@event4u/agent-config 5.7.0 → 5.9.0

package/.agent-src/skills/prediction-pool-optimizer/SKILL.md

@@ -1,7 +1,7 @@
 ---
 model_tier: high
 name: prediction-pool-optimizer
-description: "Optimize prediction-pool tips (kicktipp etc.): pool rules + market odds → the expected-points-maximizing tip per match. Triggers 'optimize my pool tips', 'best kicktipp picks', 'predict'."
+description: "Optimize prediction-pool tips (kicktipp etc.): rules + multi-book consensus odds → expected-points-max answer for every question, scores AND bonus. Triggers 'optimize my pool tips', 'predict'."
 domain: product
 personas: []
 workspaces:
@@ -18,19 +18,22 @@ install:
 
 # prediction-pool-optimizer
 
-> Turn a prediction pool's **scoring rules** plus **market odds** into the
-> tip that maximizes **expected points** — not the most likely outcome.
-> Sport-agnostic core with per-sport probability blocks. Consumed by
-> [`/prediction-pool`](../../commands/prediction-pool.md). The optimization target is
-> the pool's score, so the chain is always **rules → odds → expected value
-> → participant field → tip**, never "who wins this match?".
+> Turn a prediction pool's **scoring rules** plus a **consensus of the major
+> bookmakers' odds** into the answer that maximizes **expected points** — not
+> the most likely outcome — for **every open question in the pool**: match
+> scores AND every bonus / award / special question (top scorer, group
+> winners, champion, most cards …). Sport-agnostic core with per-sport
+> probability blocks. Consumed by [`/prediction-pool`](../../commands/prediction-pool.md).
+> The optimization target is the pool's score, so the chain is always
+> **rules → odds → expected value → participant field → answer**, never
+> "who wins this match?".
 
 ## When to use
 
-Use when someone wants the best tips for a prediction / betting pool
+When someone wants the best tips for a prediction / betting pool
 (kicktipp-style company pools — football WM, basketball WM, …) and the
-target is **pool points**, not match truth. Triggered by the
-[`/prediction-pool`](../../commands/prediction-pool.md) command (Steps 3–5) or directly
+target is **pool points**, not match truth. Triggered by
+[`/prediction-pool`](../../commands/prediction-pool.md) (Steps 3–5) or directly
 when a user asks to optimize / maximize their pool picks.
 
 **The one idea that makes this skill correct:** the highest-probability
@@ -43,39 +46,69 @@ more. **Always optimize the pool's points, never the truth of the match.**
 
 - **Rules before tips.** Never produce a tip before the pool's scoring is
   parsed (Procedure step 1). Strategy is a function of the rules.
-- **Odds are the primary signal.** Bookmaker / market probabilities already
-  fold in form, squad, injuries, travel, climate. Use them as the
-  calibration base; only override with *current* information (confirmed
+- **Answer EVERY open question.** A pool has scores *and* bonus / award /
+  special questions ("which team supplies the top scorer?", "most yellow
+  cards?", "champion?"). Scorelines only, bonus questions blank = a **failed
+  run** — enumerate every open question in step 1, carry each to an answer
+  (steps 5–6). No silent skips.
+- **Odds are the primary signal — multi-book consensus, not one book.**
+  Bookmaker probabilities already fold in form, squad, injuries, travel,
+  climate. Build the base from a **consensus across the 5–10 biggest
+  publicly-viewable books** (step 2), de-vigged, **sharpness-weighted** —
+  never mirror a single portal. Override only with *current* info (confirmed
   lineups, late injuries, suspensions, manager change).
-- **No invented numbers.** Emit no probability you cannot derive from odds
-  or from **actually executed** code. Tournament/outright numbers come from
-  real outright odds **or** the executed Poisson helper — never a claimed
+- **No invented numbers.** Emit no probability you cannot derive from real
+  odds or **actually executed** code. Tournament/outright/award numbers come
+  from real markets **or** the executed Poisson helper — never a claimed
   "I ran 10,000 simulations".
-- **One-sentence justification** per tip. Short.
+- **Scorelines computed, not guessed.** EV-max tip per match from the executed
+  grid optimiser (`score_ev.py`, step 4a), never the eye. A 3:2 / 4:1 / 1:4 in
+  the output = signature of a skipped computation.
+- **One-sentence justification** per answer. Short.
 
 ## Procedure
 
-### 1. Parse the pool rules
+### 1. Parse the pool rules AND enumerate every open question
 
 From the pool's rule page, extract and document:
 
 - Points for **exact result** / **goal (point) difference** / **tendency**.
-- **Bonus questions** (champion, top scorer, group winners …).
-- **Joker / multiplier** rules.
-- **Quote / rarity** scoring (rare correct tips score more)? — flips the
-  whole strategy toward contrarian (step 4).
-- Special scorings, **deadlines**, and **strategy limits** (e.g. max N
-  identical tips).
+- **Every bonus / award / special question** (champion, top scorer, "team of
+  the top scorer", group winners, most cards, longest unbeaten,
+  will-there-be-a-red-card, over/under totals …). **Write them all down as an
+  explicit checklist** — this list is the run's contract; every entry must
+  reach an answer.
+- **Joker / multiplier** rules, per-question point weights.
+- **Quote / rarity** scoring (rare correct tips score more)? — flips strategy
+  toward contrarian (step 4).
+- Special scorings, **per-question deadlines**, **strategy limits** (e.g. max
+  N identical tips).
 - **The goal**: place well, or *win* a large pool? (changes variance — step 4.)
 
-### 2. Build the data base
-
-Primary: current bookmaker odds, aggregated market probabilities, model
-forecasts (e.g. Opta), Elo/SPI ratings. Secondary (only when it adds signal
-the odds have not yet absorbed): confirmed lineups, injuries, suspensions,
-manager change, recent form, home advantage, head-to-head, rest/travel,
-weather. De-vig the odds (remove the bookmaker margin) before treating them
-as probabilities.
+### 2. Build the data base — a consensus across the major books
+
+Primary signal: current bookmaker odds, **aggregated across the 5–10 biggest
+publicly-viewable books**, not a single portal:
+
+1. **Collect** odds for each market (1X2, exact-score, outrights, and each
+   special/award market a bonus question needs) from several books.
+   Odds-comparison aggregators (Oddschecker, Oddsportal / Betexplorer) show
+   many books at once; supplement with named books. Book list + weighting
+   recipe in [`reference/odds-and-bonus.md`](reference/odds-and-bonus.md).
+2. **De-vig each book** independently (remove its margin) → per-book implied
+   probabilities. Raw odds sum to >100%; never treat them as probabilities.
+3. **Aggregate with a healthy weighting**, not a blind average: weight
+   **sharp, low-margin books higher** (Pinnacle, Betfair Exchange),
+   recreational books lower; weighted mean or trimmed median so one outlier
+   book cannot swing the base. Result = the **consensus probability** — the
+   calibration base.
+4. **Single-book outlier = flag, not truth** — investigate *why* (priced-in
+   injury? stale line?) before moving off consensus. Cross-portal agreement is
+   signal; one portal disagreeing is a prompt to check, not to follow.
+
+Secondary (only when it adds signal the consensus has not absorbed): confirmed
+lineups, injuries, suspensions, manager change, recent form, home advantage,
+head-to-head, rest/travel, weather, model forecasts (Opta), Elo/SPI ratings.
 
 ### 3. Per-match probabilities (sport block)
 
@@ -98,7 +131,7 @@ results. Pick the block for the event's sport:
 - Derive the outcome split straight from de-vigged moneyline odds; estimate
   a plausible score from the market total. State the model used.
 
-Cross-check the model against the market; on a large divergence, re-check
+Cross-check the model against the consensus; on a large divergence, re-check
 the data and explain the cause before trusting it.
 
 ### 4. Convert to the EV-maximizing tip
@@ -106,32 +139,69 @@ the data and explain the cause before trusting it.
 Map probabilities to the tip with the **highest expected points under the
 step-1 rules** — not the prettiest match.
 
-- **Standard fixed-point scoring + goal "place well"** → tip the EV-maximal
-  result per match. Favourites with modest scorelines dominate. **No
-  contrarian** — only your tip matters for your score, so deliberately
-  tipping "different" just burns EV.
-- **Quote / rarity scoring** → weigh rarer-but-plausible results against
-  their higher payout; take rarity when `payout × probability` wins.
-- **Goal = win a large pool** → on a *subset* of matches, take calculated
-  variance (plausible underdogs) to create upside, poker-tournament style.
-
-**Participant-field thresholds** (when two tips are close, prefer the one
-with the higher edge over the typical participant):
-
-- Pool **N < 20** → maximize EV, ignore the field.
-- **20 ≤ N < 100 and you are in the prize positions** → maximize EV.
-- **N ≥ 100, or you are outside the top ~20%** → add field-relative
-  variance (move off the consensus on a subset; rough Kelly-fraction sizing).
+#### 4a. The EV-max scoreline is computed, never eyeballed
+
+Don't hand-pick a scoreline. Run the executed grid optimiser — builds the full
+Poisson score grid, returns the EV-max tip under the step-1 point tiers:
+
+```bash
+python3 scripts/prediction-pool/score_ev.py --lh <home-xg> --la <away-xg> \
+    --tendency <t> --diff <d> --exact <e>          # one match
+python3 scripts/prediction-pool/score_ev.py matches.json \
+    --tendency <t> --diff <d> --exact <e>          # batch, prints a ranked table
+```
+
+Two facts the grid makes unavoidable, intuition gets wrong:
+
+- **High scorelines almost never EV-max.** Under partial points a moderate
+  favourite peaks at **1:0 / 2:0 / 2:1**; **1:0 wins surprisingly often**, top
+  of the surface is *flat* (1:0 vs 2:1 vs 2:0 within hundredths). 3:2 / 4:1 /
+  1:4 never optimal — such a tip means the grid wasn't run.
+- **Draws under-tipped.** A correct draw banks the goal-difference tier on
+  every draw scoreline, so in a close match (xG within ~0.4) a 1:1 can
+  out-score a 1:0 — and for low-scoring even games (λ ≲ 1.0/side) a 0:0 is the
+  EV-max. Let the grid decide; the eye tips too few draws.
+
+- **Standard fixed-point scoring + goal "place well"** → tip the grid's EV-max
+  per match. **No contrarian** — only your tip scores, tipping "different"
+  burns EV.
+- **Quote / rarity scoring** → weigh rarer-but-plausible results against payout;
+  take rarity when `payout × probability` wins (raise `--exact` or post-process
+  the ranked table by the multiplier).
+
+#### 4b. Large pool, goal "win it" — measure P(finish 1st), don't guess
+
+Goal = **win** a large pool → target flips from E(points) to **P(finish ahead
+of the field)**; pure EV-max converges with the crowd, can't open a gap.
+Measure it with the executed field simulator, not a "rough Kelly" hand-wave:
+
+```bash
+python3 scripts/prediction-pool/pool_winsim.py pool.json --runs 4000 --max-flips 4
+```
+
+Models the field as softmax-EV tippers, reports `P(win)` for EV-max-everywhere,
+then greedily reports **which few tips to flip** off EV-max (EV cost + P(win)
+gain each). Read it as the field threshold, empirically:
+
+- Pool **N < 20** → sim shows flips barely move P(win); maximize EV, ignore the
+  field.
+- **20 ≤ N < 100 and in the prize positions** → maximize EV.
+- **N ≥ 100, or outside the top ~20%** → take the sim's suggested flips: a
+  handful of higher-variance scorelines on high-consensus matches lift P(win)
+  most per unit EV given up. Flip only what the sim says pays — variance you
+  don't need is wasted EV.
 
 Respect all strategy limits from step 1 (max identical tips, etc.).
 
-### 5. Tournament & bonus questions (no hallucination)
+### 5. Tournament, bonus & special questions — answer every one (no hallucination)
 
-For group winners, KO rounds, champion, and bonus questions, use **either**:
+Walk the **step-1 checklist** and answer **each** entry. Pick the method by
+question type — full taxonomy + per-type method in
+[`reference/odds-and-bonus.md`](reference/odds-and-bonus.md):
 
-- real **outright market odds** ("to win group", "to reach final",
-  "outright winner"), **or**
-- the executed Poisson tournament simulator:
+- **Tournament structure** (group winners, KO rounds, finalists, champion):
+  real **outright market odds** ("to win group", "to reach final", "outright
+  winner") aggregated per step 2, **or** the executed Poisson simulator:
 
   ```bash
   python3 scripts/prediction-pool/poisson_sim.py <teams-xg.json> --runs 20000
@@ -141,48 +211,89 @@ For group winners, KO rounds, champion, and bonus questions, use **either**:
   advancement / title probabilities. **Run it — never report simulated
   numbers you did not actually compute.**
 
-Optimize bonus answers on the same expected-points basis. Re-run as late as
-the deadline allows: re-check confirmed lineups, injuries, suspensions, and
-odds movement, then adjust. The pool's per-match deadline is the only hard
-constraint.
+- **Award / player markets** (top scorer, most assists, "which team supplies
+  the top scorer", golden boot, most cards): use the matching **special
+  market** — e.g. aggregate per-player "top goalscorer" odds **by team** to
+  answer "which team has the top scorer". No clean market → derive from a
+  stated model (squad strength × games-expected) and **label it a model
+  estimate**, not a market number.
+
+- **Binary / over-under specials** (red card yes/no, over/under total
+  goals/cards): de-vig the consensus probability for the line, pick the EV-max
+  side under the question's point weight.
+
+Optimize every answer on the same expected-points basis as the scores. Re-run
+as late as each question's deadline allows: re-check confirmed lineups,
+injuries, suspensions, odds movement, then adjust. The per-question deadline is
+the only hard constraint.
 
 ## Output format
 
 1. **Approval table** — one row per match:
 
    ```
-   Match | Tip | Prob / EV | Risk (low/med/high) | 1-line reason | Odds used
+   Match | Tip | Prob / EV | Risk (low/med/high) | 1-line reason | Books used
+   ```
+
+   `Books used` names the consensus base (e.g. "consensus of 7 books, sharp-weighted").
+
+2. **Bonus & special answers** — one row per open question from the step-1
+   checklist, **every entry answered** (none blank):
+
+   ```
+   Question | Answer | Prob / EV | Risk | 1-line reason | Source (market / model)
    ```
 
-2. **Group standings, the full bracket, and bonus-question answers** where
-   the event has them.
-3. **Self-check note** — confirm the tips reconcile with
-   [`reference/ev-fixtures.md`](reference/ev-fixtures.md) (known pool rules +
-   market odds → a known-good EV tip). If your method disagrees with a
-   fixture, your method is wrong — find the error (usually a forgotten
-   partial-points term or un-de-vigged odds), don't ship the tip.
+3. **Group standings and the full bracket** where the event has them.
+4. **Self-check note** — (a) tips reconcile with
+   [`reference/ev-fixtures.md`](reference/ev-fixtures.md) (known rules + odds →
+   known-good EV tip); (b) bonus table has the **same number of rows as the
+   step-1 checklist** — a shorter table means a question was dropped. If your
+   method disagrees with a fixture, your method is wrong — find the error
+   (usually a forgotten partial-points term, un-de-vigged odds, or following
+   one book instead of the consensus), don't ship the tip.
 
 Handed back to [`/prediction-pool`](../../commands/prediction-pool.md) for the approval
 gate — the skill never enters or submits anything.
 
 ## Gotcha
 
-- **Tipping the modal result, not the EV-maximal one.** The single most
-  likely scoreline rarely maximizes partial points — compute EV across the
+- **Answering only the scores.** Bonus / award questions carry real points;
+  leaving them blank because they are "not a scoreline" forfeits them. The
+  step-1 checklist exists so every question is answered.
+- **Following one portal.** A single book can be stale or shaded; build the
+  base from a sharp-weighted consensus across several; an outlier is a flag to
+  investigate, not a number to copy.
+- **Tipping the modal result, not the EV-maximal one.** The single most likely
+  scoreline rarely maximizes partial points — run `score_ev.py` across the
   result grid, don't eyeball the favourite.
-- **Forgetting to de-vig.** Raw bookmaker odds sum to >100%; treating them
-  as probabilities inflates the favourite. Remove the margin first.
-- **Contrarian under fixed points.** Deviating "to stand out" only helps
-  under quote/rarity rules or a win-a-large-pool goal — otherwise it burns EV.
-- **Claimed-but-unrun simulation.** Numbers like "I ran 10,000 tournaments"
-  without executing `poisson_sim.py` are hallucinated — run the code or use
-  outright odds.
+- **Hand-picking a high scoreline.** 3:2 / 4:1 / 1:4 never EV-max under partial
+  points — moderate favourites peak at 1:0 / 2:0 / 2:1. A high tip = grid
+  skipped; run `score_ev.py`.
+- **Under-tipping draws.** A correct draw banks the goal-difference tier on
+  every draw scoreline, so a close match can want 1:1 (or 0:0). Let the grid
+  decide; the eye tips too few draws.
+- **"Rough Kelly" variance for a large pool.** Don't guess deviation amount —
+  run `pool_winsim.py`; returns the exact flips that raise P(finish 1st) most
+  per unit EV given up.
+- **Forgetting to de-vig.** Raw bookmaker odds sum to >100%; treating them as
+  probabilities inflates the favourite. Remove the margin **per book** before
+  aggregating.
+- **Contrarian under fixed points.** Deviating "to stand out" only helps under
+  quote/rarity rules or a win-a-large-pool goal — otherwise it burns EV.
+- **Claimed-but-unrun simulation.** "I ran 10,000 tournaments" without
+  executing `poisson_sim.py` is hallucinated — run the code or use outright odds.
 
 ## Do NOT
 
+- Leave any open pool question (bonus / award / special) unanswered.
+- Build the base from a single bookmaker, or skip de-vigging before aggregating.
 - Tip the most likely result instead of the EV-maximal one.
+- Hand-pick a scoreline instead of running `score_ev.py` — never emit a
+  3:2 / 4:1 / 1:4 tip, never EV-max under partial points.
 - Go contrarian under standard fixed-point scoring with a "place well" goal.
-- Report Monte-Carlo numbers without running `poisson_sim.py`.
+- Guess large-pool variance ("rough Kelly") instead of running `pool_winsim.py`.
+- Report Monte-Carlo numbers without running `poisson_sim.py` / `pool_winsim.py`.
 - Treat raw odds as probabilities without removing the vig.
 - Give betting or financial advice — this optimizes a game; the human submits.
 
@@ -190,7 +301,14 @@ gate — the skill never enters or submits anything.
 
 - [`/prediction-pool`](../../commands/prediction-pool.md) — the orchestrator (event,
   persistence, Playwright entry, gates).
+- [`reference/odds-and-bonus.md`](reference/odds-and-bonus.md) — major-book list
+  + sharpness-weighted consensus recipe, and the bonus / award / special
+  question taxonomy with a per-type method.
 - [`reference/ev-fixtures.md`](reference/ev-fixtures.md) — known-good
   rules+odds → EV examples.
+- [`scripts/prediction-pool/score_ev.py`](../../../../scripts/prediction-pool/score_ev.py) —
+  executed exact-score EV optimiser (step 4a; λ + rule → EV-max scoreline).
+- [`scripts/prediction-pool/pool_winsim.py`](../../../../scripts/prediction-pool/pool_winsim.py) —
+  executed field model + P(finish 1st) simulator and flip-finder (step 4b).
 - [`scripts/prediction-pool/poisson_sim.py`](../../../../scripts/prediction-pool/poisson_sim.py) —
-  the executed tournament simulator.
+  the executed tournament simulator (step 5).

package/.agent-src/skills/prediction-pool-optimizer/evals/triggers.json

@@ -1,6 +1,6 @@
 {
   "skill": "prediction-pool-optimizer",
-  "description": "7 should-trigger + 5 should-not-trigger queries. Should-trigger covers DE + EN phrasings and the core intent (pool tips, kicktipp, expected-points optimization across sports). Should-not-trigger covers near-miss neighbours: regulated financial advice (finance pack), plain match-result prediction with no pool, generic web research, AI video, and real-money sportsbook betting (out of scope / refuse).",
+  "description": "9 should-trigger + 5 should-not-trigger queries. Should-trigger covers DE + EN phrasings and the core intent (pool tips, kicktipp, expected-points optimization across sports, plus answering the bonus / award questions and using bookmaker-consensus odds). Should-not-trigger covers near-miss neighbours: regulated financial advice (finance pack), plain match-result prediction with no pool, generic web research, AI video, and real-money sportsbook betting (out of scope / refuse).",
   "queries": [
     {"q": "optimize my kicktipp tips for the football WM 2026", "trigger": true},
     {"q": "fill my company Tippspiel for the basketball world cup", "trigger": true},
@@ -9,6 +9,8 @@
     {"q": "maximiere meine erwarteten Punkte im Tippspiel, nicht nur wer gewinnt", "trigger": true},
     {"q": "predict our office kicktipp pool for the WM", "trigger": true},
     {"q": "mach mein Tippspiel für die WM", "trigger": true},
+    {"q": "beantworte auch alle Bonusfragen im kicktipp, z.B. welche Mannschaft den Torschützenkönig stellt", "trigger": true},
+    {"q": "use the odds from the big betting sites to optimize my pool picks", "trigger": true},
     {"q": "should we invest in this startup based on a DCF?", "trigger": false, "note": "regulated financial valuation → dcf-modeling / finance pack"},
     {"q": "who will win tonight's match?", "trigger": false, "note": "plain result prediction, no pool / no scoring rules to optimize"},
     {"q": "research the best running shoes for me", "trigger": false, "note": "generic web research → research / deep-research"},

package/.agent-src/skills/prediction-pool-optimizer/reference/ev-fixtures.md

@@ -15,16 +15,25 @@ shape you encounter so future runs catch the same class of drift.
 **Rule:** exact result = 4, goal-difference = 3, tendency = 2, else 0.
 No quote rule. No strategy limit. Goal: place well.
 
-**Match (football):** de-vigged market — Home 62% / Draw 24% / Away 14%.
-Most plausible exact results (Poisson on market xG ≈ 1.7 : 0.8):
-2:1 ≈ 9%, 1:0 ≈ 9%, 2:0 ≈ 8%, 1:1 ≈ 8%, 3:1 ≈ 6%.
+**Match (football):** Poisson on market xG ≈ 1.7 : 0.8.
 
-**Reasoning:** the single most likely *result* (2:1) and 1:0 both bank the
-tendency (2) on a home win plus goal-difference (3) on many neighbours.
-Expected points of "2:1" beats "tip the favourite to win 3:0" (lower hit
-rate on diff/exact) and beats any draw/away tip (tendency rarely banks).
+**Script-verified** (`score_ev.py --lh 1.7 --la 0.8 --exact 4 --diff 3 --tendency 2`):
 
-**Known-good tip:** **2:1 home.** (Risk: low.) **Not** contrarian — under
+```
+EV-max tip : 1:0  (EV 1.574)
+  1:0  1.574  <- EV-max
+  2:1  1.530
+  2:0  1.477
+```
+
+**Reasoning:** top of the EV surface is **flat** — 1:0, 2:1, 2:0 all bank the
+tendency (2) plus goal-difference (3) on many neighbours, within hundredths of
+each other. Grid puts **1:0 narrowly first**; eyeballing the modal *result*
+(2:1) lands a near-tie, not the optimum. Run the grid — don't assert the
+favourite's "obvious" score.
+
+**Known-good tip:** **1:0 home** (2:1 essentially tied; with the real de-vigged
+λ either can lead — the grid decides). (Risk: low.) **Not** contrarian — under
 fixed points only your own tip scores, so deviating costs EV.
 
 ---
@@ -54,15 +63,17 @@ the multiplier; take the max.
 **Match (football):** a near-coin-flip favourite, Home 52% / Draw 26% /
 Away 22%.
 
-**Reasoning:** with N ≥ 100 and you behind, pure EV converges with the
-field and cannot create the gap you need. Add field-relative variance on a
-*subset*: take a plausible underdog/draw where the consensus is heavy on
-the favourite, sized by a rough Kelly fraction. On safe matches, still
-tip EV-max.
+**Reasoning:** N ≥ 100 and you behind → pure EV converges with the field, can't
+create the gap; target is **P(finish 1st)**, not E(points). Don't guess the
+variance: run `pool_winsim.py` with the pool's `N` and your `my_lead`. Shows
+P(win) collapsing under EV-max-everywhere, returns the **specific flips**
+(higher-variance scorelines on high-consensus matches) that raise P(win) most
+per unit EV given up.
 
-**Known-good tip:** EV-max on the safe matches; **calculated underdog**
-(e.g. 1:1 or away) on 2–4 high-consensus matches to manufacture upside.
-(Risk: high — intentional.)
+**Known-good tip:** EV-max on the safe matches; the **simulator's suggested
+flips** on the 2–4 matches it names, to manufacture upside. (Risk: high —
+intentional.) Verify the sim shows a P(win) gain — flips not moving it (small
+N) → don't add variance you don't need.
 
 ---
 
@@ -78,3 +89,87 @@ Margin modelled Gaussian, mean ≈ 6.5, sd ≈ 11.
 spread). Tip the winner plus the modal margin bucket.
 
 **Known-good tip:** **Home win, margin ~5–9.** (Risk: low on winner.)
+
+---
+
+## Fixture 5 — multi-book consensus (de-vig per book, sharp-weighted)
+
+**Rule:** any — checks the **odds base**, not the EV map.
+
+**Market (football, 1X2):** two books.
+- Book S (sharp, weight 3): 1.80 / 3.60 / 4.50 → de-vig 0.526 / 0.263 / 0.210.
+- Book R (recreational, weight 1): 1.75 / 3.50 / 4.20 → de-vig 0.522 / 0.261 / 0.217.
+
+**Reasoning:** de-vig **each book** first (raw `1/o` sums to >1; normalise),
+then sharp-weighted mean per outcome and renormalise. Aggregating raw odds, or
+using one book, is wrong.
+
+**Known-good base:** **Home 0.525 / Draw 0.262 / Away 0.212.** A run that fed
+the EV grid one book's raw odds has the wrong base — fix it before the tip.
+
+---
+
+## Fixture 6 — "team of the top scorer" (aggregate player market by team)
+
+**Rule:** bonus question = 6 points: "which team supplies the tournament top
+scorer?"
+
+**Market (top-goalscorer outright, de-vigged player probabilities):**
+- Team A: A1 14%, A2 5% → team A total **19%**.
+- Team B: B1 16% → team B total **16%**.
+- Team C: C1 9%, C2 4% → team C total **13%**.
+
+**Reasoning:** the most-likely *player* (B1, 16%) is on team B, but the
+question asks the **team** — sum each squad's players. Team A 19% beats team B
+16%. Answer the asked question, not the adjacent one.
+
+**Known-good answer:** **Team A.** (Source: market, aggregated by team. Risk:
+medium.) **Not** team B — the modal-player trap.
+
+---
+
+## Fixture 7 — high-scoreline trap (the "EV-optimized" model that wasn't)
+
+**Rule:** kicktipp 2 / 3 / 5 — tendency = 2, goal-difference = 3, exact = 5.
+
+**Matches (script-verified, `score_ev.py … --tendency 2 --diff 3 --exact 5`):**
+
+| Match (λ) | EV-max | a high tip's EV | verdict |
+|---|---|---|---|
+| Senegal–Iraq (2.0:0.7) | **1:0** (1.881) | 4:1 ≈ 1.55 | high tip leaks ~0.33 |
+| Qatar–Switzerland (0.6:2.1) | **0:1** (1.981) | 1:4 ≈ 1.65 | tipping the underdog's goals = costliest move on the board |
+| Spain–CapeVerde (2.3:0.6) | **2:0** (2.033) | 3:1 ≈ 1.88 | only at λ ≳ 2.3 does 2:0 edge past 1:0; never higher |
+
+**Reasoning:** under partial points the value sits in the tendency and
+goal-difference tiers, not the exact high score. **1:0 is the optimum
+astonishingly often** (even for clear favourites at λ ≈ 2.0); 2:0 takes over
+only near λ ≈ 2.3–2.4; above that, never. **3:2 / 4:1 / 4:2 / 1:4 are never
+EV-max.** Adding goals — especially the underdog's — only shrinks the hit
+probability without protecting the diff/tendency points.
+
+**Known-good behaviour:** any 3:2 / 4:x / x:4 tip in the run → the grid wasn't
+run; `score_ev.py` is the gate. (Risk: low; correctness fixture, not strategy.)
+
+---
+
+## Fixture 8 — draws are under-tipped
+
+**Rule:** kicktipp 2 / 3 / 5 (as Fixture 7).
+
+**Matches (script-verified, `score_ev.py … --tendency 2 --diff 3 --exact 5`):**
+
+```
+λ 1.0:1.0  ->  EV-max 0:0 (1.196), 1:1 tied (1.196)   # a draw IS the optimum
+λ 0.9:0.9  ->  EV-max 0:0 (1.317), 1:1 second
+λ 1.2:1.2  ->  EV-max 1:0 (1.150), draw third (1.091)  # 1-goal win edges it
+```
+
+**Reasoning:** people tip too few draws. A correct draw banks the
+goal-difference tier (3) on *every* draw scoreline, so in a **low-scoring even
+match (λ ≲ 1.0/side) the draw — usually 0:0 — is the EV-max**, tied with 1:1.
+As λ rises past ~1.1 a one-goal win edges ahead, but the draw stays in the top
+tips. Grid surfaces this; intuition suppresses it.
+
+**Known-good behaviour:** a tip set with **near-zero draws across many
+low-scoring even matches** is a red flag — re-run `score_ev.py`, let the grid
+decide, don't default every close game to 1:0.

package/.agent-src/skills/prediction-pool-optimizer/reference/odds-and-bonus.md

@@ -0,0 +1,109 @@
+# Odds aggregation + bonus-question taxonomy
+
+Lookup material for `prediction-pool-optimizer`. Two parts:
+
+- **A — Multi-book consensus**: which books to read, how to weight them, how
+  to fold them into one calibration probability.
+- **B — Bonus / award / special questions**: a type → method table so every
+  open question in the pool reaches an answer.
+
+Not betting advice; how to read a public market as a probability prior for a
+fun pool.
+
+---
+
+## A. Multi-book consensus — read several, weight by sharpness
+
+### Why not one book
+
+A single bookmaker's line can be stale, regionally shaded, or carry a fat
+margin. A consensus across several books is a far better probability estimate,
+and cross-book agreement tells you the market's confidence. **Never mirror one
+portal.**
+
+### Which books (5–10, publicly viewable)
+
+Availability varies by region and over time — this list is **illustrative,
+refresh it at run time** and use whatever is publicly viewable from the current
+locale. Fastest way to see many at once = an **odds-comparison aggregator**:
+
+- **Aggregators (many books on one page):** Oddschecker, Oddsportal,
+  Betexplorer, OddsAlert.
+- **Sharp / low-margin reference books (weight higher):** Pinnacle, Betfair
+  Exchange (an exchange = closest thing to a true market price).
+- **Large recreational books (weight lower):** bet365, Bwin, William Hill,
+  Unibet, Betano, Tipico, Interwetten, bet-at-home, 888sport, Winamax.
+
+Aim for **5–10** spanning both groups. If only recreational books are viewable,
+say so in the run note — the consensus is then softer.
+
+### The recipe
+
+1. **Per market, per book**: collect decimal odds (1X2, exact-score, each
+   outright, each special/award market a bonus question needs).
+2. **De-vig each book independently.** For 1X2 decimal odds `o_H, o_D, o_A`,
+   raw implied probs `1/o` sum to `>1` (overround); normalise:
+   `p_i = (1/o_i) / Σ(1/o)`. Per book — never aggregate raw odds.
+3. **Sharpness-weight and combine.** Sharp books > recreational; **weighted
+   mean** — or **trimmed median** when books disagree a lot (robust to one
+   outlier). A defensible weighting:
+
+   - Pinnacle / Betfair Exchange → weight 3
+   - large recreational books → weight 1
+   - aggregator "average" column → weight 1 (already blends many)
+
+   `p_consensus = Σ(wᵢ · pᵢ) / Σwᵢ` per outcome, then re-normalise the outcome
+   set to sum to 1.
+4. **Outlier handling.** One book far off the others = **flag, not truth**:
+   check for a reason (priced-in injury, stale line) before moving the
+   consensus. Cross-book agreement = signal; one disagreeing book = investigate.
+5. **Healthy weighting overall.** The consensus is a **prior**. Blend it with
+   the per-sport model (Poisson / Gaussian) and override only with *current*
+   info the market has not absorbed (confirmed lineup, late injury,
+   suspension, manager change). The pool answer is EV under the rules on top of
+   this blended probability — the market informs, it does not dictate.
+
+### Worked mini-example (1X2)
+
+Two books, home/draw/away decimal odds:
+
+- Book S (sharp, w=3): 1.80 / 3.60 / 4.50 → raw 0.556/0.278/0.222 (sum 1.056)
+  → de-vig 0.526/0.263/0.210
+- Book R (recreational, w=1): 1.75 / 3.50 / 4.20 → raw 0.571/0.286/0.238
+  (sum 1.095) → de-vig 0.522/0.261/0.217
+
+Weighted mean (3:1), per outcome, then renormalise:
+≈ **Home 0.525 / Draw 0.262 / Away 0.212**. That is the calibration base for
+the per-match EV grid — not either book's raw number.
+
+---
+
+## B. Bonus / award / special questions — type → method
+
+Every entry on the step-1 checklist gets an answer. Match the question to a
+row; real market where one exists, a **labelled** model estimate where none
+does. Optimize each on expected points under its point weight.
+
+| Question type | Example | Method |
+|---|---|---|
+| **Outright winner** | "Who wins the tournament?" | Outright "to win" market, consensus per A; or `poisson_sim.py` `title_pct`. EV-max under the question's points. |
+| **Group / stage** | "Who wins group X?", "Who advances?" | "To win group" / "to qualify" markets; or `advance_pct` from the simulator. |
+| **Finalists / matchup** | "Who reaches the final?" | "To reach final" market per team; simulator pairing is approximate — prefer the market. |
+| **Top scorer (player)** | "Tournament top scorer?" | "Top goalscorer" outright market, consensus per A; EV-max player (favourite unless rarity scoring rewards a longer shot). |
+| **Team of the top scorer** | "Which team supplies the top scorer?" | Aggregate per-player top-scorer probabilities **by team** (sum each squad); pick the highest-summed team. |
+| **Most assists / cards / etc.** | "Most yellow cards?" | Matching special market if offered; else a labelled model estimate (discipline/aggression proxy). |
+| **Binary special** | "Will there be a red card in match X?" | De-vig the yes/no line to a probability; EV-max side under the points. No market → labelled base-rate estimate. |
+| **Over / under total** | "Over/under total goals / cards?" | De-vig the totals line at the offered threshold; higher-EV side. |
+| **Exact stat** | "How many goals in the final?" | Market totals distribution if available; else per-match Poisson on consensus xG. State the model. |
+
+### Rules for bonus answers
+
+- **Answer all of them.** The output's bonus table must have the same number
+  of rows as the step-1 checklist. A missing row = a dropped question.
+- **Market first, labelled model second.** Prefer a real special market; none
+  exists → derive from a stated model and mark `Source: model`.
+- **Rarity rules apply here too.** Under quote/rarity scoring, a
+  plausible-but-rarer answer can out-score the favourite when
+  `payout × probability` is higher — same EV logic as the scores.
+- **No hallucinated numbers.** Outright/award probabilities come from real
+  markets or the executed simulator — never a claimed-but-unrun simulation.