Reviewer Findings

Latest Quintace endpoint results on each spot, alongside the reference number the reviewer used (from GTO Wizard or another external solver). Some spots agree, some differ. The differences are open for discussion — we haven't pre-assigned whether the external solver or Quintace is correct.

PLO4 model opens UTG wider than MP at 100bb/5% rake — violating positional ordering. Brad's 2026-04-17 baseline showed a 0.8pp inversion (UTG 36.5% > MP 35.7%). The 2026-05-11 endpoint re-run on the same model shows the inversion has widened to 3.9pp (UTG 36.41% > MP 32.51%).

Cross-rake corroboration (added 2026-05-11): the PLO hygiene check in R10 reproduces the inversion at two additional rake settings — 3% rake / 3 BB cap and 0% rake — showing UTG 36.8% > MP 33.8% (3.0pp inversion). The monotonicity defect is rake-invariant, consistent with R10's separate finding that the PLO model is rake-insensitive on preflop opens overall.

B1 property gap discovered: A1 IS implemented in the production runner (with strict comparison) and SHOULD catch this. A1-strict (1pp noise floor) was added to the yaml spec but NOT implemented in the runner. A1 either doesn't sweep the rake parameter, or there's a payload-construction divergence — explaining why Nimit's Apr 23 cross-format sweep showed plo4-6max 55/55 PASS.

Quintace's latest endpoint c-bet rate on K72r matches GTO Wizard's published number within solver-noise tolerance.

Reviewer flagged that Quintace "shows little folding vs donk" compared to GTOw's ~20% fold reference at b33. Endpoint says otherwise.

Per-combo raise-action EV varies dramatically across folded hands at RFI positions — standard junk (72o, 82o, etc.) shows sensible negative raise EVs (−0.2 to −2.0 bb), but rarely-played buckets like 26o show absurdly positive raise EVs (e.g. 26o = +6.77 bb at UTG despite folding 100%). Same family as B1 KI-4 (model EV shape concerns) and G-series violations.

2026-05-11 re-run of B1 I-series properties against current production model showed I3 violation rate 28.5% (12,988 / 45,598) and I5 violation rate 66.5%. The EV-shape issue is not just 26o — it's pervasive across hand classes including premium ranges (AA call vs raise gap = 1.27 bb).

At Stand-Up val=10 / Scenario D, Klu's T9 verdict flips categorically: 100bb = call 35%, 226bb = fold 100% (−65pp swing across just 126bb of depth). NLHE cash control sweep (50-1000bb on UTG RFI) shows total swing of −1.73pp — depth-stable. The Stand-Up flip is ~16× larger in magnitude across ~8× less depth → Stand-Up-specific phenomenon, not a general model-depth issue.

B1 framework gap: B1 property B1 (stack-depth continuity) exists and applies to Stand-Up, but the production runner caps stack range at 100bb, tests only UTG : open, and doesn't sweep the val parameter. A new SQ4 property (stack-depth × val × state continuity) is needed to catch this automatically.

At deep Stand-Up Game stack depths (~1260bb in the Mariano AA hand), the model's verdicts flip fold ↔ jam ↔ call across small parameter perturbations and violate val-monotonicity. Root cause is a training-coverage gap; the OOD symptom is consistency violation. Article shipped only 500bb-validated queries as editorial mitigation.

First PLO hygiene pass mirroring the NLHE/MTT pattern. RFI VPIP vs two independent solver-grade references (100bb 6-max GTO PLO):

*BTN Upswing value: community consensus (less reliable than the two solver-grade refs); PLO Genius gives 48%. Both solver-grade references corroborate within 1.1pp at UTG/MP/CO — the divergence is robust across independent solvers, not a "Upswing happens to be tight" artifact.

Diagnosis — what the model gets right, and where the gap lives:

• Hand-strength understanding is structurally correct: trash hands fold 100% · premium hands raise 99%+ · AA/KK-in-hand opens 90% · double-suited beats rainbow (59.5% vs 15.4%) · ace-high beats low-card (60.2% vs 20–30%).
• Limps are negligible (0.02%) — the 36.8% UTG VPIP is essentially all raises, directly comparable to the "opening range" semantics in both Upswing and PLO Genius references.
• The +19.45pp UTG divergence is a shifted opening threshold — the model classifies more borderline hands as "open" than either reference solver does. Gap is largest at UTG, shrinks progressively to BTN (where Quintace lands in-range). Pattern fits "open threshold mis-calibrated wider in early positions."

Related (but distinct) findings: a separate monotonicity inversion (UTG > MP) is captured in R3. A separate rake-sensitivity diagnosis is in R16 (resolved — model is rake-flat at low rake, rake-aware at higher rake). R10 = absolute threshold too wide; R3 = positional ordering wrong; R16 = rake sensitivity correctly handled. R10's +18.9pp UTG over-loose still holds at heavy rake (R16 verified: +15.9pp gap at 10%/1cap).

Methodology notes (not findings):

• ploStrategy returns 856 canonical PLO4 classes (out of 270,725 specific combos), fixed across every parameter swept. Upswing and PLO Genius references are also class-aggregated, so comparisons are apples-to-apples.
• actionLabels[].strategy is NOT exposed for PLO (unlike NLHE/MTT) — per-combo Metrics.plo_preflop() aggregation is the only path; audit confirmed correct. Audit + diagnosis scripts: tests/probe_plo_methodology_audit.py + tests/probe_plo_query_diagnosis.py.
• Limp-inclusion confound (checked): our "VPIP" sums limp + raise, while Upswing's reference is raise-only. Model limp by position: UTG 0.0% · MP 1.0% · CO 1.2% · BTN 4.3%. At UTG (where the gap is biggest, +18.9pp), limp = 0 — the gap is pure raise-vs-raise. At MP/CO, excluding limp narrows the gap by ~1pp (CO 7.8→6.6pp, still WARN). At BTN, excluding limp closes most of the gap (4.8→0.5pp) — consistent with BTN's existing OK verdict in the table. Net: limp inclusion does not explain the UTG/MP/CO over-loose finding.
• Raise-size confound (checked): the model effectively opens at ~3.5× only (2.5× sizing used 0.8% of the time per query diagnosis). Upswing's "% of hands that open" is size-agnostic (any raise counts), so comparing open-frequency vs open-frequency is sizing-invariant.

What the model actually uses (corrected)

1. Payouts schedule shape — 5.14pp RFI spread across schedules at the same spot. Direction is physically correct (flat payouts → tighter; winner-take-all → wider):

2. Alive / entries ratio — threshold step around ~5% of field remaining. Above threshold, RFI ~13.33% (baseline). Below threshold, RFI snaps to 10.41% regardless of how it got there (16/1000, 9/1000, 200/100000 all give identical 10.41%). 3.39pp effect.

What the model still ignores (re-verified)

Implication for hygiene checks

External ICM solvers (ICMIZER, HRC) condition on per-seat ranks, payout slope, AND bounty $$ amounts. Our model captures payout slope and the alive-threshold step, but NOT per-seat ranks or prize-pool magnitude. Hygiene checks against ICM references need to match the payout schedule precisely (the dominant ICM-block input) and accept that rank-conditioned reference differences won't be reproduced.

Reference behavior — PKO and knockout bounty types correctly tighten RFI by 6–10pp (incentive to hunt elimination):

Mystery-bounty sweep — 11 configurations, every one returns RFI=11.83% (bit-identical to normal):

Verdict: spread = 0.00pp across 11 configurations including extremes. All payload JSON hashes differ (so the JSON payloads truly vary on the wire), but the model produces bit-identical output. bounty_type=mystery_bounty is a no-op — the model treats it identically to normal, with the auxiliary fields ignored regardless of value.

Verification sweep — 9 configurations, hand and mtt stack blocks varied independently:

Pattern: when hand.players[].stack is held constant, RFI is bit-identical regardless of what mtt.players[].stack contains. When hand.players[].stack changes, RFI changes dramatically. Within each "hand=X" group: spread = 0.00pp. mtt.players[].stack is silently ignored.

Classification: unlike R11 and R12 (model-behavior issues), R13 is a payload-schema redundancy. The model itself is not wrong — the payload format has two redundant stack fields where only one is actually consumed.

VPIP curve — BB defends vs BTN 2.0x by stack depth, both models on identical payloads (only stack varies):

Observations:

• Cash model is essentially depth-blind in this spot. 3.28pp VPIP range from 20bb to 100bb (a 5× change in stack depth). Theory expects substantial widening with depth — postflop playability is the main argument for defending more hands at 100bb than 20bb.
• MTT model produces the expected shape. 21.06pp VPIP range across the same depths, monotone increasing — short stacks defend tight, deep stacks defend wide. Direction and magnitude both look right.
• The two models cross around 40–50bb. At 30bb the cash model defends 9.5pp wider than MTT; at 100bb the cash model defends 2.8pp narrower. Two different shapes, not just a level shift.

Single-spot caveat: this curve is from one spot (BB vs BTN 2.0x). Before drawing strong conclusions about the cash universal model, this should be reproduced at other defense spots (BB vs CO, BB vs HJ, SB vs BTN) and at different open sizes. The pattern may be specific to BB-vs-BTN or generalize across cash defense; only further testing tells us which.

Comparison matrix

Sub-findings grouped under R15

(a) Both models under-open at deep + mid UTG stacks. Gap is 1.5–3.2pp for MTT model, 1.0–2.1pp for cash model at UTG 100/50/17 bb. Direction is consistent (both tighter than GTOw); magnitude is near solver-noise tolerance but the consistent same-direction bias across 3 stack depths suggests a real (small) calibration drift.

(b) MTT model gap widens dramatically at HJ 30bb. One non-UTG data point shows MTT model 7.1pp tighter than GTOw vs the ~2pp UTG gap. Suggests position-conditioned bias (wider gap at later positions). Caveat: single non-UTG spot; need MP, CO, BTN coverage to confirm the pattern.

(c) Cash model FLIPS sign at 5bb (over-opens by 4.66pp). Cash universal is 1–2.7pp tight at 100 / 50 / 17 / 30 bb, then +4.66pp WIDER at 5bb. Cash model uses 21.86% all-in at 5bb vs GTOw's ~20% total RFI (where almost all should be jams). The cash model isn't push-fold trained — once min-raise becomes structurally infeasible, it gambles all-in too liberally.

(d) MTT model has a sharp push-fold transition between 5bb and 17bb. At 100/50/17 bb, MTT model all-in usage is 0.000–0.004% (essentially never jams). At 5bb, all-in is 17.60% (essentially the entire RFI). GTOw transitions smoothly across this range; the MTT model has a binary switch. Caveat: no data at intermediate stacks (8–14 bb) — exact transition shape unknown.

(e) MTT model uses a different sizing tree at HJ 30bb. Both models have the same available sizes (2.0 / 2.5 / 4.5 / 6.0 / 9.5x). Cash model uses 2.0 + 2.5x as modal opens (10.97% + 15.66%). MTT model uses 4.5x as the single modal open (21.83%) with near-zero weight on 2.0x (0.003%). Same spot, very different sizing-tree resolutions.

Methodology

All 5 audits use identical parameter alignment to the GTOw reference: 8-max NLHE, symmetric stacks, ante=12 chips (0.12bb per player), 0 rake, chip-EV regime. Single known parameter mismatch: 0.5-chip ante rounding (12 vs 12.5 chips) — per ante sensitivity probe, accounts for ≤ 0.25pp of any gap. Reference numbers were re-verified by direct WebFetch of the primary GTOw blog posts (not relying on secondary citations). Side-by-side script + JSON outputs are at external-solver-benchmark/tests/single_spot_audit_*.{py,out.json}; each script prints the full payload for spot-check.

Open scope (what would strengthen this finding)

• Fill in UTG curve at 30bb, 20bb, 14bb, 2bb (all have GTOw published numbers) to validate the curve shape.
• Add MP, CO, BTN at 30bb chip-EV to confirm (or refute) position-conditioning of the MTT model gap.
• Add 8bb, 10bb, 12bb, 14bb data points to characterize the MTT model's push-fold transition shape.
• Verify the cash-model 5bb over-jam holds across positions (not just UTG).

Three-step diagnosis:

1. Initial observation (preflop, 2 rake values): PLO preflop RFI output was identical across 0% rake and 3% / 3 BB cap — hypothesized "rake silently ignored."
2. Postflop follow-up (BTN c-bet, 3 boards × 5 rake settings): postflop DOES respond to rake (2.10–4.80pp range per board). So endpoint consumes rake fields; field-strip hypothesis refuted.
3. Per-position preflop sweep (5 positions × 5 rake settings): preflop ALSO responds when swept at heavier rake values. Original "invariant" claim was a sampling artifact — 0% and 3%/3 BB cap happen to produce identical output (3 BB cap rarely binds on small preflop pots, making both effectively rake-free for preflop EV).

Full picture — preflop VPIP by position × rake:

Three takeaways:

• Low-rake equivalence is correct. 0% and 3%/3 BB cap produce identical output across all 5 positions — both are functionally "no effective rake" at small preflop pot sizes. The model correctly treats them as equivalent.
• Higher-rake response is real and directional. At 5%+ rake, the model tightens preflop ranges 3–4pp across non-SB positions; SB tightens 7.5pp. Direction is correct (heavier rake → tighter range).
• SB rake-sensitivity is notable. SB shows 2× the rake response of other positions. Consistent with GTO intuition — SB has positional disadvantage, so every BB of rake hurts SB EV more than other positions; tightening opens compensates.

Postflop response (sub-data): BTN c-bet across 3 boards × 5 rake settings — bet% range 2.10pp (Ts9s8s) / 3.80pp (KsKd2c) / 4.80pp (Ks7d2c). Postflop also responds, as expected.

The KVL register entry claimed BTN open avg_raise_bb = 4.8 BB, well above Pio/GTOW's expected 2.0–3.0 BB. The 4.8 figure was logged previously against an older model; the current endpoint output is sharply different.

Sizing breakdown (when raising): 71.5% at 2.5× · 26.9% at 3.5× · 1.7% at 5.0×. Weighted average = 2.81 BB, right in the middle of the Pio/GTOW 2.0–3.0 expected range.

Three-bet sizing is also sensible: vs 2.5x, BB mixes 11 BB (8.87% prob) and 16.5 BB (4.75% prob); vs 3.0x, BB shifts to 13 BB primary (9.99% prob) with a tiny 19.5 BB tail. Sizes scale appropriately with the open size.

Size-sensitivity sanity check: BB tightens 2.56pp from 13.82% vs 2.5x to 11.26% vs 3.0x — correct GTO direction (defense tightens against larger opens).

BTN c-bet frequencies captured across 7 paired flop classes — kept here as raw observations awaiting an external PLO solver reference to validate against:

Why this is NOT classified as an Issue:

• Each paired board has a genuinely different texture (different ranks, different kickers, different equity distributions, different BB cap dynamics, different reach configurations on both sides). The earlier "structurally equivalent" framing in my first draft was wrong — solver outputs for AsAd7h vs KsKd7c can legitimately differ substantially. PLO solver outputs are known to vary 20–40pp across boards that look similar at the rank level.
• Without a paid PLO solver reference (Monker, Pio PLO, PLO Genius, PLO Mastermind, Vision GTO Trainer — all subscription) producing comparison numbers for these exact boards, there's no external anchor to call any single number "right" or "wrong."
• Heuristics like "high pair should c-bet more than low pair" are domain knowledge, not solver-validated claims. PLO postflop GTO frequently inverts NLHE intuition.