Reproduction runs for the feasibility evidence review

The feasibility evidence review §7 re-ran three landed experiments to check the results it calls disputed or high-risk. This directory keeps the part of each re-run's collected evidence that those results rest on. Nothing here is a new experiment: every file was produced by the experiment's own harness, unchanged, at main 41c995c with 0 uncommitted inputs, on 25 September 2026.

What was run

One fixture at a time, each following its experiment's README:

Experiment	Commands	Rows / mismatches
database semantics	fixtures/bin/up; E4_OUT=<new empty dir> experiments/e4-database-semantics/run/all; fixtures/bin/evidence <E4_OUT>; copy the bundle into <E4_OUT>/bundle; E4_OUT=<E4_OUT> experiments/e4-database-semantics/run/collect-evidence; fixtures/bin/down	63 / 0
dispatch safety	the same shape with E4D_OUT and experiments/e4-dispatch-safety/run/	23 / 0
provider capabilities	fixtures/bin/up; with E5_OUT=<dir>: run/test-lib, run/all, run/collect-evidence; fixtures/bin/down	194 / 0

Every step exited 0. collect-evidence writes into the experiment's own evidence/; that output was copied here and the committed evidence left as it was.

Files

• evidence/e4-database-semantics/: cells.tsv, run.txt, bundle-summary.txt, and the S2, S5 and S7 transcripts (rows 020, 048, 061, 027 and 055).
• evidence/e4-dispatch-safety/: cells.tsv, run.txt, bundle-summary.txt, injections.log, and the fault and revocation transcripts (rows 003 and 012–017).
• evidence/e5-provider-capabilities/: cells.tsv, versions.txt, leak-scan.txt, and the row 149 transcript.
• evidence/SHA256SUMS: a digest of each file above.

Each cells.tsv keeps its harness's transcript column unchanged, so most of the paths it names are not in this directory:

• Database semantics names one file per row (transcripts/020-s5-postgres-workers-naive.txt), but its collector packs every row's transcript into one file per scenario and backend, each row after a ==> <row> <== line (transcripts/s5-postgres.txt, section 020-s5-postgres-workers-naive). The experiment's own evidence/ uses the same packing.
• Dispatch safety names the pack and section directly (transcripts/fault.txt#012-fault-netsplit-send).
• Provider capabilities names one file per row. Only row 149's is kept here.

A transcript that is not kept can be read in the same place under the experiment's own evidence/. Treat it as the committed capture's transcript, not this re-run's.

Each experiment's leak scan matched only what its report documents: for database semantics the control, its copy inside the S7 store backup, and the pattern list; for dispatch safety the control, the pattern list and the applied artifacts in E4D_OUT; for provider capabilities the 3 controls and nothing else. The first two are listed in their bundle-summary.txt, the third in its leak-scan.txt. The fixture bundles and the applied artifacts hold synthetic secrets and are not kept.

Reading them

Compare each file with the same path under the experiment's own evidence/. A harness row passes when every asserted cell matches; seen: cells are recorded, not asserted, so a row can match while a recorded value differs. The review's §7 lists every difference that bears on a conclusion.