Head-to-head prompt diff

7af7ec26df84 → 207433de91f8

	7af7ec26df84			207433de91f8
	87	- Renormalize so the three probabilities sum to exactly 1.00 at two decimal places; absorb rounding		87	- Renormalize so the three probabilities sum to exactly 1.00 at two decimal places; absorb rounding
	>	in the largest probability.		>	in the largest probability.
	88			88
	89	### Step B5 — quant cross-check		89	### Step B5 — quant cross-check
	90			90
n	91	If the recorded-stats card includes a quant reference, compare it with your numbers. Any outcome dif	n	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		>	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		>	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		>	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. Because it ignores the quality of past opponents, i		>	reference is never the baseline; never copy it.
	>	t systematically over-rates a weaker side that has lately beaten weak or friendly opposition; when i
	>	t rates the underdog at or above the favourite, that is its blind spot, not evidence — never let it
	>	lift the underdog or trim the favourite below the Step B1 anchor.
	92			92
	93	### Step B6 — freeze		93	### Step B6 — freeze
	94			94
	95	Write the `BASELINE (pre-odds)` line with the confidence grade. The ledger entry must use exactly th		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.		>	e prediction-of-record numbers that survive the disagreement protocol.
	136	- probability synthesis, recommendation, and risk flags for this fixture		136	- probability synthesis, recommendation, and risk flags for this fixture
	137			137
	138	If a piece of research is only meaningful because of the opponent, keep it in the H2H report. If it		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 — but never with market prices.		>	changes the reusable understanding of the team, update the scorecard — but never with market prices.
	139			139
n			n	140
	140	## Post-mortem learning loop		141	## Post-mortem learning loop
	141			142
	142	When settled bets or prediction rows are reviewed, feed back durable observations as calibration que		143	When settled bets or prediction rows are reviewed, feed back durable observations as calibration que
	>	stions rather than mechanical overrides. Small samples can reveal recurring blind spots, but they sh		>	stions rather than mechanical overrides. Small samples can reveal recurring blind spots, but they sh
	>	ould improve the next baseline narrative instead of adding rigid rules.		>	ould improve the next baseline narrative instead of adding rigid rules.
	143			144
	148	- When backing a longshot against a strong favourite, is the upset path football-specific and source		149	- When backing a longshot against a strong favourite, is the upset path football-specific and source
	>	d, or is it mostly a price gap created by preserving generic underdog mass?		>	d, or is it mostly a price gap created by preserving generic underdog mass?
	149	- For confidence C predictions, what fact would need to be true for the model-market disagreement to		150	- For confidence C predictions, what fact would need to be true for the model-market disagreement to
	>	be credible, and what unresolved data gap could break it? Keep the ledger disagreement, but label t		>	be credible, and what unresolved data gap could break it? Keep the ledger disagreement, but label t
	>	he evidence quality plainly.		>	he evidence quality plainly.
	150	- In mismatches, separate team-strength gap from match-shape variance: low blocks, tournament-openin		151	- In mismatches, separate team-strength gap from match-shape variance: low blocks, tournament-openin
	>	g caution, finishing variance, goalkeeper paths, travel/logistics, and game-state incentives can sus		>	g caution, finishing variance, goalkeeper paths, travel/logistics, and game-state incentives can sus
	>	tain draw/upset probability without implying the weaker team is close in underlying quality.		>	tain draw/upset probability without implying the weaker team is close in underlying quality.
	151	- If a hedge or trade later changes the realised Betfair result, keep that separate from prediction		152	- If a hedge or trade later changes the realised Betfair result, keep that separate from prediction
	>	calibration. The methodology review should still judge the original pre-match probability record aga		>	calibration. The methodology review should still judge the original pre-match probability record aga
	>	inst the football outcome.		>	inst the football outcome.
t	152		t
	153	### Calibration note from settled 2026 World Cup bets (as of 2026-06-30)
	154
	155	The settled-bet record exposes one recurring failure, large enough to treat as calibration (not a ru
	>	le): backing the weaker side to win outright has gone about 2 from 43 (~5%, far below the ~20–25
	>	% those baselines implied), while backing the draw in comparable mismatches has been net positiv
	>	e (about 6 from 17, the winners being low-block or tournament-caution results such as 0-0s). In roug
	>	hly 57 of 68 rows the frozen baseline gave the market favourite less probability than the de-overr
	>	ounded market did. Read it as:
	156
	157	- The mechanism is favourite compression — trimming a strong side below its Step B1 anchor so th
	>	e weaker side keeps "value" mass it has not earned. Anchor the favourite from squad strength and let
	>	only sourced, fixture-specific facts within the B3 caps move it. Recent form (B3 = 0) and the oppon
	>	ent-blind quant reference (B5) must never pull the favourite down or lift the underdog.
	158	- When your honest disagreement is "the favourite is less certain than the price," express it throug
	>	h the draw / avoid-defeat channel and the risk flags, not as a back-the-underdog-to-win selectio
	>	n. Backing the weaker team to win outright needs a named, sourced upset path (key absence, an exploi
	>	table tactical mismatch, must-win versus dead-rubber stakes) — preserved generic mass is not a path;
	>	absent one, prefer the draw or no-bet.
	159
	160	This stays calibration, not a ban: a real, sourced upset edge is still backed and still recorded. Th
	>	e aim is only to stop manufacturing underdog edges out of favourite compression.
	161			153
	162	These are prompts for synthesis, not automatic probability moves. Apply them only when supported by		154	These are prompts for synthesis, not automatic probability moves. Apply them only when supported by
	>	the scorecards, recorded-stats card, and sourced fixture context, and document the reasoning in the		>	the scorecards, recorded-stats card, and sourced fixture context, and document the reasoning in the
	>	baseline, market comparison, or risk flags as appropriate.		>	baseline, market comparison, or risk flags as appropriate.
	163			155
	164	## Report structure		156	## Report structure