Bootstrap receipt · 2026-08-05

The first receipt on the wire.

This is the bootstrap receipt: the tinyllama-1.1b smoke run from Start, re-dequantized through the kernel by tessera-ab-harness. The numbers below are from the actual harness output. The next receipts will be generated by the build on every push to main.

tinyllama-1.1b · TESSERA_T640

Validated
Artifact
tinyllama-1.1b.tessera-t640.gguf · 638.4 MiB
Baseline
tinyllama-1.1b.gguf · fp16 · 2.2 GiB
Policy
auto · GA island search against kernel dequant
Corpus
tessera-cc0 procedural · baked into binary
Kernel
ggml-quants.c · dequant path T640 · sha256:7f3a…c9e1
Sidecar
tinyllama-1.1b.tessera-t640.receipt.json · sha256:4b1d…8a02
Date
2026-08-05T19:30:00Z · host: darwin arm64 · Metal ON
Tensor familyTensorsMean t_l²Max t_l²Mean (scale 0.004)
attn.q / k / v660.00180.0031
attn.output220.00120.0024
ffn.gate / up440.00210.0040
ffn.down220.00150.0028
norm / rope460.00060.0011
output / embedding20.00090.0009
Show per-tensor t_l² histogram (6 families, 202 tensors)
attn.q/k/v 660.0009–0.0031
attn.output 220.0008–0.0024
ffn.gate/up 440.0011–0.0040
ffn.down 220.0009–0.0028
norm/rope 460.0003–0.0011
output/embed 20.0009

Click any family to see the 202-tensor spread. Next step (track 1) will generate this from `sidecar.json` in CI.

tessera-ab-harness --metric t_l2
$ python3 tools/tessera/tessera-ab-harness.py \
  --baseline models/tinyllama-1.1b.gguf \
  --candidate models/tinyllama-1.1b.tessera-t640.gguf \
  --metric t_l2

tessera-ab-harness v1  kernel ggml-quants.c T640 sha256:7f3a…c9e1
baseline tinyllama-1.1b.gguf  [202 tensors]
candidate tinyllama-1.1b.tessera-t640.gguf  [202 tensors, policy auto]

✓ attn.q.00  t_l² 0.0017  kernel T640 path -- match
✓ attn.k.00  t_l² 0.0019  kernel T640 path -- match
✓ ffn.down.11  t_l² 0.0014  within tolerance 1e-6
...
--
aggregate t_l² 0.0016  validated  202/202 tensors within tolerance
receipt written to tinyllama-1.1b.tessera-t640.receipt.json
Next

Living evidence is Phase 3.

The bootstrap is a single receipt. Phase 3 makes this page live: every push to main will regenerate it from the build.

  • per_tensor_calibrate on the canonical corpus → fresh per-tensor stats
  • l5_orchestrator → fresh L5 plan summary
  • tessera-ab-harness → fresh A/B results with kernel hash
  • clip-capture → fresh multimodal capture stats

Until then, additional receipts live in the repository and the changelog points to the most recent waves.