EXP-HARD-RAILS-V2

… pending

Hard rails v2 — semantic seed + explicit KEY IDEA extraction

2026-05-01 L6 level paper5 4 runs $0.02

new

premortem

mock

real

metrics

analyze

review

krit

done

In plain language

Researchers tested a new method to guide AI agents in a chain of tasks, aiming for them to focus on the core idea of each task rather than just its format. They found that the previous approach made agents stick to formatting rules, but the new method, which emphasizes extracting the main concept, is expected to lead to a deeper understanding and evolution of ideas throughout the task sequence. The results are still being collected to see if this new approach successfully creates a shared, underlying concept that no single agent could develop alone.

▶ Technical details

Research hypothesis

v1 pilot found: 100% structural constraints (format/JSON/sentences), 0% semantic. Cause: initial agent defaulted to format constraint (training template for "constraint" = formatting). Chain propagated unchanged for 30 rounds. Fix: (1) domain seed — start with explicit conceptual territory. (2) prompt explicitly says "extract the KEY CONCEPTUAL IDEA (not format/length/ style) as the constraint for the next agent." Prediction: SEQUENTIAL-V2 produces semantic constraints (concepts/ideas/mechanisms) that carry forward AND deepen across the chain. Latent invariant concept emerges that no single agent produced in isolation. Reference: 2601.15077 (constraint factorization → latent invariant structure).

Experimental setup

Type: factorial

Condition	Parameters
CHAIN_SEQUENTIAL_V2	chain_mode: sequential, n_runs: 2
CHAIN_PARALLEL_V2	chain_mode: parallel, n_runs: 2

Factors: chain_mode (sequential, parallel)

Parameters

n_agents: 3
n_rounds: 30
model: google/gemini-2.5-flash-lite
temperature: 0.85

hard-rails constraint-chain semantic-seed design-fix paper5-extension