Flow-Driven — Evolve

Outcome

The workflow gets treated as a product: probes run against real workspace data, each fired probe becomes either an improvement card with a benefit hypothesis and kill criteria or an explicit "checked, quiet" line — and nothing changes on the spot. The improvement lane is where leverage accumulates; edits made in passing are how operating models rot.

Focus argument

• focus:flow — Family A: constraint, rework, WIP, decisions, intake
• focus:evidence — Family B: watermelons, weak checks, self-graded steps
• focus:delegation — Family C: rungs, escalations, context packs, the platform
• focus:definition — Family D: policy contradictions, drift, measures nobody uses
• focus:auto (default) — read the board, flow log, and loop log, and run the 3–5 probes the data is already pointing at

Workflow

1. Load the probe library: skills/flow-driven/references/loop-probes.md.
2. Gather data first. board.md, flow-log.csv, reviews/loop-log.md, learnings/, the items' ## Escalations sections, measures.md baselines, the improvement lane, and the context packs' review dates. Probes read data; they do not free-associate.
3. Run 3–5 probes. For each: signal present? Cite the rows, items, or dates. No signal? One "checked, quiet" line. Three similar learnings count as a signal. A pass that fires nine probes produces a report nobody acts on.
4. Read the delegation ladder deliberately, every pass: which steps earned a promotion (check agreement met, escapes zero, demotion criteria written), which are running above their evidence, and which are promoted on paper but not in the habit — a human holding items in a rung-4 step is a leverage leak the board already flags.
5. Read the platform. Recurring escalations name the missing context pack. Packs past their review date, or contradicted by a learn result, are manufacturing confident wrongness upstream. questions-to-human flat over months means answers are landing in chat instead of in packs.
6. Capture bets, not fixes. Each fired probe worth acting on becomes improvements/<slug>/improvement.md: benefit hypothesis with a mechanism, the measures it claims to move, a baseline, kill criteria written now, and the cheapest discovery — a data probe, a shadow run, a one-item pilot, a replay of the flow log under a different limit. Delegation bets must include the shadow-run design and the demotion trigger.
7. Right-size the ceremony. A policy tweak needs a line and a decision. A rung promotion needs a shadow run. A change to the graph or to classes of service needs the option table with "do nothing" as a real row — most workflow changes lose to it.
8. Report. reviews/YYYY-MM-DD-meta-loop.md: probes run, fired vs quiet with citations, cards created, and which existing bets this evidence strengthens or kills.

Adoption (a separate act)

When a bet's discovery says proceed and a human decides:

1. Record the decision on the improvement card, dated and named.
2. Make the change — workflow.md, a step contract, a limit, a policy, a pack.
3. Add a dated line to the Change log in workflow.md pointing at the bet.
4. Fill the card's Watch table for at least n items. A change nobody watched cannot be reverted with confidence, and a bet that reaches adopted with an empty Watch table is itself a finding.

Rules

• Capture, never implement. Changes to the definition of workflow happen only when a card passes with a human decision. Walk the talk: the workflow rides its own lifecycle.
• Every bet gets kill criteria at birth. A change we cannot falsify or revert is a mandate — say which it is and who mandated it.
• Probes cite data or stay silent. "The team should consider…" without a signal is consulting prose; delete it.
• Respect improvement WIP. If the lane already holds more open bets than the team digests, strengthen or kill existing ones before minting new cards.
• Shrinking the machinery is a legitimate bet. If cycles keep firing the low-priority rules and nothing ages, propose removing machinery — nobody proposes it unless it is allowed.
• Simulation, replay, and shadow runs are discovery evidence, never a substitute for the human decision to adopt.

Quality Gates

• Every probe run is accounted for: fired-with-citation or checked-quiet.
• Every new card has a benefit hypothesis with a mechanism, a baseline, kill criteria, and a named, timeboxed discovery.
• Delegation bets include a shadow-run design and a demotion trigger.
• Graph and policy bets include the option table with "do nothing" as a row.
• Zero direct edits to workflow.md or step contracts in this invocation.

References

• skills/flow-driven/references/loop-probes.md — the four probe families.
• skills/flow-driven/references/evolution-path.md — stages, human roles, the platform play.
• skills/flow-driven/references/measurement.md — reading the numbers honestly.
• skills/flow-driven/templates/improvement.md — the bet card.