SOTA experiment: qualify agent team interchangeability before live replacement

## Finding

`Testing Interchangeability in LLM Agent Teams`, submitted 2026-09-04, challenges the assumption that role-compatible agents are operationally interchangeable. Across eight independently formed teams per setting, swapping role-matched agents changed task score little but increased communication per unit progress by 16% to 63%. In Hanabi, a swapped experienced agent was more expensive than an inexperienced replacement. Longer formation histories increased the swap penalty, while greedy decoding reduced both convention drift and swap cost.

Evidence class: originating-team measured, not independently reproduced by RuV. Paper license: CC BY-NC-ND 4.0. No source code is imported by this issue.

## RuV implication

Ruflo, Autogenous, Core Memory, Cognitum scheduling, and MetaHarness can replace agents for model routing, provider failover, restart, or scaling while preserving the nominal role contract. The missing state is compatibility with the conventions accumulated by the current team.

Task capability and coordination compatibility must be measured separately.

## Proposed reversible experiment

Define a `TeamCompatibilityReceipt` or equivalent benchmark artifact with:

* team protocol digest
* role identity
* candidate agent/model/harness digest
* team history horizon
* task success
* messages per unit progress
* tokens per unit progress
* coordination latency
* recovery episodes after swap
* convention divergence metric
* authority: none

Do not persist private notebook content in the receipt. Store digests and bounded telemetry only.

## Frozen conditions

A. no replacement

B. placebo restart of the same agent

C. role-matched replacement with no transferred coordination state

D. role-matched replacement with bounded protocol-state transfer

E. inexperienced replacement control

Use at least two real Ruflo multi-agent workflows with different coordination topology, plus one synthetic cooperative control.

## Required benchmark report

Record models, providers, harness versions, task corpus, seeds, team formation episodes, history size, sample size, task score, messages, tokens, latency, cost, variance, failures, regressions, and reproduction commands. Include provider failover, delayed messages, stale notebook state, missing memory, malformed handoff state, and one adversarial convention injection.

## Promotion gate

A compatibility-aware handoff only advances if it preserves task success within 2 absolute points of the stronger baseline while reducing post-swap communication overhead by at least 25% relative to an unqualified swap. Added preflight latency must remain below 5% of task step latency and no handoff may expand capabilities.

## Falsification

If a small explicit protocol summary or stateless handoff performs within variance of a learned compatibility mechanism, keep the simple protocol and reject additional model-based compatibility scoring.

## Security and governance

Team compatibility is evidence, not authority. RVM capability checks remain independent. Private team notebooks cannot be copied across tenant or authorization boundaries merely to improve compatibility. Any transferred state must satisfy existing provenance, scope, and privacy controls.

No autonomous merge, deployment, credential escalation, or irreversible migration.

The issue names no files, tests, or entry points. Begin by locating Ruflo’s multi-agent workflows and any existing benchmark or telemetry artifacts, then determine where the five frozen conditions can be measured. Done means a bounded compatibility receipt and benchmark report covering the listed workflows, controls, metrics, failure cases, reproduction commands, and promotion gate.

Written by the indexing model from the issue text.