Skip to content

Releases

Lab 0.9.19 — Nemotron lookup correction blueprint

Choose a tag to compare

@mbachorik mbachorik released this 14 Nov 20:22
a12ed58

Lab 0.9.19 — Nemotron lookup correction blueprint

The copyable host demonstrates the opt-in oMLX
correction for a schema-invalid empty-argument lookup. It allows one
authorized read and accepts only a response grounded in that read.

The dedicated manifest selects published Runtime 0.9.26, Harness 0.9.19,
Core 0.5.1, recorder 0.3.0 and store 0.4.5. The ordinary, reasoning and
guarded-recovery profiles keep their pins.

Synthetic installed acceptance covers both first-tool policies, an endpoint
ignoring the constraint, a wrong final answer and no-call resume. The release
qualifies the Lab wheel and the wheel rebuilt from the source archive.

Lab 0.9.18 — guarded workflow recovery

Choose a tag to compare

@mbachorik mbachorik released this 14 Nov 13:26
9e30ef4

Lab 0.9.18 — guarded workflow recovery

The guarded recovery blueprint shows how a trusted
host configures Runtime recovery through Harness and reads the durable
outcome. The example requires a lookup and checks the final answer against
the tool result.

The separate recovery profile pins Core 0.5.1, Runtime 0.9.25, Harness
0.9.19, recorder 0.3.0 and store 0.4.5 from released wheels. The existing
ontology and reasoning profiles retain their versions and evidence.

The installed blueprint lane covers both synthetic provider wires from
the Lab wheel and a wheel rebuilt from the source archive. It checks the
original invocation budget, exactly one authorized read, correction and
retry evidence, and resume with matching or changed host settings.
No live inference is required.

Lab 0.9.17 — catalogue ontology

Choose a tag to compare

@mbachorik mbachorik released this 13 Nov 12:41
e7a91fe

Lab 0.9.17 — catalogue ontology

Copy the catalogue ontology blueprint to attach v1
concepts and relations to existing agent and tool aliases. Explicit discovery
returns the concepts visible under the host's existing policy. The same
authored workflow reads a synthetic receipt with the attachment enabled or
disabled.

Package Version
Core 0.5.1
Runtime 0.9.23
Harness 0.9.18
Runtime PostgreSQL recorder 0.3.0
Harness PostgreSQL store 0.4.5

The two active SDK manifests pin public wheel hashes and source commits.
Lab 0.9.16's manifests, benchmark provenance and freezes are archived.
The original experimental lock still reproduces the RC example.

Attaching ontology requires no workflow readmission or store migration.
The build-only candidate matrix
passed 450 synthetic cases in each Linux x64/ARM64 and PostgreSQL 16/18
combination. It also verified cross-version recovery and both installed
package paths. The release attaches a separate qualification receipt for
the final public pins and Lab archives.

The released-SDK matrix passed 450 cases in each of four cells. Each cell also passed all 11 installed ontology cases from the Lab wheel and from a wheel rebuilt from the Lab source archive. Required integration acceptance.

Core 0.5.1 ontology format and Harness 0.9.18 ontology discovery.

Lab 0.9.16 — bounded sub-agents

Choose a tag to compare

@mbachorik mbachorik released this 12 Nov 18:59
f2337c6

Lab 0.9.16 — bounded sub-agents

Copy the bounded delegation blueprint to run a dependency analyst and reviewer from an authored YAML workflow or a conversation. Each specialist receives its own context pack. The coordinator owns the answer and can cite approved graph observations.

The release family uses Core 0.5.0, Runtime 0.9.23, Harness 0.9.17, Runtime recorder 0.3.0 and Harness PostgreSQL store 0.4.5. The store adds an owner-run migration for delegation.

The required matrix includes 24 synthetic delegation cases per Linux architecture and PostgreSQL 16/18 cell. It exercises authorization, separate packs, citations, budgets, journal replay, interruption and offline receipt adoption. Each cell requires 450 cases across 14 lanes.

Lab 0.9.15 — MCP read tools

Choose a tag to compare

@mbachorik mbachorik released this 12 Nov 01:47
4471d5d

Lab 0.9.15 — MCP read tools

The SDK family is Core 0.5.0, Runtime 0.9.22, Harness 0.9.16, Runtime recorder
0.3.0 and Harness PostgreSQL store 0.4.4. Core and both adapters retain their
published wheel bytes.

The MCP client blueprint shows how to expose a host-approved
remote read alias. The released matrix adds 24 synthetic MCP cases to each
Linux architecture and PostgreSQL 16/18 cell, bringing the required total to
426 cases across 13 lanes. Cases cover authorization, revocation, catalog
changes, journal replay and offline receipt adoption across both provider wires.

See the release gates for qualification and cleanup criteria.

Lab 0.9.14 — parallel read tools

Choose a tag to compare

@mbachorik mbachorik released this 11 Nov 14:03
c0fb8a2

Lab 0.9.14 — parallel read tools

Adopt Runtime 0.9.21, Harness 0.9.15 and PostgreSQL store 0.4.4. Retain
the original Core 0.5.0 and Runtime recorder 0.3.0 wheels.

The copyable parallel-tools blueprint demonstrates opt-in concurrency
for independent read tools. The required matrix adds 32 cases per
architecture and PostgreSQL cell, covering overlap, bounded windows,
serial barriers, cancellation, ordered evidence, citations and fenced
receipt recovery across both supported provider wires. The previous
370 cases remain required.

See the parallel tool guide for configuration and
verification commands.

Lab 0.9.13 — safe tool events and citations

Choose a tag to compare

@mbachorik mbachorik released this 10 Nov 17:21
cfcd348

Lab 0.9.13 — safe tool events and captured citations

Adopt Runtime 0.9.20, Harness 0.9.14 and store 0.4.3 while preserving original
Core 0.5.0 and Runtime recorder 0.3.0 bytes. Historical manifests and benchmark
inputs remain immutable; current adoption uses verified original public archives.

Add a complete tool-evidence blueprint and 12 logical cases across both API
protocols: 24 required cases per architecture/database cell, zero required skips.
Cover safe/denied displays, legacy returns, typed results, fabricated citations,
selected/partial evidence, journal replay, durable adoption, revoked source/history
use, guard isolation and unsupported or forged delivery. Application code uses
public SDK execution and never duplicates its model/tool loop.

All existing required gates remain. The focused lane runs on AMD64/ARM64 and
PostgreSQL 16/18 with restricted roles, explicit owner provisioning, original SDK
payload audit and verified cleanup. Citation proofs establish provenance rather
than semantic entailment; these synthetic fixtures do not establish live-model
answer quality. Parallel calls, MCP and sub-agents remain separate follow-ups.

Lab 0.9.12 — context packs

Choose a tag to compare

@mbachorik mbachorik released this 09 Nov 20:16
592ef7d

Lab 0.9.12 — context-pack acceptance

Adopts public Runtime 0.9.19, Harness 0.9.13 and Harness PostgreSQL
0.4.2, retaining original Core 0.5.0 and recorder 0.3.0 artifacts.
Complete manifests bind actual public wheel URLs, source commits and SHA256 hashes.
Earlier active manifests, packaged profiles and benchmark freezes remain archived.

The context-pack lane covers Markdown agent defaults and
workflow-node overrides, accepted service lookup records, scoped citations and
actual deferred reads, absent/partial/contradictory evidence, revocation and
foreign-scope refusal, restart adoption, default capacity and denied worker DDL.
Normal released acceptance now requires all 12 cases in every architecture/database
cell. Explicit candidates retain complete lineage and installed-byte verification.

Runtime accepts a bounded model projection with separately retained private guard
facts. Selecting model context does not silently change validation authority.
Harness owns scoped capture, accepted-record resolution and frozen context bindings;
applications own authorization and explicit schema migrations. Pack-free callers,
legacy saved records and prior published SDK trains keep their existing paths.
New context-pack records require capable readers; mixed trains are refused.

Qualification and original artifacts

All ten final-head CI jobs passed: ordinary acceptance, installed rule/feature/PostgreSQL integration, and reasoning controls.

Each of four AMD64/ARM64 × PostgreSQL 16/18 cells passed 2,221 source tests and 346 required stack cases, including all twelve context-pack cases. Four installed cells passed 141 rule, 159 blueprint and 12 PostgreSQL tests each. Both reasoning cells passed 31 tests and 58 control observations. Wheel and rebuilt-source consumers, frontend/browser/native checks and cleanup passed in CI. No required case was skipped; the explicit candidate-only job is inapplicable to this public train.

Qualified head 9c7cd88d224ab155fc21283f40fca64dbcdb255e merged as 592ef7d4a547b5e8f7fa556bbe1c4a9fdcd4c295; both have tree 60a3068e1aa1d0b9aea994dc8710c6623d11b770. These wheel and source assets are the original CI bytes, with source inventory and all eleven original CI ZIP digests verified. VERIFICATION-0.9.12.json, QUALIFICATION-0.9.12.tar.gz and SHA256SUMS retain the evidence. Normal main-branch runs are recorded separately in the verification JSON and are not a duplicate publication gate.

Qualification uses synthetic loopback responses and owned fixtures. It demonstrates
SDK contracts and persistence behavior, not model answer quality or token savings.
No paid or live inference is part of this release.

All five original release assets were downloaded anonymously without credentials or proxies and matched their qualified bytes and GitHub asset digests. PUBLICATION-0.9.12.json records this verification and preservation of all 84 earlier release records and 92 earlier tags. Its SHA256 is 8cc5aafb4ffd77d2033192265f6ca40589faa78c87f2df6a8b3270ee340c7b86.

Lab 0.9.11: sent reasoning-effort blueprint

Choose a tag to compare

@mbachorik mbachorik released this 08 Nov 22:34

Lab0.9.11 — sent reasoning-effort blueprint

Adopts actual published Runtime0.9.18 with an exact public wheel URL/source/SHA256,
retaining Core0.5.0, Harness0.9.12 and original recorder0.3.0/store0.4.1 bytes.
Adds the copyable sent-effort host, guide, installed
synthetic HTTP/native acceptance and required real PostgreSQL recovery. Every
previous blueprint remains available with unchanged resource bytes.

Behavior: Result observations separate artifact requests, original profile
defaults and actual encoded HTTP controls/launched native effort argv. Retries,
tools, repairs and safe pool fallback retain bounded owned histories. Explicit
none/false, omission, zero attempts, unknown execution and truncated observations
are distinct. HTTP call IDs join existing recorder evidence without schema changes.
Sent never proves provider delivery, acceptance or compliance. Native inner calls
and vendor settings remain unobserved; no new effort support or retry allowance.

Compatibility and identity: Artifact/rule/prepared/workflow/bundle input
identities gain no fields; independent public0.9.17 goldens remain exact. Saved
private Result bytes gain additive evidence; old receipt absence is preserved,
not invented. Published Harness and PG adapters require no new DDL or republish.
Prior active manifests/provenance/freezes are preserved verbatim as pre-sent-effort;
only current SDK profile and development/evaluation freeze hashes advance.
Corpus/prompts/scoring/live evidence remain unchanged. Public views stay filtered;
observations grant no authorization, billing eligibility or consent to retry.

Release gates: Fresh exact source and isolated installed wheel/rebuilt-source
consumers; frontend/browser/native checks; all four LinuxAMD64/ARM64×PG16/18
ordinary cells, four installed rule/feature/PG cells and both reasoning jobs.
Required database lanes fail missing infrastructure and verify restricted roles
and exact cleanup. Local candidate checks do not certify the final released train.
Require exact-head overall SUCCESS, original CI ZIP/source payload/count evidence,
expected-head merge and equal tested/merged trees. Publish only original qualified
CI wheel/source archives and sanitized proof. Anonymous public downloads, fresh
wheel/rebuilt-source consumers and all independent post-merge gates must pass.
Earlier tags/releases/assets are immutable. No live inference.

Qualified source: 5e280ee7c8792110e1a380fbf10ce7d19d07be3b; merge: 94ba18812d8ac80dc6bbeb31cda2203d83403c41; equal tree: 48cbfd3bb3e3abf56f1f8c220fd4727cd386da38.

All ten pre-merge checks and overall GraphQL status passed. Each ordinary architecture/database cell passed 2,193 source tests and 334 real-stack cases (1,336stack executions total). Each installed cell passed 141 rule cases, 159 blueprint cases and 12 real PostgreSQL cases, including authentic Runtime 0.9.17 → 0.9.18 receipt recovery. Both reasoning jobs passed 31 cases and 58 control samples without live inference. All eleven original CI archive digests and source payloads were verified.

The wheel and source archive are the original successful CI bytes from run 42525783244. VERIFICATION-0.9.11.json, QUALIFICATION-0.9.11.tar.gz and SHA256SUMS bind source, SDK lineage, architecture/database counts, cleanup and artifact hashes. Independent post-merge and public-consumer evidence is added after actual completion. No live-provider certification is claimed.

Lab 0.9.10 — accounting request metadata blueprint

Choose a tag to compare

@mbachorik mbachorik released this 08 Nov 14:03

Lab 0.9.10 adds a copyable accounting-metadata blueprint for recorder prepare(intent) labels, using actual published Runtime 0.9.17 with Core 0.5.0, Harness 0.9.12 and the original PostgreSQL adapters.

Trusted run/variant labels are immutable, available before dispatch and kept out of provider bodies/headers. Saved Harness labels stay frozen through resume; receipt recovery makes no extra model call. Empty metadata preserves old accounting bytes/hashes; nonempty labels change accounting payload/private receipt identity, without changing artifact, prepared-turn or workflow fingerprint fields. Labeled records require Runtime >=0.9.17 readers. No DDL, adapter republish or change to authorization/billing eligibility.

Copy with python -m lab_runner.blueprints copy accounting-metadata ./my-accounting-host. See the guide, feature map and release contract. Five prior provenance/freeze snapshots, all earlier blueprint resources, corpus, prompts and scoring remain unchanged.

A diagnosed historical GitHub download504 now has a bounded anonymous exact-asset API fallback. Original manifests/checksums, identity/origin bounds and fail-closed cleanup remain mandatory; no retry loop or version substitution.

Qualified source 02a7301235cf89251f0f8c63dbc6a2bacb830bd1, PR #52, merged as a579039998759d0c8a573779ef42a6531d2ee41b with identical tree d2670dae74f35e935b308c1167a1be793789a448.

All ten final-head jobs and overall rollup SUCCESS:

  • Ordinary acceptance: 2,172 source cases and 334 real-stack cases per Linux AMD64/ARM64 × PostgreSQL16/18 cell; wheel/rebuilt-source/public-index gates and exact cleanup.
  • Installed feature/recovery: each cell passed141 rules,97 features and8 PG cases with zero skips.
  • Reasoning controls: both platforms passed31 tests and58 synthetic controls.

These assets are the original qualified CI wheel/source bytes. VERIFICATION and QUALIFICATION bind all11 original CI ZIP digests, actual checkout trees, counts, SDK lineage and clean payload inventory. SHA256SUMS covers package/proof assets. Public-consumer and independent post-merge proofs will be appended after verification. No live inference or production-provider certification is claimed.