lyndonkl/wc-fitness-eval

wc-fitness-eval — risk-adjusted, rank-relative fitness

Implements footballfantasy/context/frameworks/fitness-function.md. The whole evolution engine selects on this number, so it must encode the two things that make fantasy fitness not raw points: rank is relative (ownership matters) and the rank objective sets the variance target (the selection-pressure dial).

The formula

fitness(c | θ) =   raw_xEV(c)                    # Σ player xEV (fixture-scaled) + captaincy uplift
                 + cs_corr_bonus(c)              # within-team clean-sheet correlation (BB2 stacks)
                 + progression_carry(c, horizon) # discounted future-round value of deep-team players
                 + ownership_leverage(c | θ)     # rank term — sign/shape set by θ
                 − variance_term(c | θ)          # selection pressure — sign flips with θ
                 − feasibility_penalty(c)        # 0 after clean repair; disqualifying if infeasible

Workflow

- [ ] 1. Read inputs: candidate (with block tags), round signals (player-ev, clean-sheet, fixture, ownership), θ and k
- [ ] 2. raw_xEV: sum per-player xEV (from wc-player-ev signals); add captaincy uplift (from wc-captain-ladder)
- [ ] 3. cs_corr_bonus: from wc-clean-sheet-model for each defensive stack
- [ ] 4. progression_carry: Σ player_xEV_future × P(team advances)^rounds, discounted; weight up in KO phases
- [ ] 5. ownership_leverage: apply θ (cover high-EO when protect; overweight low-EO ceilings when gain)
- [ ] 6. variance_term: estimate candidate variance; multiply by ±k per θ
- [ ] 7. feasibility_penalty: 0 if repair-clean; large if not
- [ ] 8. Sum → fitness; emit fitness + the full decomposition

Step detail

raw_xEV — sum the xEV field from each player's player-ev signal (already fixture-scaled, minutes-weighted, with the 60' tier and card/concede downside). Captaincy uplift = the expected extra from the captain ladder (from wc-captain-ladder), i.e. E[best ladder-reachable captain outcome], not a naive 2× of the top player. For a squad-build (no fixed matchday yet), use the round-1 captaincy estimate.

cs_corr_bonus — for each Clean-Sheet Spine stack (BB2), wc-clean-sheet-model returns the joint-distribution bonus capturing that a 3-defender + GK stack hauls together. Sign by θ: amplify the correlated-haul upside when gain; lightly discount when protect (they also blank together).

progression_carry — Σ_player xEV_typical(player) × Σ_{r=next..horizon} P(team alive at r) × discount^r. A steady player on a likely semi-finalist accrues several future rounds; a coin-flip nation's player accrues little. Weight ≈0 in the final, high in early knockouts (this is what makes A3 Progression Theorist competitive on fitness). Advance probabilities from wc-fixture-progression.

ownership_leverage(c | θ) — using effective ownership EO from wc-ownership-meta:

• θ=protect: − w · Σ_{high-EO players NOT owned} EO_haul_risk → rewards covering chalk (penalise gaps in the template).
• θ=gain: + w · Σ_{owned, low-EO, live-ceiling} (ceiling × (1−EO)) → rewards differentiation.
• θ=neutral: small symmetric term; mostly defer to raw_xEV.

variance_term(c | θ) — candidate variance from the spread of its players' point distributions + captaincy variance + stack correlation: | θ | term | |---|---| | protect | + k · variance (penalty) — damp swings, bank the lead | | gain | − k · variance written as a reward (subtract a negative) — you need the swing | | neutral | small penalty | k scales with how extreme the standing is (big lead + few rounds → high k_protect; big deficit + few rounds → high k_gain).

feasibility_penalty — 0 if the candidate passed repair (budget, nation cap, formation, 15-man shape). Large/disqualifying otherwise (keeps infeasible candidates out of the elite set; their blocks can still be harvested in crossover before they drop).

Revealed-preference soft terms — add the small bonuses/penalties accrued in manager-profile.md (e.g. "+ minutes-secure mids", "− benching the favourite nation"). Soft: they tilt, never dominate.

Output (the decomposition — always return it)

fitness: <number>
decomposition:
  raw_xEV: <n>
  captaincy_uplift: <n>
  cs_corr_bonus: <n>
  progression_carry: <n>
  ownership_leverage: <+/- n>     # under θ
  variance: <n>   variance_term: <+/- n>   # under θ, with k
  feasibility_penalty: <n>
objective: { theta: protect|gain|neutral, k: <n> }
variance_band: low|medium|high

The decomposition is mandatory — the board explains options through it, and the engine needs variance_band to mandate spectrum coverage in selection.

Goodhart / substitution guard (system-dynamics.md §4, invariants.md §6)

Fitness is multi-term by design, and that is non-negotiable — the systems-thinking digest's cautionary tale (held "close before writing any agent-fitness metric"): every "maximise X" rule produces a substitution effect at a node the rule did not name. Collapsing fitness to one dominant term is forbidden; if you ever find one term swamping the rest (one weight set so high the others are noise), that is the Goodhart failure in progress — re-balance.

Emit, alongside the decomposition, a one-line substitution-watch per candidate when one term is doing most of the work: name what the candidate gave up to score there. Examples:

• captaincy_uplift dominates → "ceiling-led; check the bench/minutes floor it traded away."
• ownership_leverage dominates → "differential-led; check how many nailed starts it sacrificed."
• raw_xEV (this round) dominates → "this-round-led; check the progression_carry it ignored."

The thing that quietly degraded is usually the node the objective didn't name. Surfacing it on the artifact lets round-review run the substitution-watch across rounds before it compounds.

Guardrails

• Never return raw points as fitness. If θ isn't supplied, ask for it; do not default silently to neutral on a high-stakes round.
• Never collapse to a single metric. Multi-term always (Goodhart). The decomposition ships every time.
• The same candidate must score differently under protect vs gain — if it doesn't, the variance term is mis-wired.
• Confirm the point values in scoring-rules.md are confirmed: true before trusting absolute magnitudes; otherwise flag every fitness as provisional.