Research packet

This packet is the deeper reading track. The general docs explain what Asterel is and how to run it; this section explains what can be claimed, what evidence supports those claims, and what still needs benchmarks or human evaluation.

It is intentionally narrower than the private design archive: it keeps claims, evidence classes, and references that a reader can inspect without publishing raw review logs, private transcripts, or operator notes.

Asterel is not presented here as a finished empirical paper. The current packet separates:

• implemented claims backed by code and regression tests;
• method claims backed by reproducible checks and fixture design;
• research gaps that need external benchmarks, ablations, or human studies before paper-level conclusions are justified.

Packet structure

• Claims — falsifiable design claims and their current evidence.
• Methodology — how claims are evaluated and what counts as evidence.
• Evidence ledger — publishable evidence classes and commands that reproduce them.
• Technical report v0.1 — artifact-report summary of the current runtime design, local gates, and harness ablation result.
• Public release roadmap — clean-snapshot release phases for publishing the repository without private history.
• Benchmark roadmap — external and local benchmark tracks required before empirical paper claims.
• Ablation plan — planned feature-removal studies for memory, affect, exposure, naturalness, and security controls.
• Experimental protocol — clean-snapshot runbook for benchmark, ablation, human-rating, and artifact publication.
• Reproducibility — environment, commands, and snapshot discipline for public release.
• Publication boundary — what can be public, what must stay private, and how internal material should be distilled.
• Research references — bibliography mapped to concepts and modules.

Current thesis

Asterel treats long-running companionship as a runtime problem rather than a prompting problem. The central thesis is that durable companion behavior requires transport-independent continuity state, governed memory writeback, surface-aware exposure control, and pre-send verification before a response becomes part of a relationship history.

The repository currently supports that thesis with architecture checks, boundary tests, memory/governance tests, response-verifier fixtures, and replay-based quality gates. It does not yet claim statistically validated superiority over external systems.