Policy Coherence Kit — Multi-Stakeholder Coherence Analysis

Policy coherence analysis asks a deceptively hard question: across a set of policies, which ones reinforce each other, which work against each other, and which are most influential overall? Answering it across several decision-makers — each with their own view of every policy pair — is a matrix-bookkeeping problem that doesn't fit a spreadsheet. Policy Coherence Kit captures each decision-maker's judgments, aggregates them, and runs a suite of analytical tools over the result.

Problem

For n policies, every decision-maker fills an n × n matrix rating how each policy influences every other on a 7-point coherence scale. With multiple experts that's a stack of matrices that has to be:

• entered or imported consistently,
• aggregated into a single view under a defensible rule,
• and then analysed for structure — which policies dominate, which conflict, how concentrated each one's influence is.

Done by hand this is tedious and easy to get wrong, especially reconciling disagreements between decision-makers.

Coherence Rating Scale

Cells are rated on a structured 7-point scale that maps words to a signed score:

Diagonal cells are permanently locked to Neutral — a policy doesn't influence itself.

Approach

Input works two ways: type policy names one per line and rate each cell with colour-coded dropdowns, or import a multi-sheet Excel workbook where each sheet becomes one decision-maker. The importer validates that row and column headers match exactly and accepts the 7 rating labels case-insensitively.

Aggregation across decision-makers offers three rules:

• Average — equal weighting; auto-selected for fewer than 3 decision-makers.
• Majority Rule — the most common rating wins; ties are flagged for manual resolution rather than silently broken.
• Weighted — assign each decision-maker an importance weight in 0–1 that must sum to 1.0.

Analysis runs per aggregation as a set of tabs:

The network view offers four layouts — Circular (baseline), Force-Directed Fruchterman-Reingold (default), Spectral (eigenvector clustering of similar policies), and Shell (high-betweenness policies in the centre ring) — so the same graph can be read for different structure.

Project management keeps multiple independent studies — each with its own decision-makers, matrices, and results — alive in a single session.

What It Delivers

• 7-point signed coherence scale with locked diagonals.
• 3 aggregation rules with explicit tie-handling and weight validation.
• 6 analysis tabs: aggregated matrix, coherence scores, entropy-based range of influence, PCA, centrality-ranked network with 4 layouts, and LLM interpretation.
• Multi-project sessions and multi-sheet Excel import.
• Export: one Excel workbook per project (all DM matrices, aggregated matrix, coherence scores, range of influence, network centrality), LLM interpretation as .txt or .pdf, and plots as PNG or PostScript.

Tradeoffs

• Tkinter canvas instead of a plotting library. PCA scatter and the network graph are drawn directly on a Tkinter canvas and computed with NumPy, which keeps the dependency footprint tiny (openpyxl, numpy, reportlab, optional Pillow) and makes PostScript export free — at the cost of writing layout and rendering by hand rather than leaning on Matplotlib/NetworkX.
• PNG export is optional. It needs Pillow; without it, plots still export as PostScript. Keeps the base install lean.
• In-memory API keys. Used for a single request and cleared immediately — safer, but re-entered per session.

What I'd Do Differently

• Persist projects to disk between sessions, not just export — the multi-project model is in memory only today.
• Add a guided conflict-resolution view for Majority Rule ties rather than flagging them for manual handling.

Known Limitations

• No automated test suite ships yet; correctness rests on the aggregation logic and manual checks.
• LLM interpretation requires the user's own API key (Groq has a free tier; OpenAI needs a paid key) and network access.
• Excel import requires exact header matching between rows and columns.

Architecture

Project  (independent study, multiple in one session)
  ├─ Matrix input  (manual dropdowns  |  multi-sheet Excel import)
  │     └─ one n×n matrix per decision-maker, diagonals locked to Neutral
  └─ Aggregation  (Average | Majority Rule | Weighted)
        └─ Aggregated matrix (−3…+3)
             ├─ Coherence scores (OI / II / WOI / WII)
             ├─ Range of influence (Shannon entropy per policy)
             ├─ PCA  (NumPy → 2D scatter on Tkinter canvas)
             ├─ Network  (centrality + 4 layouts on Tkinter canvas)
             └─ LLM tab → 7-section interpretation
                  └─ Export: Excel workbook · .txt/.pdf · PNG/PostScript

Stack