AI-assisted classification, grounded in Porter & Kramer's shared value framework.

Kestrel is rule-based and intentionally conservative — we prioritise false negatives over false positives. Only high-confidence, evidence-backed passages qualify as true shared value.

{steps.map((s,i)=>(

{s.n}

{s.k}

{s.v}

{i<3 &&

}

))}

Where AI helps, and where it doesn't. Language models do the heavy lifting on segmentation and contextual adjudication. The classification rubric itself is human-authored, versioned, and applied deterministically — no model is asked to decide what qualifies as shared value.

Four CSV criteria

{[ {k:'Primary label', v:'Positive, Ambiguous, or Negative. Passage-level classification against the full rubric.'}, {k:'Initiative type', v:'No initiative, Transactional, Transitional, or Transformational.'}, {k:'CSV pillar', v:'Porter & Kramer: products & markets, value-chain productivity, or local cluster development.'}, {k:'Remediation flag', v:'Marks activity that is reactive (post-impact compensation) rather than value-creating.'}, ].map(c=>(

{c.k}

{c.v}

))}

Initiative type

{initiativeTypes.map((t,i)=>(

{String(i+1).padStart(2,'0')}

{t.k}

{t.d}

))}

{/* ---------- Reliability ---------- */}

Reliability · Kestrel V2.0

We ran every passage five times. The labels hold up.

Inter-run agreement was measured with Fleiss' κ across five independent classification runs per passage. Overall agreement is substantial; on the headline question — whether shared value is present at all — it is almost perfect.

Initiative type is where we see the most instability: the boundary between Transactional and Transitional is the primary source of disagreement, and those definitions are the top candidates for refinement in V2.1.

{[ {k:'Primary label (CSV present?)', v:0.822, note:'almost perfect'}, {k:'Overall (all four dimensions)', v:0.765, note:'substantial'}, {k:'Initiative type', v:0.702, note:'substantial'}, ].map((r,i,arr)=>{ // bar range is 0.60..1.00 so values scale as (v-0.60)/0.40 const pct = Math.max(0, Math.min(1, (r.v-0.60)/0.40)) * 100; return (

{r.k}

{r.note}

{r.v.toFixed(3)}

{/* Almost-perfect threshold at 0.80 → 50% of [0.60..1.00] */}

);})} {/* Scale aligned with bar: 0.60 / 0.80 / 1.00 sit at 0% / 50% / 100% of bar track. Use same flex geometry (number col 80+14, right col 14+40) so gridline positions line up. */}

0.60

Almost perfect →

1.00

{/* Run-level agreement — Full / Majority */}

Full agreement

of passages received the identical label across all five runs (5/5).

Majority agreement

received the same label in at least 3 of 5 runs — a clear majority for nearly every passage.

{/* Human agreement — mirrors the κ card: copy left, metric card right */}

Human agreement

Kestrel agrees with human raters 74% of the time.

On a held-out sample, V2.0’s classifications matched expert human labels in roughly three of every four passages. The remaining gap concentrates at the same boundaries that drive run-to-run disagreement.

Humans will quickly become the limiting factor. Raters routinely infer beyond what the report actually discloses and lean generous — awarding shared value where the evidence in front of them doesn’t support it.

V3.0 will aim to increase this value — with a refined rubric around Transactional vs. Transitional and a larger adjudicated training set.

{[ {k:'Kestrel V2.0 · today', v:0.74, tone:'var(--ink)', note:'baseline'}, {k:'Kestrel V3.0 · target', v:0.85, tone:'var(--moss-deep)', note:'targeted', dashed:true}, ].map((r,i,arr)=>(

{r.k}

{r.note}

{Math.round(r.v*100)}%

agree

))}

V2.0

100%

Framework inspired by Porter, M. E. & Kramer, M. R. (2011). “Creating Shared Value.” Harvard Business Review.

Reliability measured with Fleiss’ κ across five independent runs per passage. Fleiss, J. L. (1971). “Measuring nominal scale agreement among many raters.” Psychological Bulletin, 76(5), 378–382. doi:10.1037/h0031619.