babel-epistemic
by mdiskint
Coding Agents & IDEsBrief
Prevents metacognitive poisoning in multi-agent handoffs.
Description
What This Skill Does
Babel Epistemic Convention tracks confidence levels across multi-agent handoffs. Without it, inferences made by one agent silently become verified facts by the time they reach a third agent. The skill provides a tagging vocabulary for confidence, intent, register, affect, grounds, and trajectory that travels with the work.
Confidence calibration propagates through the forwarding block without any shared infrastructure, so the receiving agent does not need the skill installed.
When to Use It
- Tagging a news-sourced inference as REPORTED before passing it to a downstream summarizer
- Preventing a derived metric from being treated as confirmed data in a financial pipeline
- Flagging regulatory constraints that must remain non-overridable across agent delegations
- Marking a four-month declining trend as ESCALATE rather than a neutral INFORM update
- Auditing a three-agent research chain for confidence laundering after incorrect outputs
Example Workflow
User asks: Research Acme Corp's potential Vanta partnership and hand off to the drafting agent.
- 1
Search for partnership mentions and find a single unverified news item
- 2
Label the claim REPORTED at 0.25 confidence, not VERIFIED_DATA
- 3
Set intent to INFORM and consider switching to FLAG_RISK given low confidence
- 4
Append the Babel forwarding block to the handoff output
- 5
Drafting agent inherits the label as REPORTED and does not promote it to verified in the draft
Handoff payload with per-assertion confidence tags intact; drafting agent qualifies the claim as unconfirmed in the output.
Installation
clawhub install mdiskint/babel-epistemicRun the CLI command, or paste the prompt into Claude Code, Cursor, Codex or whichever coding agent you use.