The censored figure is not a cost: 502ms and 2902ms render as the same number

[gunbai-bot]

v2.workflow.floor_cost_debt already states that an interrupted claim's reported figure is a LOWER BOUND describing when the deadline fired. That sentence is true, and it was not enough: three readers independently reasoned from the cluster's shape as though the figures were point estimates, and both live decisions on this axis were argued from that reading.

This adds the executed measurement that refutes it. No ceiling moves. No roster row is added or removed. No code changes.

The instrument

A controlled pair of floor runs differing only in the ceiling — same tree, so the join is a measurement rather than two observations of two states:

• censored — run 33191118450 (session/witty-wren-148), 500ms CPU
• uncensored — run 33190645319 (probe/ceiling-uncensor-9517), ceiling raised

Both emit required_floor_claim_cost.tsv. The table is that artifact's own cpu_ms column joined on identity.

The result

identity (tail)	censored	true
cost_coverage_dag_logic_has_three_fn_bodies	506	2897
cost_coverage_file_row_totality_holds_on_smoke	506	2907
cost_coverage_v2_ingested_body_is_not_decl_facts_skeleton	502	2902
green_control_sanctioned_reader_body_not_flagged	507	2227
emit_on_demand_classical_not_native_one_build_holds	502	2374
same_grammar_parse_ingest_bridge_holds	527	1658
nominal_distinct_control_compiles_ok	505	1596
lens_closure_question_zero_holds_live	1655	1713
frontier_cover_of_live_extdeps_tree_holds	501	559
operation_argv_corpus_every_row_is_materialized_or_counted	502	520
witness_belt_plan_equals_step_composition	502	355
witness_belt_plan_paused_holds_on_authority	502	380

502 -> 2902 and 502 -> 355 are the same reported number. One row costs nearly six times the ceiling; the other is comfortably under it and was interrupted only because the fold's charged clock crossed the line during it. The reported figure does not distinguish them, and no amount of care in reading it can — the information is not in the column.

What it refutes, by name

• "membership is decided by run-to-run variance" — Restore the operator's 500ms per-claim ceiling; freeze the over-cost population as a monotone debt contract #9517's own PR body, over four bootstrap_footprint siblings at 492/502/502/506ms. True of those four (they measure 460ms uncensored) and false of the seven rows beside them. The reasoning could not tell the two groups apart, so it was right by luck on the group it was applied to.
• "26 of 44 sit within 2% of the 5000ms ceiling, so scheduling noise decides them" — the handoff for the 5000ms-era population. Same shape one ceiling up, same refutation: those 26 true costs are unmeasured and unbounded above.

The agreement was not corroboration. Three readers reached the marginality reading independently, which reads as confirmation and is not: a shared input is a shared potential defect, and all three were reading one column that cannot express the distinction they were drawing. That is this repository's own recognition rule applied to its own authors.

What it does not license

Not an argument to move the ceiling — if anything the opposite, since the rows nearest the line are the ones most likely to be far over it. Not an argument to enrol or release any row: the four bootstrap_footprint siblings measure under the ceiling uncensored and are correctly not on this roster. It settles one thing — a censored figure may not be used as a cost, and a decision needing a cost for an interrupted row must uncensor and re-measure, as the pair above did.

The construction is declared and deliberately not built

write_required_floor_claim_cost_tsv renders a censored lower bound into the same cpu_ms/wall_ms columns a completed row's exact measurement occupies, distinguished only by the adjacent verdict_reached column — an exact discriminator (false for exactly the 12 budget_interrupted rows and nothing else) that all three readers failed to join. Making the bound unable to inhabit the measured-cost slot is the DESIGN §5 construction move. It is not landed here: forward implementation is frozen under the 2026-08-28 wind-down. TRIGGER: when that freeze lifts, a censored figure gets a rendering a numeric reader cannot silently consume, and this paragraph's warning becomes the regression control for it.

Transcription, declared

The figures are transcribed rather than named-by-instrument, as a declared exception: the probe branch is deleted, so no producer on this side of the boundary can re-derive the uncensored half, and the subject of the row is the divergence between the two runs rather than either run's state.

Two side findings, verified, not this diff's subject

• All 44 identities from this session's brief are already rostered in floor_cost_debt on main (Restore the operator's 500ms per-claim ceiling; freeze the over-cost population as a monotone debt contract #9517, merged 17:55) — that cause is closed.
• None of the roster's 276 identities is stale against main HEAD (module+fn resolved from each module's own module line), so stale_cost_debt should not block.
• Main is expected to stay red on ~13 blocking cost rows not in the roster.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QsmcgXPM8LdvEhUzadHkuL