Eighteen Factors (cultural risk 0–100) and IRA (epistemic bands) — the scoring logic behind CHIKARA’s honesty layer, with a live client-side demo and a map of which products should surface which metrics.
Rodney Vianté Keliana
How culturally risky is this situation? Rate factors 0–10; weighted sum scales to 0–100 → GREEN / YELLOW / RED.
Rodney Manyakaidze
How epistemically solid is this claim? Supply probability P; map to KNOWN / INFERRED / HYPOTHETICAL / UNVALIDATED.
IRA does not invent P. The caller (human, retrieval confidence, or model) supplies probability. IRA only maps it to an action policy.
Weights normalised to sum 1.00 (manuscript originally 1.32; relative order preserved). Omitted factors = 0.
contribution = score × weight · total_risk = 10 × Σ contributions (max 100).
| # | Factor | Weight | % | Low | Severe |
|---|
| Zone | Score | Failure band (guidance) | Action |
|---|---|---|---|
| GREEN | ≤ 30 | ~5–15% | Monitor; minimal intervention |
| YELLOW | 31–70 | ~30–50% | Quarterly focus on top factors |
| RED | > 70 | ~60–90% | Immediate comprehensive intervention |
Four Prerequisites warning: if Factor #1 and #2 both score ≥ 7, flag possible Power asymmetry + Cultural mismatch + High stakes + No cultural intelligence — recommend human review.
| Label | Range | Stance |
|---|---|---|
| KNOWN | P > 0.95 | Act with confidence |
| INFERRED | 0.70 < P ≤ 0.95 | Act carefully; keep verifying |
| HYPOTHETICAL | 0.30 < P ≤ 0.70 | Do not act; investigate |
| UNVALIDATED | P ≤ 0.30 | Reject as basis for action; demand evidence |
Runs entirely in your browser using the same formulae as the Cultural Reasoning API. Optional: point at a live API base URL to POST instead.
If set, Score calls POST {base}/score and IRA calls POST {base}/ira/classify.
Drag 0–10. Preset loads a sample athlete-transition case.
Guidance only — not every surface needs both engines. Prefer honesty over decoration.
IRA labels on corpus hits (already). Eighteen Factors via Cultural Reasoning API as a sibling builder tool — document, don’t conflate with cache search.
Primary: IRA on answers · Secondary: link/score API
IRA-style confidence on guidance; escalate before advise. Eighteen Factors only if coaching cross-cultural or high-stakes social context.
Primary: IRA · Secondary: Factors if context-heavy
Optional “cultural risk checklist” for proposals aimed at community/government stakeholders. IRA if AI Assist is on (label drafts).
Optional Factors · IRA if generative
Not a natural home for Eighteen Factors. IRA only if AI Tutor is live — label theory answers; offline corpus stays honest without fake confidence.
IRA on AI Tutor only
IRA on any recommendation text (Known vs Unvalidated). Factors when advising across community authority, land, or labour contexts — not on pure agronomic numbers.
IRA on prose advice · Factors on social context
Keep transactional honesty arithmetic-first. IRA if the app explains “why this rate” with generative text. Factors rarely, unless dispute/mediation features appear.
Minimal · IRA only on generative copy
No scoring surface. Offline player stays offline — do not bolt cultural risk UI onto a library player.
Neither (by design)
Explain IRA in Safety Before Intelligence; point builders to Forge. Do not claim live Eighteen Factors scoring on the homepage unless the API is deployed and linked.
Education + links only
| Method | Path | Purpose |
|---|---|---|
| GET | /factors | All factors, weights, anchors |
| GET | /factors/{id} | Single factor |
| GET | /ira/bands | IRA band definitions |
| POST | /score | Eighteen Factors → total, zone, top contributors |
| POST | /ira/classify | Claim + P → classification |
| POST | /assess | Score + optional IRA + Four Prerequisites warning |
| GET | /health | Liveness |
Deploy separately from Forge’s corpus app unless you deliberately merge services. Restrict CORS before production.