llms.txt
Compact machine-facing index with loading instructions and links to the current protocols.
Machine-facing research context
Load the method, not a memory of the method.
Agalmic Research publishes its active research protocols as stable, low-noise resources that can be loaded into an LLM or agent before it starts work.
Shortest instruction
Go to agalmicresearch.org/llms.txt, load the Agalmic Research Operating Pack, and use the current research completion lifecycle for this task.
The deployed website is the current protocol source unless the human curator explicitly overrides it in the active conversation.
Entry points
Choose the smallest context that can do the job.
The compact index is suitable for discovery. The full bundle is intended for agents with large context windows.
Operating Pack
The default one-human-curator + cognitive-machine research context. Load this for ordinary Agalmic research work.
Full context
Combined operating pack, research completion prompt, review swarm, scarcity-displacement prompt and external reviewer policy.
Protocol manifest
JSON index for agents and orchestration software that need explicit identifiers, purposes and URLs.
Core modes
Load stage-specific protocols when they become relevant.
Research completion: prior art, adversarial null test, credible 1 + machine design, evidence audit, analysis, review, cost accounting, publication or retirement. Load prompt →
Cognitive review swarm: independent role-diverse review at several checkpoints, including a mandatory finished-paper adversarial panel. Load protocol →
Scarcity displacement: identify the binding human/material scarcity, search inherited abundance, test the intervention and identify the next bottleneck. Load prompt →
External cognitive reviewers: recruit model-family diversity while controlling correlated failure, confidentiality, provider data use and free-tier drift. Load policy →
Cross-model review
Parallel cognition is useful only when the errors are not all the same.
For substantive work, the review process seeks diversity across model families, providers, reviewer roles, context packets, source subsets, analytical methods and reproduction implementations.
A swarm does not vote a claim into truth. One well-supported fatal objection can block publication even when every other reviewer approves.
Before unpublished or sensitive material is sent to an external free-tier provider, apply the data-governance gate in the external reviewer policy.
Authority boundary
Machine abundance can make criticism cheap. It cannot manufacture epistemic authority.
The human curator remains responsible for deciding which claims Agalmic Research will stand behind, when a genuine domain expert is required, and whether publication is warranted.