Projects · ownership · current state

Fewer projects. Clearer jobs.

Agalmic Research now separates active research, applied practice, supporting machinery and independent testbeds. A useful idea does not become an active project merely because it has a name.

Direction

One serious empirical programme. One capability practice. Bounded testbeds. Everything else stays cheap until evidence earns more attention.

Run one serious empirical research programme, one capability-building practice, and a small number of clearly bounded testbeds. Preserve the rest as options rather than pretending every idea is an active project.

Active

What Agalmic Research is actually working on.

These are the owned workstreams currently entitled to substantial attention.

A01

Attention Allocation Under Cognitive Abundance

empirical research · Agalmic-owned flagship empirical programme

Test how scarce evaluation should be allocated when candidate generation is cheap, while preserving a temporal firewall, selection-aware observability and explicit ABSTAIN conditions.

Current result: Track A v0.1 reached ABSTAIN_PRE_OUTCOME before citations were opened because none of five primary contrasts could resolve the preregistered five-percentage-point effect. The v0.2 multi-cohort redesign is now blocked on selection-neutral historical text, not on code.

Next: Recover a neutral additional historical cohort, preferably ICLR 2021, then rerun the source and precision gates before any further bulk acquisition or citation query.

Programme source ↗ · v0.1 feasibility result ↗ · v0.2 source-feasibility result ↗

active

A02

Personal Capability Practice

capability practice · Agalmic-owned applied practice

Use abundant machine cognition against real constraints in scholarliness, software engineering and engineering leadership, while measuring what remains when scaffolding is reduced.

Current result: The Cognitive Kernel has progressed from prompt conformance to a prospective consequence protocol with sealed execution receipts. The next useful evidence is consequence-bearing matched work, not more framework elaboration.

Next: Execute the frozen consequence pilot on real task pairs and test human attention, verified completion, rework and boundary outcomes before expanding the intervention.

Practice ↗ · Cognitive Kernel protocol ↗

active

Useful testbeds without ownership blur.

Independent projects can inform Agalmic Research without becoming Agalmic-owned programmes.

R01

Nemosyne / Moneta

external testbed · Independent related project and experimental testbed

Nemosyne is not owned by Agalmic Research. Where useful, Agalmic methods can be tested against its representation, verification and discovery workflows, and findings can flow in either direction without collapsing project ownership.

Boundary: The strongest Agalmic relevance is bounded: Nemosyne can test claims about representation, verification attention, provenance and discovery workflow. It cannot validate the whole Agalmic programme.

Next: Use Nemosyne only for bounded experiments whose intervention and outcome are genuinely observable; keep its product roadmap and operational ownership outside Agalmic Research.

nemosyne.world ↗ · Nemosyne source ↗

Supporting

Machinery and precursor work that should not multiply into programmes.

These objects remain reusable, but they do not currently deserve separate frontier slots.

S01

Expert-Attention Preservation

supporting research · Agalmic-owned precursor / supporting evidence

The Phase-1 ICLR preservation benchmark showed that optimizing historical acceptance can erase disagreement and topic coverage. Its core allocation question has been absorbed into the stronger Attention Allocation programme.

Current result: At a 20% review budget, post-review score ranking preserved about 60.8% of accepted papers but only 10.0% of high-disagreement cases, compared with roughly 20% under random allocation.

Next: Preserve the result and reuse the diagnostics where helpful; do not run a parallel competing programme unless a distinct question re-emerges.

Phase-1 findings ↗

absorbed

S02

Frontier Throughput Instrumentation

measurement infrastructure · Agalmic supporting infrastructure

An append-only event model for tracking generation, selection, validation, realization and assimilation without pretending the machinery itself is a research result.

Current result: The instrument exists and can make scarcity transitions replayable, but it should earn further development only when it reduces uncertainty in an active project.

Next: Use it opportunistically inside active work; do not consume a frontier slot merely to elaborate the instrumentation.

Instrumentation source ↗

supporting

Project boundary

The repository is not the organisation chart.

Agalmic Research may study, consume or contribute to independent systems without claiming ownership of them. Conversely, useful internal tooling does not become a research programme merely because it has code and documentation.

Ownership, testbed status, evidence provenance and intellectual lineage should remain separable.

See the commitment frontier →