Defend a Figure

The method lives in the GiveWell project, not here. It is the shared skill defend-a-figure of the canopy workspace project connect/givewell, so every agent and person defends a figure the same way. Load it and follow it:

project_skill "connect/givewell" "defend-a-figure"

(mcp__plugin_canopy_projects__project_skill, from the canopy plugin.) Its rule is lead with the weakest check: targeting_methodology for the exact selection, its honesty fields and four cross-checks, an outside source, the method spread, the criterion spread (targeting_compare_criteria), the traps in targeting_research, then the three paragraphs and a Sources table. Change the method there, never by copying it back into this file.

In an ACE run

• Products: <phase>/defend-a-figure_summary.md (the three paragraphs), a Sources table with a live URL per input from the methodology's own source table, and a named conservative figure where the methods disagree. build-targeting-model takes the paragraphs and the table as its document body.
• Auto: run the checks, write the defence, notify the admin group with the headline figure and any check reading worth watching or worse.
• Review: present the cross-check verdicts, the method spread and the draft paragraphs for approval before anything is published.

Change Log

Date	Change	Author
2026-08-29	Initial version — weakest-check-first rule, external validation step, method spread, three-paragraph pattern	ACE team
2026-09-01	off_method_units became inherited_units	ACE team
2026-09-07	Added the criterion spread (targeting_compare_criteria) as step 6	ACE team
2026-09-07	Corrected the ADM2 caveat: the ceiling is geoBoundaries coverage (18/55 countries), not the indicators	ACE team
2026-10-25	Method moved to the GiveWell project's shared defend-a-figure skill (connect/givewell); this file keeps ACE's products and modes	Eva