Skip to content

v2.0.2

Choose a tag to compare

@epicoun epicoun released this 23 Oct 11:31
· 1 commit to main since this release

Release 2.0.2

Support for Combination Treatments

AI4L can now create evidence reviews for two or more interventions given together, titled after the pattern Combining A, B & C [for goal].

The combination is treated as one intervention: the review states why the agents are given together, and every benefit and risk says whether the evidence tested the combination or a single named agent.

Already live in production on evipedia.ai

Improved Evidence Handling

The AI4L prompt was reviewed against multiple leading frontier models (Opus 5.5, GPT-6, Grok 4.7, Gemini 4 Argon & Muse Wonder Full) on about 100 evidence reviews of the same topics, and refined where the comparison showed systematic strengths/weaknesses

  • Research results: contrary trials and reviews are searched for, cited in the item, flagged ⚠️ Conflicted and weighed; evidence from a different dose, form, route, or compound is graded as indirect; High now needs more than one independent research group; retracted papers are excluded, and errata are applied

  • Weighing contrary evidence: a systematic review or meta-analysis outranks single trials, a larger trial a smaller one, and several consistent trials a single one; evidence outweighed by stronger contrary evidence is graded Low; a trial too small to detect the effect is noted, not weighed

  • Clinical trial results: effect sizes are the difference between groups with their confidence interval, uncontrolled figures are labeled, side-effect rates come with the placebo rate, and absolute figures are reported only from the source, never calculated; finished trials are no longer described as ongoing, and each ongoing trial states what its result would change

  • Interactions: published interaction studies are searched for first, and a study that found no interaction is reported as such; interactions are listed in two groups under fixed labels, documented (product label, regulator, or human study) and theoretical (inferred from a named mechanism, not studied in humans), replacing per-item "(theoretical)" markers; "Caution" and "Monitor" labels are dropped, since every listed interaction warrants them, and a severity label remains only for an absolute contraindication or an "Avoid"; thresholds for populations to avoid are sourced

  • Citations: the link text of every cited paper names its first author and year (e.g., "Kim 2025") instead of a generic "trial" or "study", so a reader can see which and how many studies a section rests on; linking and author rules are consolidated in the Hyperlinks section, and every cited paper now carries a link anywhere in the review

  • Reference values: an optimal functional range is given only with a named source; otherwise, the standard lab range or "No established target"; each biomarker is a safety check or a marker the intervention changes; thresholds, doses, and protocol parameters are cited or stated as common practice

  • Sources: regulatory and legal facts carry a link to a source that states them (a news report, labeled as such, may stand in until the regulator publishes the fact), and no cited source is overstated or stretched to a comparison it did not make; where a paper's abstract and full text disagree, the review follows the full text and notes the discrepancy

  • Completeness: an omitted benefit fails the audit only if it would be graded High, is an approved indication, is found by a systematic review or meta-analysis, or bears directly on a specific goal; an omitted risk fails if it would be graded Medium or High, is on the label, or is serious. An item that already cites contrary evidence fails only for contrary evidence that would change its grade; other findings are noted without failing, so audits can reach 100%

  • Checklist consolidated: several contradictions between items resolved, several duplicate items merged, hints now apply to authors as well

  • Sections 16 (Benefits) and 18 (Risks) reordered: items and grading scale first, then grade checks, completeness, contrary evidence, layout, magnitude, and sources.

  • Empty evidence levels: every level without items, Low and Speculative included, now carries one sentence naming the class of evidence that falls short, instead of a bare heading

New Models successfully tested

  • Google Gemini 4 Argon - using Antigravity

  • Meta Muse Wonder Full - using Muse Code

New Environments Supported

  • Google Antigravity: (agy), with the agents and MCP servers configured as the ai4l plugin in .agents/plugins/ai4l/ and the skills registered in .agents/skills.json. Its permissions live only in the global settings, so merge the provided template once — see Using CLI Environments

  • OpenCode 2: (opencode), with its own installer; the agents run as subagents and web search works out of the box — see Using CLI Environments

  • Cursor: the MCP servers come from .cursor/mcp.json (a link to .mcp.json), and the agents run as Cursor subagents from .cursor/agents/, model-free so they inherit the session's model

Updated Environments

  • Muse Code now runs the agents as its own typed subagents from .agents/agents/. Muse keeps no project-level permissions, so start interactive sessions with muse --approval-mode never — see Using CLI Environments

  • Codex now spawns the agents with an underscore task name (er-creator → er_creator), the only form it accepts, and no longer asks for approval on every MCP tool call, as each MCP server sets default_tools_approval_mode = "approve"

  • Claude Desktop now runs the four local MCP servers (d-browser, d-fetch, d-pubmed, d-clinicaltrialsgov) via npx and uvx, no Docker needed, under the names ER.md records in Verified By; the built-in Clinical Trials and PubMed connectors stay off, as in the CLIs — see Using Claude Desktop

  • Gemini CLI is no longer supported, as Google no longer serves it to individual accounts; GEMINI.md and .gemini/ are removed. Use Antigravity for Gemini models

Other Changes

  • The main prompt is renamed from prompts/AI4L.md to prompts/ER.md; update any local copies or links

  • Project instructions now live in AGENTS.md for all environments; CLAUDE.md imports it, so Codex and OpenCode now also get the shared globals and rules

  • Examples now link to the live reviews and audits on evipedia, which stay current with the prompt; the outdated examples folder is removed

  • The creator, fixer, and auditor use web search and the research MCPs as the prompt requires; their read limits apply to local files only

  • Model names: the creator and auditor take their model name and version from their system prompt instead of guessing it, without product, effort, or pricing names (e.g., Opus 5.5 / Opus, Gemini 4 Argon / Gemini-Argon); variants now also differ in the filename. A model whose system prompt names no version is recorded by its bare name (e.g., Grok for Grok Build subagents); one that names no model at all is recorded as Unknown

  • Obstacle Protocol: a run stops only when a sub-agent itself cannot continue (it crashes, hits a hard limit, or a required agent, tool, or file is missing); a failed call such as an unreachable website, a 403, a bot wall, or a timeout is recorded and retried or passed to the next retrieval tier, never a reason to stop

  • Frontmatter: initial_topic and topic are always wrapped in double quotes (6.2–6.5, 6.34); other values are quoted only where YAML requires it