agentfootprint
Version:
The explainable agent framework — backtrack a wrong answer to the exact context that caused it (evidence, not guesses). Built on footprintjs.
43 lines • 1.93 kB
JavaScript
export const removeAndRetry = {
name: 'removeAndRetry',
meta: {
label: 'Removes each piece and re-runs; keeps the ones whose removal fixes the answer',
method: 'leave-one-out ablation (counterfactual necessity)',
paper: 'CausalArmor (Kim 2026) / ContextCite (Cohen-Wang 2024)',
},
async find(input) {
if (!input.rerun)
throw new Error('removeAndRetry needs input.rerun');
const rerun = input.rerun;
const results = [];
for (const s of input.suspects) {
const r = await rerun([s.id]);
results.push({ id: s.id, recovered: r.recovered });
}
const flippers = results.filter((r) => r.recovered).map((r) => r.id);
// flippers first (causally necessary); stable for ties
const ranked = [...results].sort((a, b) => Number(b.recovered) - Number(a.recovered));
const explanation = [
`removeAndRetry: removed each of ${input.suspects.length} pieces and re-ran (${results.length} checks).`,
flippers.length
? `Removing these flipped the outcome (causally necessary): ${flippers.join(', ')}.`
: `No single removal flipped the outcome — possible over-determination (no necessary single cause).`,
flippers.length > 1
? `NOTE: ${flippers.length} pieces each flip — multiple necessary causes or incidental reroutes; narrow with a slice (rankSuspects) first.`
: '',
]
.filter(Boolean)
.join('\n');
return {
finder: 'removeAndRetry',
suspects: ranked.map((r) => ({ id: r.id })),
lead: flippers[0],
shortlist: flippers,
evidence: 'proven',
granularity: 'piece',
checks: results.length,
explanation,
};
},
};
//# sourceMappingURL=removeAndRetry.js.map