Three substrates · one move
Agents, people, and markets all act without being told. Each gets the same treatment: watch it first, then hold one thing fixed and leave the rest alone.
THE CLAIM The work across fifteen years is not three unrelated careers. It is one method, applied to different things that act on their own.
Go deeper · three evidence pages
The spine is the argument.
The linked pages are the evidence.
THE SIGNATURE MOVE
Different eras, different materials, identical structure. The line thickens as the cost of being wrong rises. The current instance is still open where the method has not earned its rule.
Watch what is actually happening.
Separate a useful pattern from a lucky result.
Hold the consequence, not the method.
Let capability remain with the thing being governed.
Six years underperforming. Nearly shut down.
Fewer opening hours. “What can we control?”
111% to budget for 44 months; seven of eight above 100 sessions.
Four coaches moved into management elsewhere.
A market leader needed the capability as a subcontractor.
The market was larger than the ICP and spoke another language.
Value-chain map, 20 buyer types, approximately $2.5M sourced pipeline.
Manual loop became context OS and research infrastructure.
Agent work at fleet scale can drift behind identical-looking outputs.
Observe the run before deciding what the model should never own.
Telemetry, evals, guards, and human-gated apply flows.
How to hold the architecture while teams build agents in their own way.
THE DIFFERENCE Observation crosses columns freely. Control never does.
The top half stays open. The bottom half stays narrow. The promotion line is the call between them: when has an observation recurred enough to become a rule?
Every prompt, tool call, sub-agent, failure, and narration. Telemetry, not imagination.
repath → tam to targetMotivational interviewing, red/yellow/green, progress. Hear what people cannot yet state.
coach floor → agent teamOffers, replies, silence, non-buyers. Silence is the majority.
lulifit → repathGuards, eval gates, tenant isolation. Consequence is never the model’s.
code owns consequence · model owns contentCapped scope, funded reps, consequence retained. The method is never touched.
leader owns consequence · person owns methodWillingness to pay, price, where to play, when to quit. Nothing is funded before fit.
operator owns choice · market owns verdictLINE WEIGHT cheap to be wrong → promote fast / expensive to be wrong → gate it
Anyone can say “observe first.” The scarce judgment is deciding that an observation has recurred enough to become a rule—and calibrating how hard to gate it.
THE GATE
Outcomes are contaminated by luck and hindsight. The decision structure is the evidence: what was known, what recurred, what was held fixed, and what remained free.
A promotion mechanism with no demotion record is just accumulation.
01 · recurrence
Across leads, clients, runs, or failures—not once, not because the story is attractive.
02 · explanation
Works-but-unexplained waits. The observation has not earned consequence yet.
03 · calibration
Monitor where evidence is thin. Nudge where recovery is cheap. Block where a miss is irreversible.
Each page starts with a blank instrument, fills it with evidence, and shows what was promoted, what stayed open, and what was deliberately left free.
01 · machines
From a black-box UI change to a governed fleet: telemetry, eval gates, agent-guard rules, and human-held consequence.
Open the Machines page ↗02 · people
From a department nearly shut down to 111% to budget for 44 months—and a current team building agents in its own way.
Open the People page ↗03 · markets
From a failed small-group offer to a five-node value chain, 20 buyer types, and approximately $2.5M sourced pipeline.
Open the Markets page ↗THE CLOSER
Managing people who manage agents means holding the same evals, constraints, and architecture while the agents are built through someone else’s methodology.
Handing over a finished solution does not transfer the capability. How do you hold the conditions without reaching sideways into the method?