TinyTitan 5.12

TinyTitan 5.12 — the pre-production audit: crash paths closed, and a verification suite that cannot pass silently

This release is the remediation pass the project ran before its first production cut: a full audit
of the tree, drained to zero open findings, plus the gates that keep it there. No kernel, model or
API behaviour was retargeted — what changed is crash paths, tests, gates and documentation. The
DeepSeek-V4.1-Flash port reference is the one new artifact.

What is new

• A DeepSeek-V4.1-Flash port reference (docs/deepseek-v41-flash-port.md,
  docs/deepseek-v41-flash-reference.md): what the checkpoint is, its byte and quantization layout,
  the integration surface it would need here, and the two honest conversion paths — re-quantize to
  the affine 4-bit/8-bit pair this runtime serves, or preserve the released FP8/FP4 weights, which
  needs two new decode kernels.
• Eleven pinned gates in tools/lint.sh, the exact command CI runs: force-cast, function
  length, unchecked-Sendable documentation, converter expert order, architecture paths, shell
  portability, shellcheck, SwiftLint, swift-format, eslint/prettier for the plugin packages, and
  ruff for Python. Each fails when its tool is missing or is a different version, so none can pass
  by skipping.
• A committed audit ledger (AUDIT/ledger.json, rendered to AUDIT/ledger.md): every finding
  with its evidence before and after, the commit that closed it, and a written reason wherever a
  rule was deliberately configured rather than followed.

What is fixed

• Every force unwrap is gone. 171 sites in sources/ and 139 in tests and benchmarks now bind,
  guard or throw a typed error: a bad state reports itself instead of taking the process down.
• Chunked prefill with a producer that cannot run it raises the existing chunkedUnsupported
  error; it previously force-cast and crashed.
• A scanner test that could never fail (aSecondScanReportsWhatJoined ended in … == false || true) now grows the fleet and asserts the joined-member path — the behaviour it was named for.
• Swift SAST runs again. CodeQL had failed before compiling a single file since the toolchain
  moved; the extraction build now completes and scans 249 of 464 Swift files with 0 alerts.
• Two CI regressions: the Markdown link check no longer reads installed dependencies' READMEs,
  and the verification host installs the Python suite's pinned dependencies instead of failing on a
  missing module.
• Formatting and lint debt paid: SwiftLint 4,479 → 0 findings under --strict, a committed
  .swift-format with a 442-file sweep, eslint/prettier with lockfiles in both plugin packages, and
  ruff pinned for the Python scripts.

Verification

• tools/lint.sh — all eleven gates, every tool pinned, clean.
• swift test --no-parallel — 1,493 tests in 223 suites; the benchmark suite — 299 tests; the two
  plugin suites — 66 (1 skipped) and 107.
• 7 of the 16 stored golden baselines compared byte-identical (qwen36-{4,8}, qwen38-4,
  qwen35-{4b,9b}-{4,8}) — every installed model that has a golden target — then a clean scratch
  release build, warning scan clean.
• the internal-speed record (benchmark/internal-speeds/v5.12.json, 4B dense): one first-pass
  reading breached the 10% gate — prefill 28.0 → 23.3 tok/s and TTFT 0.25 → 0.30 s — and the repeat
  pass (v5.12-run2.json) read 28.0 tok/s and 0.25 s with all thirteen metrics inside the gate. The
  7-token prefill is fixed-overhead dominated and this release changes no kernel arithmetic, so the
  first pass is measurement noise rather than a regression; no other metric moved past the threshold
  in either pass.
• Not checked, because their install is not under models/ and nothing may be fetched to change
  that: ornith-8, ornith-4, qwen38-8, agentworld-4, agentworld-8, katcoder-4,
  katcoder-8, qwen35-2b-4, qwen35-2b-8.

Checksum

tinytitan-5.12-macos-arm64.tar.gz sha256: ec93f640f1d49dfac02d129626117d47f38585b616645886e36d1c8a8040b5e0
tinytitan-5.12-macos-arm64.tar.gz size: 15386068 bytes