Source and Execution Evidence — Agent Lectures¶
Checked 2026-10-18 UTC. This note separates package/source observations from execution evidence. Moving documentation and model names are not immutable release pins.
Executed locally¶
The provider contract lab runs a strict local arithmetic dispatcher and real SDK clients with fixture HTTP transports:
- 8 standard-library tests: exact argument types/keys/bounds, duplicate JSON keys, unknown tools and whole-batch call budgets.
- 20 optional SDK tests: actual request serialization/response parsing, multi-call correlation, preserved reasoning items, refusals/incomplete output, turn budgets, explicit continuation signals, duplicate IDs, unsupported server-tool scope and structured-output/local validation.
- Two fixture conversations: actual tool sums and captured fictional request bodies in provider-demo.json.
- 134 fenced Python/shell blocks: parsed with
ast.parseorbash -n; syntax checks are not execution or dependency/API validation.
The validation record records actual commands, return codes, package versions, lock hash and timestamps. The 20-package hashed wheel lock was installed in an isolated CPython 3.12 Linux x86-64 environment. Anthropic 1.15.0 uses httpx2; OpenAI 3.35.1 uses httpx. Passing httpx.Client to the current Anthropic SDK, and passing its removed temperature argument, were both tested to raise TypeError.
Not executed: hosted model inference, current model quality/price/latency, Chroma deployment, OpenClaw Gateway/CLI, browser-use benchmark, local Qwen/Granite inference, live multi-agent orchestration, confidential hardware attestation or zero-knowledge proving. Fixture responses are intentionally authored data, not a benchmark dataset. The tool arithmetic is real; the model response text is fictional.
Immutable source observations¶
The machine-readable source manifest records 33 downloaded files with exact revision, URL, byte count and SHA-256.
| Project | Revision | What it supports |
|---|---|---|
| OpenClaw | 24aac3c15a01feb1946dad5b212282e78dc3941e |
Package version 2026.10.4; Node engine; current loop/writer fencing, automation/storage, SDK and Gateway schemas |
| OpenAI Agents Python | 9abc9182d06dbd28330af888306031269d57a175 |
Blocking versus parallel input guardrails and tool-side timing distinction |
| Reflex benchmark | 3528c2cc11d561599fa92df0316bbe222b50269f |
Historical author's small-sample browser/API results, operation units and cache accounting |
| Agent Skills Eval | f10ae2b3c93055cadd82c02a3b7fe4cfbd9a4645 |
Source package version 0.1.1; optional CLI case study, not executed evaluation |
| LLM workshop | 63069f99af8ce9a6ab6e261000f4f9f1daaa1ee9 |
Workshop configurations and author's M3 Pro training-time estimates |
OpenClaw's package version, wire protocol 4 and device-signature payload v2/v3 are distinct. The private @openclaw/sdk manifest is 0.0.0-private; this does not contradict separately published gateway protocol/client packages. The current SDK implements scoped artifact methods, environment methods (with delete explicitly unsupported) and tools.invoke. Exported method existence does not establish remote authorization or compatible Gateway configuration.
Model metadata and worked arithmetic¶
Qwen3.5-4B revision 851bf6e and Granite embedding revision 835ad140 were checked through config and pipeline metadata only. Both repositories declare Apache-2.0; the bundled JSON retains its upstream identity through manifest URLs/hashes. No weights or training data were downloaded.
For Qwen, the actual text config has 8 full-attention and 24 linear-attention layers. The manifest records the evaluated BF16 full-attention KV arithmetic: 32 KiB/token/sequence and 128/256/512 MiB for 4K/8K/16K. It separately records an illustrative 48 MiB recurrent-matrix term. These are calculated components, not measured process allocations. Convolution state, padding, rollback, weights, workspaces and other model/application components are excluded explicitly.
Granite's metadata specifies 384-dimensional CLS pooling followed by normalization. Changing to mean pooling would change the embedding pipeline. Encoder maximum positions do not imply that equally long RAG chunks are useful or fast.
Moving primary documentation checked¶
- Claude model lifecycle and removed sampling parameters, structured outputs, tool-result handling.
- OpenAI function calling, structured output, reasoning, agent architecture choices.
- Chroma clients and collection metrics/configuration.
- RISC Zero receipt types, Groth16 trusted setup, and EZKL.
These pages support the stated API/protocol distinctions, not a claim that every integration example was run. HTTP source checks, including blocked requests, are preserved in source-links.json. A 403 from a raw HTTP probe is recorded separately from a successful browser-tool read; it is not relabeled as a successful download.
Revalidation exercise¶
Choose one pinned source. Fetch it at its recorded revision, recompute SHA-256 and compare the exact field used in the lecture. Then inspect a newer release and write a three-line migration note: changed contract, affected example, test that would detect the incompatibility. Do not update a date or model name without checking the behavior it changes.