Code review caught a regression in the single-Elo blend: resolveSourceElos /
resolveRatings clamped every resolved value to [eloMin, eloMax] /
[ratingMin, ratingMax], including a directly entered base Elo. The original
resolver only clamped derived values (projections, odds, fallbacks) and passed
direct inputs through untouched. With the default 1900 Elo ceiling this
truncated manual Elos for sports that enter higher ratings (e.g. snooker ~2450),
collapsing their spread.
Clamp only odds-derived / blended / fallback values now; direct Elo and rating
inputs pass through as entered. Adds a regression test, and fixes two comments
that still referenced the removed `sourceEloPriority` field (now `oddsWeight`).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QhNeB7gN6VKene7sdbBVQT
Replace the two parallel odds/Elo mechanisms with one model: every strength
source (raw Elo, projected wins, projected table points, futures odds) converts
to an Elo, those blend by weight into a single Elo, and that one Elo feeds every
simulator. This supersedes the earlier source-priority + per-match probability
blend approach.
Resolution (app/services/simulations/input-policy.ts):
- SimulatorInputPolicy gains baseEloPriority (order among the substitutable base
sources: raw Elo / projected wins / projected table points) and oddsWeight
(0–1). resolveSourceElos/resolveRatings now compute baseElo, derive an
odds Elo via convertFuturesToElo, and blend: 0 = base only, 1 = futures
override, between = weighted blend (method "blend").
- Drops the odds-inclusive sourceEloPriority and the prefersFuturesOdds helper.
Simulators consume the single resolved Elo:
- UCL, World Cup, NCAA Football, MLB drop their separate normalized-odds signal,
convertFuturesToElo calls, and per-match probability blend; they read the
resolved sourceElo (preserving each sim's hardcoded fallback Elo table). The
optional SimulationContext oddsWeight plumbing (types.ts/runner.ts) is removed.
- UCL is routed through the central blend (requiredInputs sourceElo, derivable
from sourceOdds) so any entered Elo and futures blend uniformly.
- College hockey already blends odds into Elo internally (and uses NPI rank the
central resolver can't), so its central oddsWeight is set to 0 to avoid
double-counting; the simulator is unchanged.
Manifest: per-profile oddsWeight defaults (World Cup/UCL/MLB 0.3, NCAA FB 0.4,
college hockey 0; global default 0.3).
UI: the Input Policy card exposes one "Futures vs. Elo — Odds Blend Weight"
control; the /admin/simulators inventory badge shows the effective blend
("Elo only" / "NN% blend" / "overrides Elo") with the odds participant count.
Tests: input-policy blend math (0/0.5/1) for Elo and ratings, baseEloPriority
and oddsWeight parsing/clamping, manifest per-profile weights; obsolete
source-priority and oddsWeight-context tests removed/replaced.
Note: this intentionally shifts the calibrated EV outputs of the four sims that
previously blended at the probability level (accepted in design discussion).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QhNeB7gN6VKene7sdbBVQT
Make the futures-vs-Elo relationship an explicit, configurable rule and fix
futures odds failing to override stored Elo.
Root causes addressed:
- resolveSourceElos/resolveRatings used a hardcoded precedence (direct Elo ->
projectedWins -> projectedTablePoints -> odds). The prior fix nulled sourceElo
on futures entry but not projections, which still outranked odds.
- The override relied on a destructive null-on-save hack that also wiped
manually entered Elo.
- Blended simulators buried their Elo/odds weight in module constants.
- Futures odds were not surfaced on the /admin/simulators inventory.
Changes:
- Add a configurable source policy: sourceEloPriority (ordered) and oddsWeight,
parsed/clamped in getSimulatorInputPolicy. resolveSourceElos/resolveRatings now
resolve each participant by the configured priority instead of a fixed order.
Futures-centric simulators (ncaam, ncaaw, world_cup, ncaa_football,
college_hockey) default to a futures-override priority.
- Drop the destructive nulling in batchSaveFuturesOddsForSimulator and
batchSaveSourceOdds; override is now governed by policy, preserving stored Elo.
- Thread an optional SimulationContext (oddsWeight) through the Simulator
interface so blended sims (UCL, World Cup, NCAA FB, MLB) read the blend weight
from the season policy; defaults preserve prior calibration when no context is
passed.
- Add a "Futures vs. Elo" strategy control and Odds Blend Weight input to the
Simulator Setup input-policy card, persisted via save-input-policy.
- Surface futures on /admin/simulators: a source badge and a Futures Odds quick
link; extend listSportsSeasonSimulatorSummaries with odds source info.
- Tests: configurable priority override (Elo/projections/rating), oddsWeight
parsing/clamping, prefersFuturesOdds, manifest defaults, an NCAA Football
context-blend behavioral test, and an updated non-destructive save test.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QhNeB7gN6VKene7sdbBVQT
When an admin entered futures (preseason) odds for a season that already
had Elo ratings stored, the simulator kept using the old Elo and silently
ignored the new odds. This affected any Elo-based simulator (e.g. NHL).
Root cause: resolveSourceElos() ranks a direct sourceElo above the
sourceOdds -> convertFuturesToElo branch, but batchSaveFuturesOddsForSimulator()
only cleared the bracket-seeding `rating`/`ratingMethod` — never the stale
`sourceElo`/`sourceEloMethod`. A manually entered Elo (method "direct") is not
treated as generated, so it survived and short-circuited the resolver.
Fix:
- batchSaveFuturesOddsForSimulator now also nulls sourceElo and strips
sourceEloMethod (both the pre-update and upsert-conflict paths), so the
existing futures -> Elo conversion drives the run.
- resolveSourceElos' sourceOdds branch now guards for >= 2 participants
(mirroring resolveRatings), so a lone-odds season falls through to the
configured missing-Elo strategy instead of getting a flat ~1500.
- batchSaveSourceOdds clears the legacy EV sourceElo and marks source as
futures_odds so the elo-ratings page won't resurrect a stale rating.
Adds unit coverage for odds-derived Elo, the single-participant guard, the
post-clear regression, generated-vs-direct sourceElo suppression, and the
new clearing behavior in batchSaveFuturesOddsForSimulator.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YNfUEd9RzD3zm84oLHBHUH
- Revert ncaaw-simulator to Barthag win probability formula; set
realistic rating bounds (ratingMin: 0.70, ratingMax: 0.97) so derived
ratings stay in the range where the formula behaves well
- Add batchSaveFuturesOddsForSimulator which clears all ratings (manual
and generated) before upserting sourceOdds, so futures odds always
drive the simulation rather than being silently overridden by existing
Barthag ratings from Simulator Setup
- Add clearSourceOddsForParticipants to zero out both tables for
participants excluded from a bulk import
- Add "Clear existing odds" checkbox to the bulk import card; applies
client-side on match and server-side on submit
- Fix missing sportsSeasonId filter in batchSaveFuturesOddsForSimulator
pre-clear UPDATE (could have wiped ratings across other seasons)
- Fix race condition: run batchSaveSourceOdds then
batchSaveFuturesOddsForSimulator sequentially so the simulator inputs
table always ends in the correct cleared state
- Fix Math.round in convertFuturesToElo collapsing Barthag-scale ratings
to 0 or 1; Elo callers already round after clamping
- Handle all-identical-odds edge case in convertFuturesToElo (assign
midpoint instead of throwing)
- Add missingRatingStrategy: worstKnownMinus to ncaaw_bracket manifest
so fallbackRatingDelta is live config, not dead
- Log warning in resolveRatings when only 1 participant has odds
- Reset clearExisting checkbox after applyMatches
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Introduces three new schema tables (simulator_profiles,
sports_season_simulator_configs, season_participant_simulator_inputs),
a central model layer (app/models/simulator.ts), and a single runner
entry point so every simulator run follows the same prepare → simulate
→ persist → snapshot → recalculate flow.
Key additions:
- manifest.ts: per-simulator display names, default configs, required/
optional inputs, derivable-input declarations, and setup sections
- input-policy.ts: resolves sourceElo from projectedWins,
projectedTablePoints, or sourceOdds; resolves ratings from sourceOdds;
supports block / fallbackElo / averageKnown / worstKnownMinus strategies
- runner.ts: single entry point for admin simulation runs; materialises
derived inputs, normalises result columns, zeroes omitted participants,
snapshots EVs, and recalculates linked fantasy standings
- /admin/simulators: inventory page with per-season readiness and bulk run
- /admin/sports-seasons/:id/simulator: per-season setup page with readiness
summary, input-policy editor, raw JSON config override, and CSV bulk input
- NCAAM/NCAAW simulators now read ratings from season_participant_simulator_inputs,
falling back to the hardcoded name-keyed maps while DB data is being populated
- Clone flow copies simulator config by default; volatile inputs (odds, Elo)
only copied when explicitly requested
Code-review fixes included in this commit:
- source field in compatibility bridge checked with !== null instead of !== undefined
- sourceEloRequirementLabel no longer appends "configured fallback" when the
participant is already excluded from all resolved sources
- Duplicate inline label maps in input-policy.ts replaced with simulatorInputLabel
- save-config preserves existing inputPolicy when the submitted JSON omits it
- Input table truncation label added (Showing 20 of N)
- CSV description notes values must not contain commas
- N+1 comment added to listSportsSeasonSimulatorSummaries
- assertRegistrySchemaDriftFree called in manifest tests
- Runner test suite added covering happy path, already-running guard,
readiness failure, empty results, and error recovery with status reset
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
