Head-to-head prompt diff

aa608c898bb5 → 043fa3fcfd52

	aa608c898bb5			043fa3fcfd52
f	1	# Head-to-head football report	f	1	# Head-to-head football report
	2			2
n	3	Prepare a fixture-specific prediction by grounding the analysis in the two current team scorecards,	n	3	Produce a fixture-specific 1X2 prediction in three strictly separated layers:
	>	then adding match-context that does not belong in those scorecards.
	4			4
n	5	Scorecards are team-level snapshots. They should be updated first, but they should remain about each	n	5	1. Recorded facts — deterministic data retrieved from the local stats ledger via `football-model
	>	team's baseline state rather than this specific opponent. The head-to-head report is where opponent		>	stats h2h-card`, included verbatim. Code computes these; you do not.
	>	interaction, venue, market, and matchup-specific reasoning live.
				6	2. Model baseline — your own W/D/L probabilities built from the team scorecards plus odds-free f
				>	ixture context, frozen before you see any market price.
				7	3. Market comparison — odds, implied probabilities, and value assessment, strictly downstream of
				>	the frozen baseline.
				8
				9	The order exists because market prices anchor whatever they touch. The baseline must be your indepen
				>	dent estimate of the match; the market is what you compare that estimate against afterwards. A predi
				>	ction whose "model" probabilities were back-solved from odds measures nothing and can never reveal a
				>	n edge.
	6			10
	7	## Required order		11	## Required order
	8			12
	9	1. Read `instruction-updating` and this `instruction-head-to-head` document.		13	1. Read `instruction-updating` and this `instruction-head-to-head` document.
n	10	2. Read both team scorecards.	n	14	2. Read both team scorecards. They must already be updated for this run under `instruction-updating`
				>	; refresh them first if current evidence has moved a team's baseline state.
	11	3. Refresh each team scorecard under `instruction-updating` if current evidence has changed the team		15	3. Generate the recorded-stats card:
	>	's baseline state.
				16
				17	```
				18	football-model stats h2h-card --team1 "<team1>" --team2 "<team2>" --event-date <YYYY-MM-DD>
				19	```
				20
				21	This card is the only permitted source for recorded H2H meetings, recorded form counts, rest
				>	days, and the quant reference. Where it says "none recorded", that data does not exist for this run
				>	— do not substitute remembered results or invent any. Record any results you have separately verifie
				>	d with `football-model stats record-result` first, then regenerate the card.
				22	4. Research fixture-specific context without odds: venue and neutrality, expected lineups for th
				>	is match, availability changes since the scorecards were written, rest/travel/congestion, motivation
				>	and stakes, tactical interaction, weather/referee if material. Do not look up odds, prices, lin
				>	es, or market sentiment in this step. If a source happens to show prices, ignore them and do not wri
				>	te them down.
				23	5. Build the baseline with the method below and freeze it by writing this line in your working notes
				>	and report:
				24
				25	```
				26	BASELINE (pre-odds): <team1> 0.xx / draw 0.xx / <team2> 0.xx — confidence A\|B\|C
				27	```
				28
				29	After this line is written, the numbers may change only through the disagreement protocol, never
				>	by drifting toward the market.
				30	6. Only now fetch market odds for the three outcomes:
				31
				32	```
				33	football-model odds --market "<team1> vs <team2>" --selection "<team1>" --sport <sport-key> --reg
				>	ion <region>
				34	football-model odds --market "<team1> vs <team2>" --selection "Draw" --sport <sport-key> --reg
				>	ion <region>
				35	football-model odds --market "<team1> vs <team2>" --selection "<team2>" --sport <sport-key> --reg
				>	ion <region>
				36	```
				37
				38	Convert to implied probabilities with the overround removed: `implied_i = (1/odds_i) / ((1/odds_1
				>	) + (1/odds_X) + (1/odds_2))`. Example: odds 2.50 / 3.40 / 3.00 give raw 0.400 / 0.294 / 0.333, tota
				>	l 1.027, so implied 0.389 / 0.286 / 0.324.
				39	7. Run the disagreement protocol (below). The probabilities that come out of it are the prediction o
				>	f record.
	12	4. Re-read/capture the two current team blob hashes plus the current `head-to-head` and `updating` i		40	8. Re-read and capture the two current team blob hashes plus the current `head-to-head` and `updatin
	>	nstruction hashes.		>	g` instruction hashes, immediately before persistence.
	13	5. Research fixture-specific context: venue, expected lineups for this match, tactical interaction,		41	9. Create the ledger entry with `football-model predictions create ... --scorecards-updated`, passin
	>	market odds, weather/referee if material, motivation, and relevant recent H2H.		>	g the prediction-of-record probabilities. The command fetches its own market odds, stores normalized
				>	market probabilities, applies the edge/Kelly gates, and prints the recommendation — you never type
				>	odds into the ledger. It also captures the current recorded-stats card onto the row automatically.
	14	6. Produce W/D/L probabilities that sum to 1.0, explicitly grounded in the scorecards plus the added		42	10. Write the full report (structure below), save it to a file, and attach it: `football-model predi
	>	match context.		>	ctions set-report <ledger-id> --from-file <path>`. Attach it immediately: the report can only be att
				>	ached or replaced while the prediction is unsettled; once the result is recorded it is locked as the
				>	pre-result record, and corrections belong in post-mortem notes.
	15	7. Create the prediction ledger entry with `--scorecards-updated`.		43	11. Return the report and ledger id. If any required step is blocked (missing scorecard, no odds ava
				>	ilable), report the exact blocker instead of improvising.
	16	8. Return a report with assumptions, risk flags, market comparison, and the ledger id.		44
				45	## Baseline method
				46
				47	Follow this recipe exactly; it is designed to be reproducible by a careful generalist, not to requir
				>	e modelling judgement.
				48
				49	### Step B1 — strength-gap anchor
				50
				51	Judge the pure quality gap between the two squads at current full strength, ignoring venue, availabi
				>	lity, and this week's news. Pick the closest row (venue-neutral numbers; favourite may be either tea
				>	m):
				52
				53	\| gap \| favourite \| draw \| underdog \|
				54	\|--------------------------------------------------------\|-----------\|------\|----------\|
				55	\| even — same tier, no clear edge \| 0.37 \| 0.26 \| 0.37 \|
				56	\| slight — same tier, one side somewhat stronger \| 0.45 \| 0.26 \| 0.29 \|
				57	\| clear — about one tier apart \| 0.55 \| 0.23 \| 0.22 \|
				58	\| strong — about two tiers apart \| 0.65 \| 0.19 \| 0.16 \|
				59	\| very strong — elite vs weak \| 0.72 \| 0.16 \| 0.12 \|
				60	\| overwhelming — elite vs minnow \| 0.80 \| 0.12 \| 0.08 \|
				61
				62	Interpolate between adjacent rows when the gap sits between them. State in one sentence which row yo
				>	u chose and the main scorecard evidence for it.
				63
				64	### Step B2 — venue shift
				65
				66	- True home fixture: move +0.04 to +0.07 from the away side to the home side. Use the top of the ran
				>	ge for strong effects (altitude, extreme travel for the visitor, notoriously strong home crowd); def
				>	ault +0.05.
				67	- Tournament host, or heavy crowd asymmetry at a nominally neutral venue: at most +0.03.
				68	- True neutral: no shift.
				69
				70	### Step B3 — bounded adjustments
				71
				72	Apply only adjustments backed by a stated, sourced fact. List every adjustment with direction, size,
				>	and the fact behind it. Caps per category, as probability moved between the two teams (or to/from d
				>	raw where that is clearly the effect):
				73
				74	- availability (missing or returning key players): up to ±0.05
				75	- rest/travel/congestion edge: up to ±0.03
				76	- motivation/stakes asymmetry (must-win vs dead rubber): up to ±0.03
				77	- tactical matchup between the two scorecard identities: up to ±0.04
				78	- recorded H2H pattern: up to ±0.02, and only if the recorded-stats card shows at least 3 meetin
				>	gs inside its window; otherwise exactly 0. H2H history is weak evidence; squads, managers, and conte
				>	xts change. Never exceed this cap no matter how striking the pattern looks.
				79	- recent form as such: 0 — form is already inside the scorecards and the anchor choice; adjusting fo
				>	r it again double-counts.
				80
				81	Total movement across all adjustments: at most ±0.12 on any single outcome. If one underlying cause
				>	shows up in several guises (an injury that also explains a form dip and a tactical hole), count it o
				>	nce.
				82
				83	### Step B4 — draw sanity and renormalization
				84
				85	- The draw must land in 0.15–0.33, unless the favourite is at or above 0.70, in which case it may fa
				>	ll to 0.10–0.15 (big mismatches kill draws).
				86	- A draw outside 0.20–0.30 needs a one-sentence justification.
				87	- Renormalize so the three probabilities sum to exactly 1.00 at two decimal places; absorb rounding
				>	in the largest probability.
				88
				89	### Step B5 — quant cross-check
				90
				91	If the recorded-stats card includes a quant reference, compare it with your numbers. Any outcome dif
				>	fering from the reference by more than 0.15 requires one sentence explaining why (the reference is v
				>	enue-agnostic, blind to opponent quality, and underestimates draws — differences are often right, bu
				>	t you must be able to say why yours is). If the card says "not computed", note that and move on. The
				>	reference is never the baseline; never copy it.
				92
				93	### Step B6 — freeze
				94
				95	Write the `BASELINE (pre-odds)` line with the confidence grade. The ledger entry must use exactly th
				>	e prediction-of-record numbers that survive the disagreement protocol.
				96
				97	## Confidence grades
				98
				99	- A — both scorecards updated this run with availability essentially known, and the recorded-sta
				>	ts card has form data for both teams.
				100	- B — one notable gap: unconfirmed lineups, thin or missing recorded data, or one scorecard area
				>	known stale.
				101	- C — major gaps: a scorecard mostly empty or stale, availability unknown, or key facts contradi
				>	ctory.
				102
				103	With grade C: cap the most likely outcome at 0.75, widen the draw toward the middle of its range, an
				>	d recommend no-bet in the report narrative unless the market disagreement is extreme (0.15+) and sur
				>	vives the re-check.
				104
				105	## Disagreement protocol
				106
				107	Compare the frozen baseline with the overround-removed market probabilities, per outcome:
				108
				109	- All outcomes within 0.10: no action. Report the comparison.
				110	- Any outcome 0.10 or more apart: do exactly one targeted re-check of team news for a fact you m
				>	ay have missed (late team news, key injury, rotation announcement, motivation report). Then:
				111	- A specific, nameable missed fact found: name it in the report, change only the adjustment it
				>	affects within the normal caps, renormalize, and record both the original frozen baseline and the r
				>	evised prediction of record.
				112	- No nameable fact found: keep the baseline unchanged as the prediction of record. The gap is
				>	your measured disagreement with the market — that is exactly what the ledger exists to test, and the
				>	edge/Kelly gates in `predictions create` decide whether it is bet-worthy.
				113	- Never split the difference with the market, and never revise more than once.
	17			114
	18	## Boundary between scorecards and H2H context		115	## Boundary between scorecards and H2H context
	19			116
	20	Put in team scorecards:		117	Put in team scorecards:
	33	- how Team A's traits interact with Team B's traits		130	- how Team A's traits interact with Team B's traits
	34	- fixture venue and game-state incentives		131	- fixture venue and game-state incentives
	35	- predicted lineups specifically for this match		132	- predicted lineups specifically for this match
	36	- opponent-specific tactical edges or vulnerabilities		133	- opponent-specific tactical edges or vulnerabilities
n	37	- whether H2H history is relevant to this matchup	n	134	- recorded H2H relevance for this matchup
	38	- market odds and model-vs-market comparison		135	- market odds and model-vs-market comparison
	39	- probability synthesis, recommendation, and risk flags for this fixture		136	- probability synthesis, recommendation, and risk flags for this fixture
	40			137
n	41	If a piece of research is only meaningful because of the opponent, keep it in the H2H report. If it	n	138	If a piece of research is only meaningful because of the opponent, keep it in the H2H report. If it
	>	changes the reusable understanding of the team, update the scorecard.		>	changes the reusable understanding of the team, update the scorecard — but never with market prices.
	42
	43	## Prediction method
	44
	45	Use the team scorecards as the base layer. Treat head-to-head history as weak contextual input, not
	>	a stand-alone model.
	46
	47	A practical weighting is:
	48
	49	- 50-60% team baseline strength and underlying metrics from scorecards
	50	- 25-35% recent form and availability reflected in scorecards
	51	- 10-20% fixture-specific context: venue, lineup interaction, rest/travel edge, tactical matchup, mo
	>	tivation, weather, market movement
	52
	53	Avoid double-counting. If a market move, lineup downgrade, and xG trend all reflect the same injury,
	>	explain the single underlying cause rather than stacking three independent penalties.
	54
	55	Compare final model probabilities with market-implied probabilities after removing overround. If the
	>	model and market strongly disagree, re-check team news, lineups, motivation, venue, and data freshn
	>	ess before persisting.
	56			139
	57	## Report structure		140	## Report structure
	58			141
n	59	Return:	n	142	Return, in order:
	60			143
n	61	1. fixture: competition, date/time, venue	n	144	1. fixture: competition, date/time, venue, neutrality
	62	2. provenance: team doc hashes and instruction hashes		145	2. provenance: team doc hashes, instruction hashes, stats-card `generated_at`
	63	3. scorecard-grounded baseline: concise snapshot of each team		146	3. recorded-stats card: the `stats h2h-card` output, verbatim — including any "none recorded" lines
	64	4. fixture-specific context: availability, lineups, venue, rest, motivation, weather/referee		147	4. scorecard-grounded team summaries: a few sentences per team, from the scorecards
	65	5. tactical matchup: how the two scorecards interact in this fixture		148	5. fixture-specific context: the odds-free research findings, each with source and as-of date
	66	6. market check: odds, implied probabilities, line movement		149	6. baseline (pre-odds): anchor row chosen, venue shift, every bounded adjustment with size and fact,
				>	draw check, quant cross-check sentence, then the frozen `BASELINE (pre-odds)` line with confidence
				>	grade
				150	7. market comparison: odds seen, overround-removed implied probabilities, per-outcome gaps, disagree
				>	ment-protocol outcome including any named missed fact and revised numbers
	67	7. prediction: Team1 / Draw / Team2 probabilities, fair odds, confidence		151	8. prediction of record: final Team1 / Draw / Team2 probabilities and fair odds
	68	8. recommendation: bet/no-bet and why		152	9. recommendation: the bet/no-bet result and reasoning from `predictions create`
	69	9. risk flags and data gaps		153	10. risk flags and data gaps
	70	10. ledger id and recommendation text after persistence		154	11. ledger id
	71			155
	72	## Guardrails		156	## Guardrails
	73			157
t	74	- Do not predict from stale or missing scorecards.	t	158	- No odds, prices, or market-implied probabilities may be looked up, quoted, or written down before
				>	the `BASELINE (pre-odds)` line exists. This includes "checking quickly" during research.
	75	- Do not write opponent-specific analysis into team scorecards.		159	- Never write market prices into team scorecards, in any step, ever.
	76	- Do not overweight old H2H meetings.		160	- Every "recorded" claim must appear verbatim in the stats card. If the card says none recorded, you
				>	have no recorded data — say so and grade confidence accordingly.
	77	- Do not invent odds, injuries, lineups, or sources.		161	- H2H history influence is capped at ±0.02 and requires at least 3 recorded meetings in window.
	78	- If live odds are unavailable, report the blocker and do not fake a ledger entry.		162	- Never average the baseline toward the market without a named missed fact; never revise twice.
	79	- If critical lineup/news is uncertain, cap confidence and make the uncertainty explicit.		163	- Do not invent odds, injuries, lineups, results, or sources. Missing data lowers the confidence gra
				>	de; it is never filled in.
				164	- If live odds are unavailable from every configured provider, report the blocker and do not create
				>	a ledger entry with fake market data.
				165	- Probabilities must sum to 1.00, and the ledger entry must exactly match the prediction of record.
				166	- Do not predict from stale or missing scorecards; fix them first under `instruction-updating`.