Repository navigation
A /goal that the verifier panel refuted and then verified on a later round
is a behavioral contrast between two attempts at the same objective: same
model, same tools, one gap closed. Those verdicts were written under
~/.blockrun/goals/<id>/verdicts/ and never read. Now a reflector on the
free extraction models turns the contrast into at most one lesson per goal:
the shape of situation it applies to, the operation to perform naming its
component, and the closed gap as evidence. A goal that passes on round one
writes nothing, because adoption alone is not evidence. A goal that hits
the round cap writes a confounded row that is never rendered.
Lessons are scoped to the wallet for trading goals and to the workspace
for everything else, never global. They render after the learnings block
under a 1,200-character budget, lose a confirmation when a later goal in
the same scope is refuted on round one while they were in the prompt, and
turn into a ⛔ rule after two such events. Injection patterns, addresses,
amounts and the objective's own tickers are rejected before anything is
saved, and lessons are prose only: nothing here writes hooks, skills or
scripts. FRANKLIN_NO_EXPERIENCE=1 removes the component, the harness
bench gains a no-experience config, and /learnings lists the verified
lessons (/learnings clear leaves them; /learnings clear lessons
removes them). The design follows Harness² (google-research/harness2,
arXiv 2610.14224); the prompt injection scanner moves to its own module so
the lesson path and the project-config path share it.