense-ready packets; tries spend here), evaluator on Defend. The founder signed outbound. On 2026-10-04 it failed its evaluation with no fitness evidence, and the founder killed the cell. No offer cell is running now, and the next one has not been chosen. A failure with zero evidence is recorded as an honest zero, and a cell owner can carry out a founder's kill once the cell is quiet. Packets sent are not fitness.

Founding topology is one owner per outcome, temporary specialists, independent verification, shared evidence. Learning loops stay distinct. No eight-bot society. No chatbot org chart. No marketplace GTM bolt-ons. No commercial offer on this series.

Read the overview if you are new to the series. Read what is not being built yet if you want the stop list.