[Research]: qualify hierarchical stride control across effect, delivery, and authority boundaries

[huangruiteng]

Problem

Long-horizon agent work is often tuned as one generic "turn length" problem. In practice, three different feedback intervals are being changed at once:

1. work between model-visible environment observations, such as tool results;
2. work between light control interventions, such as scheduler wakes and durable Todo settlement;
3. work between heavy authority interventions, such as user steering, protected-operation gates, or accepted supervisor proposals.

Making all three intervals larger increases drift and delayed correction. Making all three smaller creates model roundtrip cost, fragmented Todos, repeated scheduling, and human-attention churn. A model may safely compose many read-only tool effects while still needing a narrow delivery checkpoint for a risky code change and no human intervention while goal authority remains stable.

LoopX has the relevant owners already: EffectProgram and host execution, Turn/Todo/settlement, and goal/vision/gate authority. It does not yet have one provider-neutral vocabulary or measurement model that can distinguish which layer is too short, too long, or coupled to the wrong signal.

Research direction

Define three nested strides:

• effect stride: one bounded effect program before the next decision-relevant observation returns to the model;
• delivery stride: one governed Turn that aligns a coherence slice with a bounded, evidence-backed delivery slice;
• authority stride: one autonomously terminable segment under a stable authority snapshot.

Semantic stop conditions should outrank numeric ceilings. A scheduler tick is not a semantic boundary. Replan is a transition between delivery directions, not a fourth stride. A report with no authority delta is not heavy steering.

Research questions

• Which typed observations can detect decision-relevant feedback without model self-report or prose matching?
• When does a Todo encode one stable postcondition, and when should it split or merge?
• Which model capability dimensions justify widening effect stride without widening delivery or authority stride?
• How should control-plane cost be measured alongside useful outcome, correction latency, defect escape, and recovery loss?
• Can event-driven authority escalation reduce human attention without increasing late-steering loss?
• Which cross-layer mismatches best predict long-horizon failure: observation debt, settlement lag, fragmentation tax, authority churn, authority drift, or scheduler coupling?

Smallest useful scope

The first implementation should be observation-only:

• add a public architecture RFC and stable three-layer vocabulary;
• derive a read-only, public-safe stride observation from existing Turn, settlement, Todo, quota, and gate receipts where available;
• preserve missing host metrics as unknown instead of inferring them from prose;
• add an offline evaluator for layer metrics and cross-layer mismatch signals;
• characterize current standard and fine-grained profiles;
• add at least one real model behavior qualification path using actual packets and tool schemas;
• emit only shadow recommendations.

This phase must not alter effect execution, Todo selection, quota spend, scheduler frequency, notification behavior, replan closure, or user authority.

Experiment requirements

• Pin task start state, LoopX release, host, tool catalog, and scheduler implementation.
• Do not reinstall or change policy during a run.
• Vary one stride layer at a time before running factorial combinations.
• Compare count, token, wall-time, semantic outcome, human attention, and failure mode.
• Use repeated runs (N >= 5) before promoting a profile.
• Keep raw trajectories, credentials, private task text, and internal operating evidence outside the public repository.

Qualification criteria for the observation phase

• Stable goal/agent/Todo/Turn/authority lineage where those identities exist.
• Deterministic replay produces the same observation and mismatch classification.
• No keyword, command-name, or file-count heuristic is used as semantic truth.
• Local Todo completion cannot become goal or autonomous-segment terminal.
• Reports without authority delta are not counted as heavy intervention.
• Shadow mode is provably read-only.
• English and Chinese RFCs remain semantic mirrors.

Non-goals

• One universal optimal number of tool calls, Todos, turns, or minutes.
• Forcing every host to use programmatic tool calling.
• Replacing EffectProgram, Turn settlement, Todo, replan, goal vision, or user gates.
• Granting authority to a scheduler, supervisor, model score, or adaptive policy.
• Building an adaptive runtime before measurement and stable baselines exist.

The draft RFC will be linked here when its review PR is opened.

[huangruiteng]

Draft RFC review is open in #3204. The PR is documentation-only and keeps the first implementation milestone observation-only/shadow-only; runtime control changes remain gated on repeated qualification.