Cross-model continuation · one selected public session

Retrieve the history.
Check the delivered program.

Qwen3-4B continues a recorded Qwen3-8B implementation, with or without agent-requested Funes MCP retrieval. Inspect what was recovered, what the agent did, and what independent execution actually accepted.

The task normalizes coordinates and clocks in robot recordings. Review the centimeter and millimeter coordinate checks alongside the protocol outputs: the diagnostic checks responses; the independent grader checks the numbers.

__RESOLVED__ / 6tasks fully resolved
__RETRIEVED__successful native retrieval results
__CHANGED__ / 6starter programs changed

The supplied program scores __BASELINE__ in a separate preflight under the current grader. All six attempts remain visible, including unchanged programs and errors. A successful retrieval or an agent's finish signal is separate from task acceptance. Inspect the starter check.

Follow the actual retrieval route

  1. Local model
  2. JSONL recorder
  3. Scoped MCP server
  4. Native Funes MCP

The experiment uses the standard MCP entrypoint provided in Skills Anywhere's source example. The recorder retains both negotiated protocol eras and native responses. Memory paths and session IDs are fixed to the reviewed source.

Selected session: __SESSION__
Source manifest SHA-256: __SOURCE_SHA__
Prior agent record · Prior program · Source inventory · Actual MCP and missing-source controls

Compare all six continuations

Both conditions share the task, prior program, initial instructions, protocol diagnostic and interaction budget. Memory tools are available only in the Funes condition. Three model seeds are paired in alternating order.

__ROWS__
One public development task · independent execution under two public evaluation seeds
Condition / seedScoreTaskRetrievedWritesCommands
__CARDS__

Read operation counts in context

Command matches compare exact strings after trimming surrounding whitespace. Write matches compare normalized workspace paths and content hashes. The report separates matches with the prior session from repeats inside a continuation. Repeating a diagnostic can be necessary; fewer commands can also mean skipped checks. These counts do not estimate wasted work or human time saved.

Methods, failure controls and limits

Each attempt has up to 12 model turns, 4,096 generated tokens per turn and a 600-second interaction budget. Bridge startup and final grading are recorded separately. The same public protocol probe checks persistent JSONL responses in both conditions; numerical correctness comes from the independent grader.

Preflight controls exercise real Funes retrieval, an empty range, a rejected memory override and deliberate removal of the selected public Parquet. They are scripted checks, not extra agent trials. The exported prior session stores historical message strings; it does not reconstruct dedicated Funes tool-result blocks.

This is one public development task and one selected prior session. It does not establish general memory efficacy, held-out performance or native session restore in branded agent applications. Earlier handoff trials used pre-injected context and remain a separate experiment. No retries or successful-attempt filtering are used in this cohort.

Preselected plan · Every result · Reviewed model files · Retrieval model cache check · Offline verification · Record checksums