THE RANEX EVIDENCE LAB / v0.1.6

Big claims.
Hard evidence.

An AI says it's done. Ranex makes it prove it.
Explore the real checks, measured performance, and dogfood runs behind the gate.

Explore the benchmarks ↗Inside the dogfood loop ↓

REPO-SOURCED DATA · SYNCED HOURLY · RECEIPTS INCLUDED

DETERMINISTIC PROOF BOARD● RECORDED

43/43CHECKS PASSED

10 kernel areas. Every published check accounted for.

REPORT #034Sep 13, 2026

GATE EVALUATION / 100 CALLS4.34 ms

20 claims per gate · median of 3

PROOF-PILE HISTORY48 runs

Recorded at c1ca06f01

DOGFOOD LOOP34 iterations

Published benchmark ledger

REAL-WORLD AUDITS18 sessions

Commit-pinned evidence to inspect

Timing above: 4.27–4.42 ms · CPython 3.14.7 · x86_64 · 16 CPUs · Linux 7.0.0-30-generic. Single-machine measurement; pure gate evaluation excludes command execution.

01PERFORMANCE, WITH RECEIPTS

Proof has a price.
Here's the measured one.

Real workloads, real units, the full observed range. No synthetic competitor speedups. No invisible outliers.

DETERMINISTIC GATES01

100 decisions. One rulebook.

Evaluate acceptance rules against evidence, with a repeatable verdict.

4.34ms

100 pure evaluations · 20 claims per gate

4.27–4.42 ms3 runs · median

Inspect raw samples +

4.417 ms / 4.268 ms / 4.336 ms

evaluate-20claims-100calls

SIGNED EVIDENCE02

Trust the signature.

Check the evidence signature before a record can count toward acceptance.

21.26ms

200 signed records admitted · Ed25519 verification

21.19–23.31 ms3 runs · median

Inspect raw samples +

21.189 ms / 23.306 ms / 21.258 ms

admission-200

TAMPER-EVIDENT HISTORY03

Recheck every link.

Recompute the decision chain so edits to recorded history are detectable.

16.65ms

2,000 journal rows · full chain recomputation

16.28–18.65 ms3 runs · median

Inspect raw samples +

16.277 ms / 16.652 ms / 18.647 ms

journal-verify-2000

TEST BENCH

CPython 3.14.7 / x86_64 / 16 CPUs / Linux 7.0.0-30-generic
single machine, single run environment; timings are not deterministic and must not be compared as if they were

Every workload. Including the slower ones.13 measurements ↗

All published workloads, in milliseconds. Min and max are observed samples, not confidence intervals.
Workload	Median	Min–max	Runs
append 100 chained evaluation rows	455.22 ms	326.63–961.49 ms	3
append 500 chained evaluation rows	1,784.73 ms	1,707.06–1,789.74 ms	3
append 2000 chained evaluation rows	4,147.97 ms	4,086.23–6,201.15 ms	3
verify chain of 100 rows (recompute all links)	1.17 ms	1.11–1.29 ms	3
verify chain of 500 rows (recompute all links)	4.97 ms	4.50–8.42 ms	3
verify chain of 2000 rows (recompute all links)	16.65 ms	16.28–18.65 ms	3
admit 200 signed records (Ed25519 verify each)	21.26 ms	21.19–23.31 ms	3
Ed25519-sign 200 evidence payloads	11.47 ms	11.38–11.60 ms	3
Ed25519-verify 200 evidence payloads	19.78 ms	17.97–20.72 ms	3
100 pure evaluations of a 20-claim gate	4.34 ms	4.27–4.42 ms	3
select wheels over the real committed uv.lock	4.54 ms	4.36–6.61 ms	3
canonicalise 1000 medium documents	6.95 ms	6.74–8.78 ms	3
ranex --help cold start (interpreter + imports)	142.54 ms	138.49–152.43 ms	3

02BUILT TO BE CHECKED AGAIN

More history.
Still accountable.

Measured at 100, 500, 2,000 rows. Full-chain verification and append cost are separate workloads.

INTEGRITY VERIFICATION

Recompute the whole chain

● median

Journal rows → · recomputes every link, including historical decisions.

JOURNAL APPENDS

Cost per recorded decision

● derived

Batch size → · batch median ÷ rows; individual appends were not timed separately.

Same test bench as above. 3 repeats per workload. The full workload table includes sample ranges; these points show medians, not a performance guarantee.

03WHEN GREEN ISN'T ENOUGH

Break the evidence.
Watch the gate respond.

Controlled fault injections from the published proof artifact. These demonstrate specific defenses; they are not a universal detection rate.

FAULT INJECTION / TEST REMOVAL

Tests disappear.
CI stays green.

4 tests removed from the task. Compare the recorded checks on the modified tree.

Bare CIGREEN

Ranex manifest-bound gateFAIL

Read the recorded refusal ↗

FAIL  gate=landing  rule=TASK_TESTS
      tests-executed: missing test ID(s): test_txnkv.py::test_autocommit_without_transaction, test_txnkv.py::test_commit_keeps_changes, test_txnkv.py::test_commit_without_transaction_raises, test_txnkv.py::test_keys_reflects_transaction_state, test_txnkv.py::test_nested_commit_then_outer_rollback_undoes_everything, test_txnkv.py::test_nested_inner_rollback_leaves_outer_intact, test_txnkv.py::test_rollback_replays_multiple_writes_in_reverse, test_txnkv.py::test_rollback_restores_deleted_key, test_txnkv.py::test_rollback_restores_prior_value_and_absence
      subject=sha256:34bf5b5389e5e5fffefb6e11aef844e6040eb9b27d712322aea729ea4c06dac7

FAULT INJECTION / STALE EVIDENCE

Code changes.
Old proof stops counting.

Evidence is bound to the subject digest. The recorded verdict changes after the subject changes.

Original subjectPASS

Changed subject, old evidenceFAIL

Read the recorded refusal ↗

FAIL  gate=landing  rule=TASK_TESTS
      evidence bound to a different subject digest: tests-executed
      subject=sha256:36bdd85aaa74c00146a491ad6c8667c8506df3be15fa465a06de5096404eaa72

04WE RUN IT ON OURSELVES