Nine questions the firm keeps open, and writes down whichever way the evidence falls.

How does a system carry a lifetime of experience without carrying that lifetime into every thought, and still let the right part of it change what it does when the moment comes?

The usual answer is to store more. This program asks what to reconstruct, when, and under whose control. It treats memory as a path from experience to behavior and finds the first stage that failed, rather than saying the system forgot. The test is behavioral: take the memory away and see whether the action changes. The first laboratory is the firm's own: the memory its engineering agents work from, 159 files with a checker that runs on every write.

What has to change in or around a system for experience at one time to cause reliable, general improvement at a later time, and how do we prove it was learning and not more context, retrieval or luck?

A system can explain its mistake well and stay exactly as likely to repeat it. We measured that in our own work: by 29 September 2026, of 207 logged process failures, 92 named themselves a repeat of a class already on file. The program studies where a lesson should live, in memory, in an instruction, in code or in the model; how credit is assigned across a long run of actions; and when the right lesson is to learn nothing. A change in behavior is the only evidence that counts.

How does a system reliably turn an intent into a verified outcome in the world?

Architecture, tool use, planning, verification, recovery from failure, state, observability and cost, compared rather than summarized, with expensive intelligence spent only where intelligence is required. The laboratory is the firm's own desk agent, studied as it is and not as it is described.

Can a machine do science well: find, test and keep reliable knowledge, and improve its own method, without mistaking sophistication for evidence?

This program reads the other programs' cycles as its data. It measures how often a claim changes when its source is read a second time, how often a cycle finds again what was already known, and whether the method fit the question. On our own record, across four research reports, 15 of 60 load bearing claims changed when their sources were read a second time. Its output is a proposed change to how the firm does research, written as a diff, never applied by itself.

Why does a person take action A instead of action B at time T, and which observable variables, and which interventions that leave the choice with them, change that probability?

It began as a science of selling. Three hundred questions are answered in its first phase; a second registry numbers its questions to 920, and 250 of those were worked in five daily cycles between 24 and 28 September 2026. The same machinery of attention, belief, incentive, identity, uncertainty and friction is at work in hiring, quitting, learning, trusting and delaying, so the question is being widened to its full size. The charter writes its line down: the person's understanding and autonomy come first, and deception, manufactured scarcity and dark patterns are marked do not implement.

What sequence of challenge, explanation, practice, retrieval, feedback, spacing, variation and difficulty produces the most durable skill per hour?

The human twin of program 02. If reasoning can be compiled into skill in a machine, the same question applies to a person on a sales floor, in a clinic or in a trade. Program 05 already requires the practice science for selling: spacing, retrieval, stress, transfer tests. This program lifts it out and asks it of any skill worth having.

Why do some firms produce far more from the same people, capital and tools, and which parts of that can be designed?

Structure, incentives, decision rights, who talks to whom and how information moves, treated as an experimental system: a state, an intervention, a measured outcome, an inference, an update. This is the research under the applied work. The Pilot Study is how a question like this is tested on one firm's own data.

How much useful intelligence runs on hardware a firm already owns, from a desk machine to a microcontroller, with nothing leaving the building?

A thirty five billion parameter model first went live on a desk machine in this office on 31 August 2026. A study of distributed intelligence on microcontrollers is on file, and two boards were running on a bench by 17 September. The program asks what that buys a firm with sensitive data, a thin connection or a thin budget, and where it stops.

What structure lets a group reliably outthink its best member?

Every firm is a group deciding things. Independent judgments before discussion, a record of who said what, a way of combining views that is not the loudest voice: the program asks which of these hold up when tested on real decisions, and whether a team of agents and people can be arranged the same way. It sits one level under program 07.