feat(web): the Evidence Matrix, read rather than computed (#708 view 2)

[eaitbrahim]

Closes #708. Closes the honest cockpit milestone.

The last of #708's four views, and the one that waited for an engine record instead of being built on a guess.

Why it is a read

Building the matrix on request costs 11.9 / 12.9 / 14.3 seconds per session on the real ledger — ~39 s of CPU for three sessions, on a page the console re-polls every 15 s. And over the ledger as a whole it doesn't run at all: build_matrix requires synchronous columns, so a PBO is only ever defined within a session whose trials share a bar count, and a page can't pick that scope without inventing an operator's decision.

So #726 made trials pbo record every field of its PBOResult, and this reads them. The console displays results an operator ran and never runs one on their behalf — pinned by a test that makes build_matrix raise if it's called.

An unrun matrix is not an empty one

Three states, which is why the report carries candidate_sessions at all: no ledger · a ledger with columns and no recorded run · a ledger with runs. The middle one is the state this ships in, and the only one that can be acted on.

The guidance names a session that actually has columns. keel trials pbo --session all is the obvious thing to print and would filter to a session literally called "all", find nothing, and refuse — teaching an operator that the page doesn't know what it's talking about. So the report counts the trials build_matrix would accept and names one of those. On the tracked ledger that's pbo-grid-entry-lookback-2026-07-20, one of three genuine candidates.

It deliberately doesn't check synchronicity, which would mean reading every series — most of the cost this module exists to avoid. Suggesting a session that might turn out ragged, and getting a clear refusal from the command, is a much smaller harm than a page that costs twelve seconds to render.

PBO carries no judgement

It's the one figure a reader will want graded, and grading it is what the rail refuses: a high PBO beside a flat, positive OOS scatter is the good outcome — a broad plateau of near-identical configurations produces high PBO by construction. trials pbo's own closing sentence says to read it alongside the degradation slope, never alone, and both cross plainly so a reader can.

The dominance flags do carry a state, because they're already verdicts — and they're three-valued: False says the distribution did not dominate, which is not what an absent field says.

⛔ No sortable column on the route and no sort key in the view. A matrix ordered by PBO is a leaderboard of overfitting scores, and cscv.py forbids PBO as a ranking key in its own source.

Two tests that weren't testing anything

test_a_recorded_run_stops_telling_the_operator_to_run_one seeded a ledger with no columns, so suggested_session was empty and the invocation was empty whatever the rule said — it passed against a page that prints the command forever. And the bool-is-an-int guard was pinned through a fixture no writer can produce; it's checked at the reader now.

Both found by mutation. So was a stale __pycache__: the mutant MIN_COLUMNS_FOR_A_RUN = 2 → 1 is the same byte length, so restoring the source within one mtime tick left Python reusing the compiled mutant. Worth knowing for any same-length mutation.

Gates

• 6,292 passed / 3 skipped; mypy clean on 446 files; ruff clean (the one E501 is pre-existing in docs/experiments/)
• 11 mutants killed: the matrix built on request · pre-Engine: persist comprehensive gauntlet artifacts (Monte Carlo quantiles, DSR inputs and score, full CSCV result) when the gauntlet runs #726 gauntlet rows read as matrices · a dominance flag read as a column count · an absent flag read as a denial · a backfilled session suggested · a one-column session suggested · the command still printed after a run · PBO gaining a judgement · the section dropped from the view · the table gaining a sort control · the route becoming sortable

🤖 Generated with Claude Code

https://claude.ai/code/session_01KZZxmspQXe5qJ9FAsG13s6