Recorded by commits ad46cf9 through f2cb470
Experiment Note
VARA V2 was introduced as an opt-in controller without changing V1 configs or legacy runs. During the same day, cavity continuation diagnostics and safeguards were repaired repeatedly, reliable overlay configs were accepted, topology validation was added, and a strict Re=100 sanity gate was implemented.
The frozen V2 protocol used a neutral warm-up followed by fixed control blocks. The documentation records 100 neutral steps and six 50-step blocks by default, with 10 probe steps included inside each block budget.
Technical Essay
V2 formalized the method as a matched counterfactual controller:
- Diagnose reference-free channel-by-patch severity.
- Track persistence and trend, rather than responding to one noisy map.
- Propose bounded sampling, local-loss, or joint actions.
- Estimate target and guard gradient compatibility and prefilter conflicts.
- Snapshot model, optimizer, controller allocation, sampler, and random state.
- Run a neutral probe and an action probe from matched initial state.
- Accept only if the action beats the neutral trajectory on the target while protected metrics remain within Pareto tolerances.
- Commit the best safe action, or restore the neutral trajectory and shrink the trust radius.
Sampling probability mass was conserved. Patch-loss multipliers were normalized to conserve configured loss mass. Bounds enforced minimum uniform sampling, maximum patch mass, and multiplier ranges. Action memory recorded effectiveness, allowing repeated good actions to rank higher and harmful ones to be damped.
The controller source also encoded a reference firewall. Names associated with Ghia, CFD, test error, full-field error, relative L2, topology, or benchmark scores were forbidden as controller inputs. Allowed guards were training-time PDE, continuity, momentum, boundary, sparse-data, and other explicitly reference-free diagnostics.
Reading Note
Trust-region optimization and safe control influenced the V2 design. A proposed adaptation should be small, measured, expanded after reliable success, and shrunk after failure. The counterfactual probe resembles an online A/B test from an identical checkpoint: ordinary Adam progress is factored out, so the action must beat what would have happened without it.
The topology gate reflected cavity benchmark practice. A low scalar residual is not enough if the primary vortex is reversed, misplaced, or replaced by a spurious bubble. Reference-free topology and speed checks became prerequisites for treating a run as physical evidence.
Progress Update
This was the principal architecture milestone. V1 established localized intervention; V2 made the intervention budget-neutral, counterfactual, guarded, auditable, and PDE-independent. The strict Re=100 gate prevented higher-Re continuation from building on a physically invalid first stage.