mlb-faab-sizer

MLB FAAB Sizer

Table of Contents

• Delegation Chain
• Example
• Workflow
• Common Patterns
• Guardrails
• Quick Reference

Delegation Chain

This skill is a baseball-specific orchestrator. It does NOT compute auction math inline. It composes two domain-neutral sibling skills:

Step	Who	Responsibility
1	this skill	Compute base_value from Yahoo-adjusted projection, pos fit, role certainty, urgency, season pace (with league-inflation calibration)
2	this skill	Classify target as common_value / private_value / mixed
3	auction-winners-curse-haircut	Return adjusted_valuation (Bayesian haircut for common-value)
4	this skill	Estimate N from opponent profiles
5	auction-first-price-shading	Return shaded_bid ((N-1)/N + distribution + risk adjustment)
6	this skill	Apply baseball guardrails (April 40%, speculation 20%, $1 floor, role-cert floor)
7	this skill	Emit faab_rec_bid, faab_max_bid, rationale naming both delegations

Invariant: this skill never computes (N-1)/N or a common-value haircut directly. Any change to those primitives is made in the sibling skills.

Example

Scenario: Roki Sasaki called up, likely Dodgers rotation spot. $100 FAAB remaining, week 4.

1. base_value (this skill): 28 x 0.70 x 0.65 x 1.2 x 0.7 = $10.70
2. Classify: common_value (headline prospect).
3. Invoke auction-winners-curse-haircut with (raw=10.70, type=common_value, N=6, dispersion=60) -> adjusted_valuation = $7.39, haircut 31%.
4. N estimate: 6 opponents have SP need + budget.
5. Invoke auction-first-price-shading with (true_value=7.39, N=6, dist=log-normal, risk=0.2, budget=100) -> shaded_bid = $7, shade 0.90.
6. Baseball guardrails: all clear.
7. Output: faab_rec_bid=$7, faab_max_bid=round($7.39 x 0.90)=$7. Rationale cites both sibling skills.

Full trace in resources/methodology.md and resources/template.md.

Workflow

FAAB Sizing Progress:
- [ ] Step 1: Collect input signals and budget state
- [ ] Step 2: Compute base_value (baseball layering)
- [ ] Step 3: Classify value_type
- [ ] Step 4: Invoke auction-winners-curse-haircut -> adjusted_valuation
- [ ] Step 5: Estimate N from opponent profiles
- [ ] Step 6: Invoke auction-first-price-shading -> shaded_bid
- [ ] Step 7: Apply baseball guardrails
- [ ] Step 8: Emit signal + rationale with delegation trace

Step 1: Collect inputs (ask caller if missing):

• acquisition_value ($, 1-100 scale) -- from mlb-player-analyzer
• positional_need_fit (0-100) -- from mlb-waiver-analyst
• role_certainty (0-100) -- from mlb-player-analyzer
• FAAB remaining, week number, situation label

Step 2: Compute base_value (baseball-specific layering):

base_value = acquisition_value
           x (positional_need_fit / 100)
           x (role_certainty / 100)
           x urgency_multiplier          [0.7 - 1.4]
           x season_pace_multiplier      [0.6 - 1.4 after inflation calib]

Urgency: 1.4 (closer loses job / prospect called up); 1.2 (new opportunity); 1.0 (steady); 0.8 (wave); 0.7 (speculation).

Pace base: 0.6 (Apr wk 1-4), 1.0 (May-Jun), 1.2 (Jul-Aug), 1.4/0.5 (Sept contending/eliminated). Then multiply by league_inflation_ratio from tracker/faab-log.md (see methodology.md). Min 5 valid rows; else skip calibration and flag low_calibration_data.

Step 3: Classify value_type:

• common_value: headline prospect, named closer, star off IL (same info for all teams)
• private_value: handcuff, platoon fit, punt-category-specific (only we weigh this way)
• mixed: record common/private weight

Default to common_value when uncertain (conservative; triggers haircut).

Step 4: Invoke auction-winners-curse-haircut — keyed off CONTESTED bidders, not the full field:

inputs = { raw_valuation: base_value, value_type, n_informed_bidders: n_contesters,
           signal_dispersion: 40 (default) }

n_contesters = rivals realistically expected to bid on THIS specific target (see Step 5). Do NOT pass the full plausible field here. The winner's-curse haircut (Step 4) and the first-price shade (Step 6) MUST key off different counts — otherwise one "many bidders" signal is counted twice and we systematically under-bid contested closers/call-ups to ~58% of true value (audit 2026-06-02, the verified mlb-faab-sizer double-count). If n_contesters <= 1 the target is effectively uncontested → the haircut short-circuits to 0 (treat as private-value). Consume adjusted_valuation; preserve classification_rationale. Dispersion defaults: 60 prospects, 30 established, 40 otherwise.

Step 5: Estimate the TWO bidder counts from opponent profiles (they are deliberately different numbers; they converge only when a target is genuinely hotly contested):

• n_contesters (drives Step 4 haircut) — rivals with positional_need > 50 AND faab > 20% original AND activity >= moderate AND a concrete reason to want THIS player. In our deflated market (modal winning bid $1-2) this is typically 1-2, occasionally 3-4 for a confirmed closer change. Clamp [0, 6].
• N_field (drives Step 6 shading) — the broader set who could plausibly bid. Clamp [1, 8]. Defaults: common superstar 6, role-player 3, private 1-2.

Step 6: Invoke auction-first-price-shading — keyed off the full field:

inputs = { true_value: adjusted_valuation, n_bidders_estimate: N_field,
           value_distribution: "log-normal" (MLB default),
           risk_aversion: 0.2 (bump to 0.4 for contending September),
           budget_remaining: faab_remaining }

shaded_bid becomes pre-guardrail faab_rec_bid. Set faab_max_bid = round(adjusted_valuation x 0.90).

Step 7: Apply baseball guardrails (see below). Never silently violate.

Step 8: Emit via mlb-signal-emitter. User-facing rationale MUST name both sibling skills by purpose. Validate with rubric. Minimum 3.5.

Common Patterns

1. Hot common-value call-up (early season): N=5-7, pace 0.6-0.7, haircut ~25-30%, shade ~0.80-0.85. Typical $5-$12 rec.

2. Private-value handcuff: N=1-2, haircut=0 (short-circuit), shade 0.0-0.5. Typical $1-$3 rec.

3. Closer change (mid-season, common-value): N=5-8, urgency 1.4, haircut ~25%, shade ~0.83. Typical $10-$30 rec.

4. September contender stretch target: pace 1.4, risk_aversion bumped to 0.4, shade ~0.88-0.92. Can reach 40-60% of remaining FAAB.

Guardrails

1. April 40% cap: weeks 1-13, faab_max_bid ≤ 40% of FAAB remaining. Flag april_40pct_cap_triggered.
2. Speculation 20% cap: if situation=speculation or role_certainty<30, cap at 20%. Flag speculation_20pct_cap_triggered.
3. $1 floor: if shaded_bid rounds to $0 but positional_need_fit >= 30, bid $1 (rolling-list tiebreak). Otherwise bid $0 and flag zero_bid_preservation.
4. Role certainty floor: if role_certainty < 20, force faab_rec_bid = $0. Flag role_certainty_floor.
5. Regression override: if regression_index < -30, cut faab_rec_bid by 30%. Flag regression_luck_discount.
6. Variant divergence: advocate/critic differ >30% -> take critic. Flag variant_divergence_applied.
7. Budget floor (post-July): if week >= 14 and remainder < $5, flag budget_floor_near_zero.
8. Log the decision: every computation via mlb-decision-logger (including $0 bids).

Do NOT duplicate sibling caps: the 0.9 x true_value ceiling is enforced by auction-first-price-shading; the 35% haircut cap is enforced by auction-winners-curse-haircut. Trust them.

Quick Reference

Pipeline:

base_value = acq_value x (pos_fit/100) x (role_cert/100) x urgency x pace_calibrated

adjusted_valuation = auction-winners-curse-haircut(
    raw_valuation=base_value, value_type, n_informed_bidders=N, signal_dispersion)

shaded_bid = auction-first-price-shading(
    true_value=adjusted_valuation, n_bidders_estimate=N,
    value_distribution="log-normal", risk_aversion=0.2, budget_remaining)

faab_rec_bid = round(shaded_bid)          # then baseball guardrails
faab_max_bid = round(adjusted_valuation x 0.90)

Inputs required: acquisition_value, positional_need_fit, role_certainty, FAAB remaining, week, situation label, regression_index (optional).

Outputs: faab_rec_bid, faab_max_bid, value_type, N, adjusted_valuation, shaded_bid, multipliers, guardrail flags, user-facing rationale.

Sibling skills:

• @skills/auction-first-price-shading/ -- (N-1)/N + distribution + risk-aversion
• @skills/auction-winners-curse-haircut/ -- Bayesian common-value haircut

Key resources:

• resources/template.md: Input block, output template with delegation trace, per-target brief, worked delegation flow
• resources/methodology.md: Baseball layering, inflation calibration, value-type classification, baseball guardrails. Auction math is in the sibling skills.
• resources/evaluators/rubric_mlb_faab_sizer.json: Eight-criterion rubric including Delegation Integrity.