A run does not describe itself: results/ holds 27 files named by version, with no config, seed, conditions or git stamp — two runs of one version overwrite each other

[dim-sous]

What is wrong

results/ today (27 files, 33 MB, gitignored):

v5_regulation_activation_results.npz
v5_regulation_activation_results.png
v5_regulation_activation_scalars.json      (444 bytes)
v5_regulation_activation_stress_tests.png
v5_comparison.json
phase1_milp_smoke.png  …

Every name identifies a version, not a run. Nothing on disk records the config that produced a file, the seed, the horizon, the strategy, or the checkout it came from. Two runs of the same version silently overwrite each other. A figure cannot be traced to its numbers; the numbers cannot be traced to their conditions.

So results/ is generated but not reconstitutable — the exact folder kind Wilson et al. say must be deletable and regenerable, and it is neither.

Why this is the highest-leverage item

It is the single precondition for all three problems this repo has:

• reproducibility — a number with no recorded conditions is not a finding (RCWS [A7]: "a run at one named condition is that condition, not a capability");
• comparison — two runs can only be compared if each carries what it was run with;
• a Lab — the RCWS Lab is 1,300 lines of stdlib with no logic of its own because every record already carries its config, metrics, provenance and drawable panels. Built over today's results/, a Lab has nothing to show.

What it should be

RCWS's record layer, transcribed:

• One record per run, written by the harness at the end (and, later, on failure): kind, name, config (the whole experiment config from run_simulation takes 15 positional parameters and 9 config dataclasses from one file, with no runtime schema — every caller is welded to the core's shape #5), metrics, provenance (git sha, date, duration, dirty flag), and any drawable series it declares.
• Records are the currency. Every figure and every printed number reads a record; the table and the plot cannot disagree (RCWS #47).
• A record declares its own panels (Series), so a new quantity gains a figure with no renderer edit.
• One store, one drift guard (a record whose class has moved on still lists), one records view of what the project has run.
• Naming: a figure's stem is the target that drew it, enforced at the write chokepoint — the invariant that lets a view map figures to runs with no table.

Done when

• run_simulation (or its wrapper) writes a record; nothing else writes to out/.
• Existing plots read records instead of live arrays.
• sim records (or equivalent) lists every record with kind, date, sha.
• Proof: the record for the canonical synthetic day round-trips — same config in, same metrics out.

Phase 2 of the platform epic. Depends on #5 (the config it records) and #1 (where it lives).