Fixed skill version · actual MCP delivery · recorded GPU continuations
You are viewing unpacked evidence. Verify the included record inventory before reuse.
Carry the reviewed skill.
Check what the next agent delivers.
A recorded Qwen3-8B session loaded a robot-recording skill but left an unfinished program. Qwen3-4B receives that exact skill through MCP, with or without access to the earlier session through Funes. Follow the version, the retrieved evidence and the independently checked output.
The workflow preloads the skill before model generation in both conditions. Historical retrieval is requested by the model. These are separate actions: delivery and retrieval receipts establish what was returned; task acceptance comes from executing the delivered program.
Follow the handoff
- Qwen3-8B: original MCP load
- Selected public session and program
- Workflow: same pinned skill over MCP
- Qwen3-4B: continue and verify
Session: __SESSION__
SKILL.md SHA-256: __SKILL_SHA__
Source manifest SHA-256: __SOURCE_SHA__
The unchanged starting program scores __BASELINE__ in a separate control executed after the six continuations. All six attempts remain visible, including early finish signals and unchanged programs. Inspect the control and its timing.
Compare all six continuations
Both conditions share the prior program, authoritative task, pinned skill, protocol diagnostic and interaction budget. Only the Funes condition offers historical retrieval tools. Three public generation seeds are paired in alternating order. This is a development workflow study on one task.
| Condition / seed | Score | Task | Retrieved | Writes | Commands |
|---|
Showing all six attempts
Inspect version and source failures
A new provider process rejects changed skill bytes against the earlier pins. The native memory controls also exercise an empty result, a rejected memory override and removal of the selected source. These controls use scripts, without additional model generations.
Methods, runtime provenance and interpretation
The predecessor is the earliest published Qwen3-8B trial with a successful MCP load of this skill. Its result was already known. Each successor attempt allows 12 generations, 4,096 generated tokens per generation and 600 seconds of interaction. Skill preload, memory startup and grading have separate timing records. The public protocol probe checks JSONL responses; the grader checks numerical facts. All six attempts are retained.
During this cohort, the shared editable Python environment imported EvalArc's core from the canonical main checkout. Observations during and after execution matched all 54 core source files to the frozen copy. These observations were collected at the recorded times; the subsequent recorder now requires its own checkout at startup. During-run mapping · Post-run mapping · Frozen input checks.
The successful preloads and retrievals do not establish that the model followed the guidance. Full task acceptance is reported separately above. Earlier handoff experiments use another source and no additional skill, so their results stay separate. These records do not establish general memory or skill benefits, hidden-test performance, or native state restoration in branded clients.
Repeated commands compare trimmed strings; repeated writes compare paths and content hashes. The counts distinguish matches with prior work from repeats inside the continuation. Repeating a diagnostic can be useful, and fewer operations can mean skipped checks. The counts do not estimate human time saved.