Skip to content

Shieldbreaks

A "shieldbreak" is a single project query — one research question about a mechanism of resistance to a cancer immunotherapy (checkpoint blockade, adoptive cell therapy, vaccine, cytokine, bispecific, oncolytic virus, or other immuno-oncology modality), with its own search parameters, extraction schema, trial table, and manuscript critique. Each lives in its own subdirectory so it can evolve independently of the others.

Active shieldbreaks

Shieldbreak Rows Last updated Critique
Treg Depletion and/or Inhibition 51 2026-04-23 critique
Tumor-associated Macrophage Depletion, Inhibition, or Repolarization 37 2026-04-24 critique

Adding a shieldbreak

Invoke the trialist_screener Claude Code subagent and answer "new shieldbreak" when prompted. The agent will:

  1. Ask for a project name and propose a slug (kebab-case)
  2. Elicit the search parameters → write to prompts/shieldbreaks/<slug>/search.md
  3. Elicit the extraction schema → write to prompts/shieldbreaks/<slug>/extract.md and data/shieldbreaks/<slug>/schema.json
  4. Run the search, screen, extract → append to data/shieldbreaks/<slug>/trials.jsonl
  5. Generate docs/shieldbreaks/<slug>/index.md with the trial table
  6. Add a row to this index page linking to the new project

Once a shieldbreak has trial rows, the companion trialist_skeptic subagent reads the ingested manuscripts in full, verifies the extracted fields against the source, and publishes docs/shieldbreaks/<slug>/critique.md — a per-paper and cross-paper methodological appraisal. The skeptic also assesses, for every paper, counter-productive mechanisms by which the intervention's MoA may undermine the shieldbreak's broader target effect (e.g., anti-CCR4 depleting CCR4+ CD8 effector-memory cells alongside Tregs) even when the proximal endpoint is met.

With the screener and skeptic runs complete, the trialist_prescriber subagent then synthesizes a top-level narrative: it reads trials.jsonl and critiques.jsonl, groups rows into interventions or intervention classes under PI-confirmed granularity, and for each selected intervention (3–7 total) writes a short section covering evidence base, likelihood of the desired effect, toxicity profile, counter-productive mechanisms, practical considerations, and a ranked-position rationale. The output is published as docs/shieldbreaks/<slug>/scope_summary.md and inlined at the top of the shieldbreak page under the ## Scope summary heading; a Ranked prioritization table at the end weighs Likelihood × Toxicity × Counter-productive MoA into an Overall column. Rankings reflect the shieldbreak's Target effect as written; re-scoping it toward a downstream goal shifts the weights.