Skip to content

Releases: tensor-systems/droste

v0.26.0

Choose a tag to compare

@github-actions github-actions released this 20 Aug 22:42
79d41d5

What's Changed

  • Streamline launch documentation by @shanev in #212
  • Streamline and verify documentation by @shanev in #213
  • Retitle the OG card tagline to the public positioning by @shanev in #214
  • Make ModelRelay a first-class provider by @shanev in #215

Full Changelog: v0.25.0...v0.26.0

v0.25.0

Choose a tag to compare

@github-actions github-actions released this 04 Aug 18:28
315f69c

What's Changed

  • Converge the release instead of executing it once by @shanev in #209
  • Make a run say what it did without by @shanev in #210

Full Changelog: v0.24.0...v0.25.0

0.24.0

Choose a tag to compare

@shanev shanev released this 04 Aug 00:49
24d9bf9

Trace ABI v9 reports which ready-time gates a run armed

startup gains an optional ready_gates array naming the gates the run enabled: any of policy_hints, ready_metadata_validator, ready_answer_validator. An empty array means the run enforced nothing.

Enforcement is opt-in and defaults to off, so a host that never passes policy_hints runs with ready_violations and contract_violations permanently inert — and until now emitted a trace identical to a fully gated run. One embedder shipped months in that state while its benchmarks passed hints and measured the enforced configuration; nothing could fail, because nothing recorded which system was running.

Embedders should assert on ready_gates in whatever checks their production configuration, and treat an unexpectedly empty array as a misconfiguration rather than a default.

The v8 → v9 rename moves the trace_v8_* conformance helpers and trace-v8-* fixtures to trace_v9_* / trace-v9-*.

ready_answer_validator gates on observed run state

run_rlm accepts an optional ready_answer_validator, called with a ReadyAnswerState (metadata, answer content, iteration, calls made, successful calls) instead of metadata alone.

ready_metadata_validator sees only model-written metadata, so it can verify the model's claims about its work but not the work. Behavioural rules — "do not answer without consulting the data" — need what the engine observed. The older hook is unchanged and still supported; both run when both are supplied, and violations from either revoke readiness identically.

See UPGRADING.md for full notes.

v0.23.0

Choose a tag to compare

@github-actions github-actions released this 03 Aug 22:15
12d0841

What's Changed

  • Report liveness while a provider call is in flight (Trace ABI v8) by @shanev in #205
  • Release 0.23.0 by @shanev in #206

Full Changelog: v0.22.0...v0.23.0

v0.22.0

Choose a tag to compare

@github-actions github-actions released this 03 Aug 19:47
eb3c95f

What's Changed

  • Answer-state checkpoints, and finishing an interrupted run into an answer by @shanev in #203
  • Release 0.22.0 by @shanev in #204

Full Changelog: v0.21.1...v0.22.0

v0.21.1

Choose a tag to compare

@github-actions github-actions released this 27 Jul 14:05
4cbc50c

What's Changed

  • Preserve stop reasons in streamed unary responses by @shanev in #202

Full Changelog: v0.21.0...v0.21.1

v0.21.0

Choose a tag to compare

@github-actions github-actions released this 27 Jul 06:12
98e686f

What's Changed

Full Changelog: v0.20.1...v0.21.0

v0.20.1

Choose a tag to compare

@github-actions github-actions released this 25 Jul 03:02
bc259c9

What's Changed

  • Validate ready answer metadata before completion by @shanev in #198

Full Changelog: v0.20.0...v0.20.1

v0.20.0

Choose a tag to compare

@github-actions github-actions released this 24 Jul 19:24
bf4aebf

What's Changed

  • Emit cumulative usage at root and subcall boundaries by @shanev in #196

Full Changelog: v0.19.2...v0.20.0

v0.19.2

Choose a tag to compare

@github-actions github-actions released this 20 Jul 17:31
afa62bb

What's Changed

Full Changelog: v0.19.1...v0.19.2