#Hold & Win Reference

Reference pack for the reusable hold_and_win mechanic. This game exists to specify, implement and verify the mechanic — it is not a themed product release.

#Design gate

DESIGN approved 2026-08-09 by user; scope bound to design spec c44bc01.

Implementation follows docs/superpowers/plans/2026-08-09-hold-and-win-reference.md.

#Mechanic at a glance

Grid:            3 columns × 5 rows (15 cells)
Cell order:      column 0 rows 0-4, column 1 rows 0-4, column 2 rows 0-4
Trigger:         6 or more BONUS symbols visible
Initial respins: 3
Landing chance:  15% per empty cell (draft)
Respin reset:    any newly landed prize resets remaining respins to 3
Miss-only step:  remaining respins decrement by 1
Termination:     respins reach 0, or all 15 cells locked
Ordinary award:  sum of locked prizes, capped once at 500× total bet
Full grid:       exactly Grand 500× total bet, replacing the sum

Locked cells are never redrawn and consume no RNG. The feature charges no additional wager.

#Current math

Confirmed on 2,000,000 rounds with an independent master seed — draft-v0.2 confirmation. Base RTP and trigger frequency are EXACT from enumeration of all 8000 reachable base grids.

total RTP     96.0031% +/- 0.1189pp   target 96.0% +/-0.3pp   PASS
feature RTP   20.8871%                required 20.8840%       PASS
cond mean     32.4593x +/- 0.1848     required 32.4545x       PASS
full-grid     0.090053% of triggers    band 0.05%-0.25%        PASS
              = 1 in 1,110 triggered features = 1 in 172,570 base rounds
base RTP      75.1160% EXACT          target 76%              unchanged
trigger       1 in 155.4035 EXACT     target 1 in 150         unchanged
max win       500.0000x observed, never exceeded              reachable and contained

Total RTP is quoted as exact base + exact trigger x simulated conditional mean; the all-simulated figure for the same run is 95.6695%, and the difference is attributed in the report.

Reached via two tuning passes. Pass 1 swept LandingChancePPM alone and proved no value works (the two bands are disjoint by ~13,500 PPM); pass 2, option A3, additionally moved the regular-value weights, leaving coin values, jackpot values and prize-type weights untouched.

#Runtime

The mechanic is played by the live spin orchestrator, not only by the simulator.

paid base spin, 6+ BONUS
  -> the durable feature is opened INSIDE the paid gameplay transaction
  -> each respin is its own request, carrying expected_hold_and_win_step_index
  -> no additional wager; PostgreSQL is authoritative, Redis is a post-commit cache
  -> settlement writes exactly one credit intent

Reference/runtime parity is executed, not asserted: the orchestrator's complete state — every cell, prize, newly_landed delta, respin counter, step index and settled award — is compared with RunHoldAndWinToCompletion at 0, 43,000 and 1,000,000 PPM. That is what makes the RTP figures above statements about production play rather than about a second implementation.

The client contract is gdd/hold-and-win-backend-response.md, generated from fixtures captured on the live orchestrator path.

#Evidence routing

• evidence/simulation/ — seeded aggregate simulation reports and manifests.
• evidence/parity/ — reference/runtime comparison.
• evidence/max-win/ — max-win bound validation.
• test-vectors/ — deterministic portable fixtures and their generation commands.

#Scope exclusions (v0.1)

No collector, no multiplier, no feature buy, no additional wager. Art, audio, animation and market/compliance matrices are out of scope for a reference pack.

#Known limitation

hold_and_win is RNG-driven per empty cell, so it is not grid-determined and exact enumeration is unavailable. All mechanic RTP evidence is seeded Monte Carlo, labelled estimated — full engine capability is a statement about what the runtime does, not a claim that the figure is exact. Internal tests and simulations are not external certification.