Summary / The 27 components
23 Workflow and state machine
Controlling which phase may happen next, instead of letting the model improvise the process.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
What is built now
driver.py: the Round 1 graph INTAKE → RECON → LOCALIZE → REPAIR×N → VERIFY → [REVIEW] → ACCEPT → MEMORIZE → DONE, with REJECT on budget or no patch. Control is deterministic, model calls happen only inside REPAIR, and every transition writes a ledger row (about 36,000 so far).
What is missing
Artifact freshness (a new commit stales old observations), the graph checker test (every state reaches a terminal, every cycle has a budget), and the full 18-state graph.
Reviewer Claude, 2026-10-04
Reviewer's read: Round 1 graph built
The 09-06 review's biggest systemic finding was that no spec owned the agent loop; four of them claimed it. Settling it (the driver owns the states, pi owns the turn inside one) made everything else buildable. Two of the states (RECON, LOCALIZE) are still stubs, so in practice the graph is "run N workers, verify, accept".
How it is measured built (R1)
Graph checker in the unit tests; ledger completeness audited per bench run.
Adoption rules shared by every arm: How the tests work.
Research & attempts
Score history
No scores yet. The first entry is the baseline for suite v1.
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.