Versioned
Pinned, not moving
Each package is tagged to an immutable version. Re-runs land on the same tag or open a new one — published numbers never silently change.
How the evidence room works
Each entry below is a published evidence package (EVP). Every package carries a version tag, a signed manifest, and a reproduction kit that regenerates the reported numbers from raw inputs. Nothing here asks for your trust — it hands you the means to check.
Protected artifacts are excluded by policy. Only public, scrubbed evidence appears in this room. Internal architecture, private model weights, and pre-publication data are held back as background IP and never referenced by name in any package listed here.
Public evidence packages
EVP-001 … EVP-004EVP-001 · v3.2.0
Benchmark run, cold-start harness
End-to-end scoring log with per-item timing and seed record. Manifest digest sha256:9f2a4c71b0e8.
→
EVP-002 · v3.2.0
Reproduction kit, pinned environment
Container digest, dependency lockfile, and a single command that regenerates EVP-001 numbers. Manifest digest sha256:1c88de03a7f4.
→
EVP-003 · v2.9.1
Ablation table, public split only
Component-off comparisons over the released evaluation split, with variance across five seeds. Manifest digest sha256:6b0417ea25d9.
→
EVP-004 · v2.9.1
Scrub log and exclusion manifest
Line-item record of what was removed before publication and why, so the omission itself is auditable. Manifest digest sha256:d3f5a09c8412.
→
What every package guarantees
Hashed
Signed manifest
A SHA-256 manifest digest covers every file in the package. Any edit after publication breaks the hash and is visible.
Reproducible
Kit included
Pinned environment plus one command. You regenerate the reported result on your own hardware, from raw inputs.
Scrubbed
Public by construction
Protected artifacts are stripped before release and logged in the exclusion manifest. What ships is safe to share.
Observation
Hypothesis
Demonstrated result
SHA-256 pinned
Reproduction kit
The headline result lives with the numbers
Evidence packages carry the raw material. The framed benchmark result — with its labels, caveats, and comparison context — is published in the results room.
Read the headline benchmark result →
See the reported figure alongside its evidence class, seed record, and the reproduction kit that produced it. Labeled as observation, hypothesis, or demonstrated result — never overstated.