Experimental research
ScopeLabs / research notebook
Learning rules
from worlds.
Experiments in building scoped symbolic models from observations, actions, context, and outcomes — and determining which learned structure can safely transfer into another context.
Can a system learn not only what happened, but the predicates and rules that made an outcome possible — without being given those concepts explicitly?
See current experiments ↓The working model
Observe. Induce. Test. Repair.
ScopeLabs studies operational models of bounded worlds and systems: what entities, predicates, scopes, and rules are supported by evidence.
Transfer path
Learned local structure→Rebind→Local validation→Provisional transfer
Transfer is earned, not assumed.
Current benchmark / synthetic world
The berry / poison world
A deliberately small environment where agents receive structural perceptions, context, available actions, and outcomes. There is no oracle is_food or is_poison predicate.
Safe and unsafe resources expose the same high-level action —gather(Resource). The learner must use perceptual distinctions and evidence from success or harm to discover what matters, and in which scope it applies.
RESOURCE α
mark:x / smoothgather(Resource)→ task successsame action≠same outcome
RESOURCE β
mark:y / dottedgather(Resource)→ harm / task lossExperimental record
Five questions. Including the failure.
The useful boundary is evidence versus aspiration. Results below are empirical findings from bounded synthetic environments.
| RQ | Question | Verdict |
|---|---|---|
| RQ1 | Does structural chunking accelerate induction? | Partial |
| RQ2 | Do learned predicates improve behavior? | Supported |
| RQ3 | Does learned structure transfer? | Naive transfer failed; bounded rebinding supported |
| RQ4 | Can learned predicates compose? | Supported in tested context |
| RQ5 | Can outcome feedback repair rules? | Supported |
RQ1 / structurePartially supported
Does structural chunking improve concept induction?
Structural chunks improved early sample efficiency, but did not demonstrate a meaningful final-performance advantage.
+4.0pp@ 100 samples
+2.4pp@ 250 samples
+0.2pp@ 500 samples
flatat convergence
RQ2 / predicatesSupported
Do learned predicates improve action selection?
Locally induced predicates provided behaviorally useful structure rather than merely descriptive labels.
Task success+12.9pp
Harmful actions−14.1pp
learned predicates vs. baseline behaviorRQ3 / transferFailed → reframed
Original / opaque zero-shot transfer
Failed and rolled back.
Symbol names or opaque identities were not enough to justify semantic transfer.
Transfer82.0% task18.0% harmful
Local scratch87.6% task8.4% harmful
Revised / earned perceptual structure
Supported with bounded conditions.
Explicit rebinding and local validation produced the canonical result below.
Seed 092%
Seed 196%
Seed 2100%
Safe symbolic rebinding under earned perceptual structure has been demonstrated; general semantic transfer has not.
Mechanisms: snapshot of learned structure · explicit rebinding · local validation · harm registration · statistical shadowing · scoped registration · narrow escape behavior.
RQ4 / compositionSupported in tested context
Can learned predicates compose?
Learned predicates participated in useful context-specific compositions: 94.7% success on the critical case versus 0% without the relevant composition.
Evidence for composition inside this bounded environment — not general symbolic compositional reasoning.
RQ5 / outcome repairSupported
Can outcomes repair an overgeneral rule?
berry-like → useful food+harmful counterexample→scoped exceptionOutcome evidence introduced a targeted exception without forcing complete relearning — approximately one repair per seed.
Task success89.1 comparison94.8%+5.7pp
Harmful actions10.9 comparison5.5%−5.4pp
Current boundary
What we've learned so far
Local symbolic behavior induction works.
Learned predicates measurably improved action selection.
Outcomes can repair rules.
Counterexamples narrowed overgeneral rules while preserving useful structure.
Composition works in bounded contexts.
Learned predicates formed useful higher-order contextual conditions.
Naive semantic transfer does not.
Opaque zero-shot transfer was unsafe and explicitly rolled back.
Earned transfer is promising.
Rebinding plus local perceptual validation performed strongly with zero harm in the canonical bounded experiment.
The current boundary is not learning useful local symbolic structure. The harder problem is determining when that structure is valid somewhere else.
Evidence boundary
What we are not claiming
Research credibility depends on saying where the result stops.
- General intelligence
- Autonomous ontology discovery across arbitrary environments
- Unrestricted world-to-world transfer
- Zero-shot semantics
- Robust learning from realistic distributed observers
- Learning natural social knowledge
- Repository-to-repository rule transfer
- Production readiness
- Safety guarantees outside the tested benchmark conditions
Research direction
Next questions
Conceptual direction
Learning specifications / executing specifications
ScopeLabs and Kalo approach explicit system knowledge from opposite directions.
INPUTEvents / outcomes
→RESEARCHScopeLabscandidate scoped model / rules
→GATEValidation + human governance
→INFRASTRUCTUREKaloimplementations / validators / artifacts
ScopeLabs asks: What rules can be inferred from what actually happens?
Kalo asks: Given an explicit specification, how can it be transformed into concrete implementations?
This is an architectural and research direction. ScopeLabs does not currently emit production-ready Kalo specifications.
A future operational bridge
Velota
Operational systems offer a future real-world testing ground for questions ScopeLabs studies synthetically today: bounded process rules, authoritative evidence, drift, and local exceptions. Current Velota customers are not part of ScopeLabs experimentation.
Explore Velota ↗Research record
Publications
None yet. Formal write-ups are in preparation. The results above summarize the current experimental program and should be treated as empirical findings from bounded synthetic environments, not peer-reviewed general claims.