TT Claude Memory

feedback_display_numbers_in_scorecards

Properties1
descriptionEvery forthcoming scorecard / decision-analysis doc must include an explicit numerical scoring matrix — per-option × per-criterion scores (0-5 typical) + weighted sum + margin to second-best + sensitivity perturbations. Verbal reasoning alone is insufficient. Reinforces / extends `feedback_always_display_full_scorecard`.

4 min read

Display numerical scoring matrices in forthcoming scorecards

Rule. Every forthcoming scorecard / decision-analysis doc / scorecard-equivalent artefact (any doc whose deliverable shape is “Question + Options + Trade-offs + Recommendation + Effects-per-option awaiting Rich-input”) MUST include an explicit numerical scoring matrix:

  • Per-option × per-criterion scores (typically 0-5 scale; configure per scorecard)
  • Per-criterion weight (typically 1.0× equal-weighting unless explicit weighting specified)
  • Weighted sum per option
  • Margin to second-best (in pp or absolute points)
  • Sensitivity perturbations (drop highest-weight criterion / drop lowest-weight criterion / flip-tie-break — does ranking hold?)

Verbal reasoning + trade-off tables alone are NOT sufficient. The numbers force discipline (effect-size test; scarcity test; double-counting test) and surface ranking-flips that prose can hide.

Why (locked Monday 27 April 2026). Rich directive in response to decision-12 + decision-13 docs which both included options + trade-offs + effects-per-option + verbal recommendation but NO weighted scoring matrix. Rich’s wording: “A. Please display numbers in forthcoming scorecards” (D-13 confirmation followed by directive applying to D-14 onwards).

This is a sharpening of the existing feedback_always_display_full_scorecard (locked 2026-04-23 — “every scorecard displayed inline in full”). The 2026-04-27 sharpening adds numerical-matrix-mandatory.

How to apply:

  • Boundary test (pre-author): “Does my decision-analysis doc include a numerical scoring matrix?” If no → STOP. Add the matrix before surfacing.
  • Boundary test (during authoring): “Have I made a recommendation backed by numbers, or by prose alone?” If prose-alone → STOP. Score the options.
  • Default scoring scheme (use unless otherwise specified):
    • 0-5 scale per criterion (0 = fails / 5 = best)
    • Equal 1.0× weighting unless explicit per-criterion weight given
    • Weighted sum = per option total
    • Margin = pp difference between top option and second-best (e.g., “5.2pp margin”)
    • Sensitivity = drop top criterion / drop bottom criterion / flip-tie-break — does ranking hold?
  • Cross-module primitive promotion scorecards (D-13 onwards) — recommended criteria set:
    • C1: AC-1 cross-cutting test (≥4-module reach)
    • C2: Introduction-module rule (cleanest module fit per feedback_introduction_module_rule)
    • C3: Phase-1.5 jurisdictional-coverage fit (8-jurisdiction stress-test)
    • C4: Probate-Alloy / Phase-1.5 LATE-STAGE compatibility (per A-118)
    • C5: Schema impact (additive vs invasive)
    • C6: Sovereignty boundary test (feedback_v2_ik_ias_build_plan_sovereign)
    • C7: Audit-evidence quality (verbatim quotes; library coverage)
  • Architecture-options scorecards — use the 17-criterion frame from criteria-frame-sharpened.md v1.0
  • Sub-decisions inside a parent decision — apply per-sub-decision scoring (one matrix per sub-decision per feedback_decision_analysis_one_at_a_time)

Companion disciplines (already locked):

  • feedback_always_display_full_scorecard — display inline in full (parent rule; this memory adds numerical-matrix-mandatory)
  • feedback_always_save_scorecards — every scorecard saved to disk
  • feedback_actively_use_t_files_in_scorecard_authoring — verbatim T-file quotes; numerical scoring sits ALONGSIDE verbatim quotes, not in place of them
  • feedback_weak_criteria_to_avoid_in_scorecards — 5-test checklist (scarcity / clean-break / double-counting / temporal-mismatch / effect-size); apply BEFORE assigning numbers
  • feedback_reframe_beats_reweight — when scorecard stalls at 85-92% with narrow margin, reframe (NOT just re-weight). Numerical matrix surfaces the stall; reframe is the response.
  • feedback_decision_analysis_one_at_a_time — surface ONE numerical scorecard at a time

Anti-patterns (avoid):

  • ❌ Trade-off table with verbal “Strong / Medium / Weak” descriptors — convert to 0-5 scores
  • ❌ Effects-per-option section without scoring — numbers go alongside effects
  • ❌ “My recommendation: Option B with five reasons” — must show Option B’s score AND second-best’s score AND margin
  • ❌ Hand-waving on sensitivity — explicit “drop c1 → ranking holds / flips” required
  • ❌ Skipping the matrix because “the answer is obvious” — Rich has explicitly asked for numbers; the obvious-ness should be visible AS the margin

Boundary cases to remember:

  • Re-confirm scorecards (e.g., decision-11 RE-VERIFY) — numerical matrix still required even if confirming an existing lock; the confirmation IS the score
  • Composite decisions (e.g., decision-13-16 covering 4 primitives) — one matrix per sub-primitive (per feedback_decision_analysis_one_at_a_time)
  • Visibility-projection docs (e.g., decision-10 seam articulation) — these are NOT scorecards (they project locked content); no matrix required
  • Memory-saves / process-directives — NOT scorecards; no matrix required

Graph View

Backlinks