After the Spill

How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.

View the Project on GitHub higuseonhye/research-os

Paper 003 — Research question v0.1

Status: design · not yet run · Experiment ID TBD Program L0: When reality cannot be explained by the agent's model class, how should an embodied system revise parameters and the architecture/composition of its world-modeling system? Builds on: Paper 002 description · expansion taxonomy


Central question (v0.1)

When repeated failures reveal that the agent's world model is missing not a mode but a relation between entities, does adding a prepared relation-module expansion open task capability that neither parameter repair nor a single mode-expert reaches — measured as growth of the achievable-task space, not only reduced prediction error?

Paper 002 asked whether a missing dynamic mode (drift vs static) justifies structural expansion. Paper 003 asks the next taxonomy cell — a missing causal relation (line 84/100 of the expansion taxonomy: add_relation_module) — and adds a stronger success criterion: does expansion cross capability thresholds (task variants that go from unachievable to achievable), not merely shrink error on already-partially-solvable variants.


Framing note (external vs internal — keep separate)

Two audiences read this RQ differently; do not collapse them into one sentence:

Audience Framing Use
Lab intro / advisor conversation (e.g. execution-horizon, replanning-trigger language) "When should an autonomous system revise its model class rather than repair within it?" Entry point — familiar engineering vocabulary, useful for first contact
Actual RQ this paper answers "After a structural gap is filled, does the system do something qualitatively new — not just fail less?" What gets measured and written up

The entry framing is not wrong, it is just narrower than the question being asked. See vision_narrative_v0.2.md § Domain generality for the same distinction applied to the program's long-term framing, and paper003_lit_positioning_v0.1.md for the papers behind the entry framing and exactly where they stop short of this RQ.


Sub-questions

ID Question Role
RQ-1 After L1 (parameter) and L3-mode repair both fail to absorb structured residual, does the residual signature specifically implicate a missing relation (correlates with a second entity's state) rather than a missing mode (correlates with time/regime only)? Detection / diagnosis
RQ-2 Does invoking a prepared relation-module expansion reduce residual and improve control on the novel relational task beyond L1 and L3-mode arms? Expansion validity
RQ-3 Does expansion convert task variants with ~0% baseline success into >0% success — i.e. cross a capability threshold — rather than only lowering error on variants that were already partially solvable? Core novel contribution
RQ-4 Does relation expansion avoid regression on Paper-002-style nominal and single-mode-drift tasks that do not require the relation? Guardrail

Program position

L0         Representation reconstruction under model-inadequate reality
Paper 001  Recoverability @ S (optional credential)
Paper 002  Missing dynamic mode — detect · expand · validate
Paper 003  Missing causal relation — detect · expand · validate CAPABILITY (this doc)
Later      Human-provided representation · surgical exceptions

Paper 003 is not required to follow Paper 002 sequentially in execution, but reuses its two-encounter protocol and expansion-gate logic (see method draft).


What Paper 003 is not (draft)


Doc Path
Description / method draft paper003_description_v0.1.md
Related work paper003_related_work_v0.1.md
Paper 002 description (parent method) paper002_description_wm_expansion_v0.1.md
Expansion taxonomy paper002_wm_system_expansion_v0.1.md
Naming / program roadmap NAMING.md

Version history

Version Date Note
v0.1 2026-07-31 Initial RQ draft — missing relation + capability-threshold framing