mlb-opponent-profiler

MLB Opponent Profiler

Table of Contents

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

Example

Scenario: Monday morning of Week 5. Refresh profile for Springfield Isotopes (manager Nikolay) after observing another high-activity week (7 moves, $14 spent on adds, 4-1 matchup win).

Inputs:

team_name: "Springfield Isotopes"
yahoo_session: <authenticated chrome context>
prior_profile:                       # from context/opponents/springfield-isotopes.md (Week 4)
  archetype: balanced
  archetype_confidence: 0.50
  posterior: {balanced: 0.42, stars_and_scrubs: 0.18, punt_sv: 0.02, punt_sb: 0.04,
              punt_wins_qs: 0.03, hitter_heavy: 0.12, pitcher_heavy: 0.08,
              inactive: 0.00, frustrated_active: 0.11, unknown: 0.00}

Step 1 -- Yahoo scrape (4 pages): teams index, team page for team_id=8, draftresults, transactions filtered by tid=8. URLs in methodology.md.

Step 2 -- MLB feature extraction yields:

sp_roster_share: 0.35; closer_count: 2; sb_speed_count: 4; power_bat_count: 6
moves_per_week: 3.3; faab_spent_pct: 0.32; faab_avg_bid: 2.5
lineup_set_daily: true; trade_offers_sent: 1; trade_offers_received: 0
record_last_2_weeks: "8-2-0"; days_since_last_login: 0

Step 3 -- Invoke opponent-archetype-classifier:

archetype_taxonomy:            # see resources/template.md -- 10 archetypes
  balanced: {...}
  stars_and_scrubs: {...}
  ...
observed_features: <from step 2>
archetype_prior: <prior_profile.posterior>     # sequential: last week's posterior is this week's prior
observation_weight: 0.65        # Week 5 -- see methodology.md observation_weight_calibration
correlated_feature_pairs:
  - [moves_per_week, faab_spent_pct]            # active managers do both
  - [sp_roster_share, closer_count]             # they trade off

Step 4 -- Classifier returns:

posterior:
  balanced:            0.58    # up from 0.42 -- evidence accumulating
  stars_and_scrubs:    0.12
  hitter_heavy:        0.10
  frustrated_active:   0.08    # record is winning, so frustrated_active down
  pitcher_heavy:       0.07
  punt_sb:             0.03
  punt_sv:             0.01
  punt_wins_qs:        0.01
  inactive:            0.00
  unknown:             0.00

map_archetype: balanced
classification_confidence: 37.7      # 0.58 * 0.65 * 100 -- still just under 40
best_response_hints:
  - "Match cat-for-cat; decided on execution"
  - "Include in N-estimate for every common-value FAAB target"
  - "Active manager -- probe for fair consolidation trades"

Step 5 -- Read existing context/opponents/springfield-isotopes.md, preserve manual notes, update machine-generated sections:

The "Summary", "Apparent weaknesses", "Best response" and "Open questions" sections contain hand-authored user notes -- keep verbatim. Only these get refreshed:

• YAML frontmatter: last_updated, confidence, source_urls
• Section 2 (Archetype): archetype, archetype_confidence, archetype_evidence
• Section 3 (Category strength): cat_strength, presumed_punts, likely_pushes
• Section 4 (Behavioral): activity_level, last_active, waiver_aggression, faab_remaining, faab_avg_bid_pct, trade_propensity

Step 6 -- Emit weekly signal at signals/wk05-opponent-profiles.md:

---
type: opponent_profile_summary
date: 2026-04-20
week: 5
emitted_by: mlb-opponent-profiler
confidence: 0.65
source_urls: [<11 team pages + transactions + teams>]
---
## Archetype shifts since Week 4
- Springfield Isotopes: balanced 0.42 -> 0.58 (confidence 0.50 -> 0.66). Nikolay continues active pattern, record confirms execution.
- Jennys Team: inactive 0.78 -> 0.84 (confidence 0.80 -> 0.85). Still zero moves. Profile hardening.
- ...9 more

## New reactivity triggers fired
- Team 7 (Kenyi) -- outbid us on Wade Miley at $14; shift `faab_avg_bid_pct` up

Workflow

Copy this checklist and track progress:

Opponent Profiler Progress (per team, x 11 for all_opponents):
- [ ] Step 1: Resolve team identity (team_name -> team_id, manager_alias)
- [ ] Step 2: Scrape 4 Yahoo pages for this team
- [ ] Step 3: Extract MLB-specific features (9 features; see template.md)
- [ ] Step 4: Load prior_profile.posterior as archetype_prior (sequential update)
- [ ] Step 5: Invoke opponent-archetype-classifier with 10-archetype MLB taxonomy
- [ ] Step 6: Compute cat_strength (0-100 per cat) from roster composition
- [ ] Step 7: Read existing context/opponents/<slug>.md; identify manual-notes sections
- [ ] Step 8: Write updated file atomically (preserve manual notes)
- [ ] Step 9: (at end of batch) Emit signals/wkNN-opponent-profiles.md summary

Step 1: Resolve team identity

Input is either a single team_name (case-insensitive, fuzzy match against canonical Yahoo names) or all_opponents: true (iterate over team_ids 1-12, skipping user's own team_id). Map to team_id and team-slug (lowercased, hyphenated).

• If team_name does not match any Yahoo team, return error -- do not guess
• team-slug must match the filename on disk at context/opponents/<slug>.md (or create new file if absent)

Step 2: Scrape 4 Yahoo pages

Exact URL flow documented in resources/methodology.md. Pages needed:

• https://baseball.fantasysports.yahoo.com/b1/23756/teams -- manager, last-login, W-L
• https://baseball.fantasysports.yahoo.com/b1/23756/<team_id> -- roster + FAAB remaining
• https://baseball.fantasysports.yahoo.com/b1/23756/<team_id>/draftresults -- draft pick distribution
• https://baseball.fantasysports.yahoo.com/b1/23756/transactions?tid=<team_id> -- adds/drops/bids/trades
• Each scrape records the exact URL in source_urls (for citation)
• If any page returns 4xx/5xx or auth failure, degrade gracefully (see Guardrails #3)

Step 3: Extract MLB features

Computed from scraped pages. See resources/methodology.md for exact formulas.

• sp_roster_share, closer_count, sb_speed_count, power_bat_count (roster composition)
• moves_per_week, faab_spent_pct, faab_avg_bid (waiver activity)
• lineup_set_daily (bool -- any benched-but-MLB-starting players in last 7 days?)
• trade_offers_sent, trade_offers_received
• record_last_2_weeks, days_since_last_login

Step 4: Sequential-update prior

Read prior_profile.posterior (if supplied) and pass as archetype_prior to the classifier. This is the key sequential-Bayes move: last week's posterior becomes this week's prior.

• If no prior_profile, use taxonomy priors (Week 1 only)
• If prior_profile.posterior sums to != 1.0, normalize and flag in assumptions_flagged
• Observation weight rises with week number: Wk1-2 = 0.25, Wk3-4 = 0.45, Wk5-7 = 0.65, Wk8-11 = 0.80, Wk12+ = 0.90

Step 5: Invoke opponent-archetype-classifier

Pass the 10-archetype MLB taxonomy from resources/template.md along with the observed features, prior, and observation_weight. Do not re-implement Bayesian math in this skill. The classifier returns posterior, map_archetype, classification_confidence, best_response_hints, feature_contribution_breakdown, assumptions_flagged.

• Pass correlated_feature_pairs (see Guardrails #1)
• If classifier returns map_archetype: "inconclusive", write archetype as unknown with confidence as returned
• Do not override or reinterpret classifier output -- store it as-is

Step 6: Compute cat_strength (0-100 per cat)

Classifier returns the archetype; the 10 per-cat strength scores come from a separate roster-based estimator (this is MLB-specific and stays in this skill). Method in resources/methodology.md.

• Sum projected season totals for each cat across roster
• Normalize to 0-100 where 50 is league average
• Derive presumed_punts: cats where score < 35
• Derive likely_pushes: cats where score > 65

Step 7: Read existing file, identify manual-notes sections

Critical: this skill never overwrites manually-authored notes. See resources/methodology.md.

• Open context/opponents/<slug>.md; if not present, emit new file from template
• Preserve Sections 1 (Summary), 5 (Apparent weaknesses / surpluses), 6 (Best response), 8 (Reactivity triggers), 9 (Open questions) verbatim
• Refresh only: frontmatter (last_updated, confidence, source_urls), Section 2 (Archetype), Section 3 (Category strength), Section 4 (Behavioral), Section 7 (Matchup history -- if we played them)

Step 8: Write file atomically

• Write to temp file <slug>.md.tmp, fsync, rename
• Validate output against yahoo-mlb/context/frameworks/opponent-profile-schema.md frontmatter + section order before rename

Step 9: Emit signal file

After all 11 updates complete (for all_opponents: true mode), emit signals/wkNN-opponent-profiles.md. Format in resources/template.md.

• Include every archetype posterior that shifted by > 0.05 since last week
• Include every confidence change > 0.10
• List any reactivity-trigger events fired this week

Common Patterns

Pattern 1: inactive (dormant manager -- e.g. Jenny's Team)

• Signals: moves_per_week < 0.5, faab_spent_pct < 0.05, days_since_last_login > 3 or lineup_set_daily = false
• Best-response: Don't include in N-bidder FAAB estimates. Send consolidation trade offers. Expect roster atrophy -- exploit lineup-neglect advantage week-over-week.
• Confidence grows fast: inactivity is unambiguous; expect archetype_confidence above 0.80 by Week 4.

Pattern 2: punt_wins_qs (Marmol strategy -- all-hitter + RP-only staff)

• Signals: sp_roster_share < 0.25, closer_count >= 3, moves_per_week > 3 (cycling for hitter production)
• Best-response: Concede K and QS (two free cat losses); lock 6 of remaining 8 cats. Do not stream SPs against them.

Pattern 3: punt_sv (no closers, elite SP + hitting)

• Signals: closer_count = 0, sp_roster_share > 0.40, high-end SP drafted in round 1-3
• Best-response: Concede SV; push all 5 hitting + K + QS + ERA + WHIP. Do not chase closers as streamers.

Pattern 4: frustrated_active (active but losing -- motivated trader)

• Signals: moves_per_week > 4 AND record_last_2_weeks W% < 0.35
• Best-response: Prime trade partner. They'll respond to offers. Target their cooling stars (sell-high for them, buy-low for us).

Pattern 5: balanced (the default)

• Signals: No strong punts; roster composition within 1 std of league average across all features
• Best-response: Match cat-for-cat; matchup decided on weekly execution. Use variance-seeking when we're the underdog (matchup_win_probability < 0.40).

Guardrails

1. Classifier delegation purity. This skill MUST NOT implement Bayesian math. If you find yourself computing posteriors, likelihoods, or normalizations, stop -- those belong in opponent-archetype-classifier. This skill contributes the taxonomy (MLB-specific), feature extraction (Yahoo-specific), and output formatting. Any math beyond "sum projected season totals for cat_strength" is a smell.

2. Manual-notes preservation is non-negotiable. Users hand-author Sections 1, 5, 6, 8, 9 with strategic notes this skill cannot reproduce (e.g. "Jennifer's email bounced, suggest probe via DM"). Blind-overwriting those sections destroys user work. Always read-modify-write. Validate diff before commit: only frontmatter + Sections 2, 3, 4, 7 should change.

3. Graceful scrape failure. If a Yahoo page returns 4xx/5xx or the session is unauthenticated: do NOT guess. Mark the corresponding features as null, record the failed URL in a scrape_errors: block in frontmatter, and drop confidence by 0.2. The profile should still emit, just with lowered confidence. A partial profile is better than a missing one.

4. Sequential-update discipline. Last week's posterior becomes this week's prior -- always. Do NOT restart from taxonomy priors each week (that throws away 1-4 weeks of evidence). The one exception: if prior_profile.confidence < 0.20, restart (the prior profile is essentially garbage).

5. Correlated-feature handling. moves_per_week and faab_spent_pct are strongly correlated (active managers do both). Pass correlated_feature_pairs to the classifier so it down-weights. Same for sp_roster_share and closer_count (they trade off by roster-slot constraint).

6. Do not invent cat_strength from the archetype. cat_strength is a separate estimator based on actual roster composition, NOT derived from archetype. A balanced team with a thin OF has below-average SB; don't paper over that by inheriting "balanced = 50s across the board". Compute from roster.

7. Archetype unknown is valid. When classification_confidence < 40, write archetype: unknown with archetype_confidence reflecting the classifier output. Do not force a MAP on thin evidence. Downstream agents handle unknown by falling back to broad heuristics.

8. Citation requirement. Every emitted profile lists exact Yahoo URLs in source_urls. The weekly signal file lists all URLs across all 11 teams (dedupe the teams-page URL).

9. One-shot vs all-opponents mode. Single-team refreshes (on trade-offer arrival, FAAB competition observation) emit only the profile file -- no weekly summary signal. Weekly summary signal is emitted only after all 11 refreshes complete.

10. Team slugs stable. Slugs are lowercased + hyphenated team name, stable across the season (e.g., jennys-team, springfield-isotopes). If a manager renames their team mid-season, keep the original slug and store the new team_name in frontmatter -- do not rename the file (breaks downstream agent references).

Quick Reference

Pipeline one-liner:

scrape_yahoo(4_pages) -> extract_mlb_features -> invoke(opponent-archetype-classifier,
  taxonomy=mlb_10_archetype, prior=last_week_posterior, observation_weight=f(week))
  -> compute_cat_strength(roster) -> read_existing_md -> merge_preserve_notes
  -> atomic_write -> [if all_opponents] emit_weekly_signal

MLB 10-archetype taxonomy (baked in):

Archetype	Core signal	Prior (Wk 1)
balanced	no punts, pushes all cats	0.25
stars_and_scrubs	4-5 elite + bench scrubs	0.10
punt_sv	0 closers, loaded elsewhere	0.10
punt_sb	power-only offense	0.10
punt_wins_qs	SP thin, RP-heavy (Marmol)	0.05
hitter_heavy	sp_roster_share < 0.30	0.10
pitcher_heavy	sp_roster_share > 0.45	0.10
inactive	zero moves, stale lineup	0.10
frustrated_active	high moves + losing record	0.05
unknown	inconclusive threshold	0.05

Priors sum to 1.00. Informed (not uniform) per guardrail on 12-team base rates.

Observation weight schedule:

Week	Weight	Rationale
1-2	0.25	Draft-only evidence; noisy
3-4	0.45	First waivers visible
5-7	0.65	Patterns stabilizing
8-11	0.80	Behavior well-characterized
12+	0.90	Any remaining doubt is true ambiguity

Key resources:

• resources/template.md: 10-archetype MLB taxonomy YAML (passed to classifier), per-opponent .md output template (matches opponent-profile-schema.md), weekly summary signal template.
• resources/methodology.md: Yahoo scrape URL flow, feature-extraction formulas, cat_strength estimator, read-modify-write protocol, sequential-update protocol, graceful-degradation handling.
• resources/evaluators/rubric_mlb_opponent_profiler.json: 10 quality criteria (Taxonomy Completeness, Feature Extraction Correctness, Yahoo Scrape Coverage, Classifier Delegation Purity, Output Schema Conformance, Preserve Manual Notes, Sequential-Update Correctness, Signal File Emission, Graceful Scrape Failure, Citations).

Inputs required:

• team_name (string, single team) OR all_opponents: true
• yahoo_session (authenticated Chrome context)
• prior_profile (optional; if missing, bootstrap from taxonomy priors)

Outputs produced:

• Updated context/opponents/<team-slug>.md (one per refreshed team) -- manual notes preserved
• signals/wkNN-opponent-profiles.md (only in all_opponents mode) -- summary of archetype shifts + confidence changes + reactivity events

Downstream consumers (5 agents):

• mlb-lineup-optimizer -- reads §3 cat_strength + §6 best_response for leverage_vs_opponent
• mlb-category-state-analyzer -- reads §2 archetype + §3 to anticipate opponent push/punt
• mlb-faab-sizer -- reads §4 (waiver_aggression, faab_remaining) across all 11 for N-bidder estimate
• mlb-trade-analyzer -- reads §4 (trade_propensity, trade_cooperation_score) + §5 weaknesses
• mlb-fantasy-coach -- reads §6 for the week's opponent for morning brief