Desperation Polarization: The Cliff and the Texture Flip

§1 — The spot

You sit at a six-handed Squid Classic table. Fifteen hands in. Each of the other five players has scooped a main pot. Each holds a squid. You haven't, and you don't. If the game ends now, you alone pay the full penalty.

This is hero-last. Every other player has cleared the one bar the format asks them to clear. They can fold junk with no forward-looking penalty to protect. You are the only player whose next decision still carries the full weight of the game-end math.

No-Limit Hold'em has no analogue. Cash treats every hand as independent; folding costs zero until the next deal. The solver's NLHE preflop equilibrium moves smoothly — hand strength against range against sizing, one axis, no discontinuities. Pocket eights plays like pocket sevens. KQs sits in the middle of most BB cold-calling ranges.

Squid Classic adds a forward-looking penalty term. In the hero-last state that term dominates. The solver's response is not "widen a cash range." It is a new equilibrium shape, and it shows up twice. Preflop, the solver draws a raise-or-limp cliff between 88 and 77 — pocket eights raises 74.9% of the time at val=3; pocket sevens drops to 27.0%. Postflop, the same mechanism flips the texture order: a monotone K94ss flop that's 34.6% c-bet in Cash becomes 87.6% c-bet in Squid hero-last at val=3 — a +53.0pp swing, the largest positive Cash→v3 shift of any board tested.

This article walks that evidence across both layers.

§2 — The mechanism

NLHE's preflop equilibrium is one-dimensional. Each hand's expected value depends on equity against the opponent's calling range, effective stack, position, and sizing. Move one rank down in pocket pairs and EV changes gradually. KQs to KJs changes gradually. The solver mixes and the mixes drift.

Squid's hero-last state bolts a second axis onto the chip-EV calculation. Every hand now carries a forward-looking squid-equity term — a measure of how much winning this pot contributes to avoiding the game-end penalty. That term doesn't scale smoothly with hand strength. It gates. Either the hand has enough showdown equity to survive a call — collecting a main pot that would otherwise be forfeited — or it doesn't.

Against safe opponents (every opponent, in the hero-last state), the fold-equity [KB:fold_equity] math compounds. Safe opponents already hold a squid; they have no forward-looking pressure to defend marginal hands. A raise forces a decision; many marginal holdings fold. A limp gives a cheap look at the flop, preserves the safe opponent's positional advantage, and invests chips without generating fold equity. If a hand clears the equity floor, raise; if it doesn't, fold; if it's close, limp.

That is the preflop cliff.

The same mechanism reshapes the flop c-bet decision. The hands BB adds to its Squid-widened defending range — junk that would fold in Cash — can't connect on every flop. On monotone boards specifically, the added junk hands almost never have a matching suit. When CO c-bets, those junk combos fold. CO's fold equity compounds, and the c-bet frequency on monotone K-high jumps from 34.6% (Cash) to 87.6% (Squid v3) — a +53.0pp swing, the largest positive Cash→v3 delta in this article's five-board scan.

Book-2 Part 6 grades the broad polarization direction [KB:polarization_threshold] (preflop and postflop) as [T1] — meaning graded authoritative in book-2's Layer-8 verdict pass (the strongest evidence tier this article series cites). The precise location of the preflop cliff at val=3 — in the 88/77 neighborhood rather than 99/88 or 77/66 — is empirically observed from the hero-last defense batch and carries an [AMBIGUOUS] qualifier (book-2 shorthand for "direction trusted, exact threshold data-thin") on the exact floor rank; it moves with val. The aggregate postflop c-bet rates confirm the broad direction. Per-combo postflop resolution landed in v2 (see §4a and §6) with two signal clusters; turn/river and BB-as-hero postflop remain out-of-scope per the rail's API shape.

Source: fresh pulls from pull_squid_hero_last.py on the Gameplay-AI preview endpoint; cross-referenced to output/book-2/part-6-hero-last.md Research notes, causal-explanations.md §M8, and squid-deltas.md Table 3. See Methodology and caveats.

§3 — The preflop cliff

The widget above is the slice. Click any row to see the mechanism. The story the bars tell:

AA–TT raise near-pure in both Cash and Squid. Premium pairs sit far above the equity floor; no regime shift moves them. AA raises 100.0% in Cash and 100.0% in Squid v3 (under the §6 Raise% definition — action_8 + raise 16.2 BB). The top of the range is the same in both worlds.

99 is the first row where NLHE and Squid separate cleanly. Cash raises 0.9% (almost pure cold-call); Squid v3 raises 98.4%. NLHE plays 99 as a set-miner with implied odds. Squid treats it as a clean raise — above the floor, commit.

88 is the top of the cliff at val=3. Cash 0.0% raise → Squid v3 74.9% raise. Three-quarters of the time, the solver raises; the remaining 24.3% calls; 0.8% shoves (tracked separately from Raise% per §6). Nothing folds. The +74.9pp shift is the canonical cliff entry.

77 is where the cliff falls. Cash 0.2% raise → Squid v3 27.0% raise, 72.8% call, 0.2% shove. Adjacent hand, opposite action. The solver calls three-quarters of the time. This is the pair that tells you where the cliff actually sits — not a pair-rank label (88/77) but a direction: graded authoritative ([T1]) on "raise strong, fold or limp weak, with a clean equity floor somewhere in the 88/77 neighborhood" and [AMBIGUOUS] on the exact rank.

66 is below the cliff: Cash 0.3% → Squid v3 3.9% raise, 96.1% call. Call-or-fold in both worlds; Squid just refuses to fold it (the squid-equity term pulls it into a defending range that Cash folds 48.2% of aggregate).

KQs shows that the cliff isn't a pair-rank phenomenon. Cash 0.7% raise → Squid v3 87.5% raise. The mechanism is identical to 88's — showdown equity clears the floor, and the fold equity broadway generates against safe defenders is especially large. KQs is one of the best fold-equity-generating hands in the deck once BB is facing opponents with no forward-looking pressure.

JTs at Cash 3.5% → Squid v3 86.5% repeats the KQs pattern. The cliff pulls broadway into near-pure raises the same way it pulls 88 in.

Reading across val: at v1 the solver is already polarized (defense 81.4% vs Cash's 51.8%) but the cliff is softer — raise frequency climbs smoothly. At v3 the cliff is sharpest. At v5–v10 the solver flips toward raise-dominant across the board (56.4% raise aggregate at v10 vs 39.2% call). The cliff is a mid-val phenomenon; at extreme val the whole range converges to raise.

§4 — The postflop texture flip

The grid above rotates the camera. Same regime — hero-last in the Squid mix — but now seen through CO's c-bet decision after BB defends. Five board textures. Five val columns. Click any board to see what makes the curve move or not.

The canonical story the grid tells is K94ss (monotone K-high [KB:monotone_board]).

In NLHE, K94ss is the board CO c-bets least in this five-board set — 34.6% frequency, less than half of the 66.4% on A94r and less than half of the 87.0% on K72r. The intuition is familiar: monotone flops are scary because BB's defending range contains every suited broadway and suited connector, so flush draws crowd the board. CO's range advantage [KB:range_advantage] partially evaporates; check-backs preserve equity.

Squid inverts this. At val=1, K94ss c-bet jumps to 75.2%. At v3, 87.6%. At v5, 92.6%. At v10, 96.4%. The Cash→v3 delta is +53.0pp — the largest positive c-bet delta among the five boards tested. A board NLHE theorists spend careful pages convincing you to c-bet less becomes a board Squid hero-last solvers c-bet on par with the driest K-high rainbow.

The mechanism is the same one driving the preflop cliff, seen from the opposite side of the table. BB's Squid-widened preflop defending range (defense 51.8% → 95.8% at v3) adds offsuit broadway and offsuit low connectors that folded in Cash. On a monotone K94ss flop, that added junk almost never has a spade. When CO c-bets, that junk can't continue — no pair, no draw, no backdoor, no share of the flush. CO's fold equity against the monotone-added junk is nearly 100%. Because Squid's widening contributes disproportionately to the offsuit junk category, the average fold equity on CO's c-bet is disproportionately large. Cash was under-utilizing fold equity on monotone flops; Squid's widening makes the correction visible.

Dry-rainbow ceiling. K72r is already 87.0% c-bet in Cash. Squid nudges it to 97.1–97.5% across all val. Flat past v1 — Pattern A: "already maximal, no headroom."

A-high rainbow. A94r goes 66.4% → 93.4% (v1) → 98.2% (v3) — a large relative lift (+31.8pp to v3) because the ace blocks BB's continuing range, but CO was already c-betting at a healthy clip. This is the supporting case, not the canonical one.

Connected wet. T98 two-tone goes 63.8% → 79.2% at v3, then creeps to 85.3% at v10. Smaller delta (+15.4pp to v3) because connected textures give BB's wider range actual equity — straight draws, pair+gutshots, two-pair combos — that partially offsets CO's fold equity. Book-2's M8 postflop forward prediction specifically tested this connected-texture protection and found BB's defense rate below what the aggregate mechanism predicted.

Monotone low. 652ss goes 44.3% (Cash) → 91.4% (v3) — the other monotone board in the set, +47.1pp. Confirms the mechanism isn't about K-high specifically; it's about monotone textures where BB's added offsuit junk has nothing.

Reading across val on K94ss: the c-bet frequency climbs monotonically (75.2 → 87.6 → 92.6 → 96.4). Unlike the preflop cliff, which is a mid-val phenomenon, the postflop texture inversion gets more extreme as val rises. The forward-looking penalty keeps widening BB's defense, keeps adding junk, keeps feeding CO's fold equity. At v10, K94ss is effectively an always-c-bet board.

§4a — Per-combo: the short-stack bluff-shove vehicle

The aggregate grid above asks "how often does CO c-bet?" A per-combo pull asks the charter question — "which exact combos does the solver pick to c-bet, at what size, how pure?" A second scan, pull_h2_percombo_postflop.py, ran 75 cells of CO flop c-bet resolution (five boards × four vals × three stacks, plus cash baselines per (board, stack); squid hero_last state, opener-as-hero; flop-only — the API-shape caveat is logged in §6). Two signal clusters surface. They share a mechanism family with the aggregate finding but specify its shape at the combo level.

Cluster 1 — short stacks, dry K-high and monotone-low, the canonical bluff-shove vehicle. At stack = 8bb, the solver's largest available bet label is bet5.5 — roughly 70% pot and effectively a flop shove given the remaining stack depth. The allin action is almost unused in these cells (single-digit basis points for individual combos); the solver picks the bet5.5 sized label instead. Under a max-size-bet-mass reading (per §6, scoring rule: the probability mass the solver places on its largest bet label, not the nominal allin field), the shove-vehicle pattern is dense. At K72r / val=5 / stack=8, 263 low-equity combos park most of their mass on bet5.5; at 652ss / val=5 / stack=8, 207; at K72r / val=10 / stack=8, 163; at 652ss / val=10 / stack=8, 171. 98s on K72r stack=8 at val=5 is a canonical example: bet5.5 at 67.9%, bet2.8 at 21.6%, bet1.8 at 10.3%, check 0.04%. No pair, no draw, no pot-odds equity story against BB's defending range — the solver shoves the max size because the fold-equity side of the math, against squid-wielding BB with no forward-looking pressure to defend junk calls on a dry K-high, pays for the investment. This is the desperation-polarization thesis at combo resolution, on the street it was originally charted for.

Cluster 2 — T98 100bb at val=10, Jx junk as premium-sized bet. A separate pattern appears on the connected-wet texture at deep stacks and extreme val. On T98 at stack=100 and val=10, the solver picks a 13.8bb sized bet — a pot-ish sizing, not a shove — at 98%+ frequency with a cluster of Jx weak holdings: J3s (3hJh) at 98.8%, J2s (2hJh) at 98.3%, J2o (2sJh) at 98.2%, and adjacent Jx combos. Cash baseline for the same combos: 0%. 204 combos total read this way. Mechanically, this is not a shove — 13.8bb into a roughly 6bb pot at 100bb effective leaves plenty behind. It also isn't the short-stack fold-equity math. It looks like a squid-tax premium sizing on a connected-wet board where the solver has a specific backdoor-straight / pair-plus-Ten-blocker story to tell and the val=10 penalty term widens CO's aggression ceiling into territory Cash never reaches. Flagged as a second observation distinct from Cluster 1 — a separate surface in the per-combo data, not confirmation of the short-stack thesis. A v3 pass could probe turn/river resolution for the Jx line (not buildable through the current rail — see §6).

§5 — The diagnostic

Four rounds. Two preflop, two postflop. Pick before you reveal. The answers the rounds teach assemble into the diagnostic.

Rules that come out of the data:

1. Count the squids before every decision, preflop and postflop. Am I safe? How many safe opponents around me? If I'm the only no-squid player — hero-last — NLHE preflop habits misread 88 / KQs / JTs, and NLHE postflop intuition misreads monotone boards. The regime shift is active on both streets.

2. Preflop: raise or fold-limp, not in between. The cliff makes polarized play mandatory. If you're hero-last and reaching for a "standard 2.25bb open" with a marginal hand, the solver is telling you raise pure or not at all. A small raise with a hand below the equity floor is the worst of both worlds — chip investment without the fold equity to justify it.

3. Postflop: the texture order flips. In Cash, K94ss is a low-c-bet board and K72r is a high-c-bet board. In Squid hero-last, they converge — K94ss rockets to 87.6% and K72r saturates at 96.6%. When you enter a hero-last postflop spot on a monotone flop, c-bet is the default. When the flop is connected wet (T98), slow down — the mechanism's magnitude is smaller there and BB's widened range has more to work with.

4. Don't trap with premiums. AA raises 100% at v3; 99 raises 98.4%. The big postflop shift is not value — those hands always go — it's bluff frequency. CO's c-bet range in Squid includes far more bluffs on monotone boards than Cash, because the fold-equity math calls for them. Hero-last is a win-a-pot-now regime on both streets.

Based on general poker theory: the cliff shape this article describes is a structurally new equilibrium object. It has no direct NLHE analogue because NLHE has no forward-looking penalty term. What it most formally resembles is ICM-adjusted play [KB:icm_math_vs_chip_math] in a late-stage MTT — a regime where chip-EV no longer tells the whole story and a second equity axis dominates. But Squid hero-last is a cash-shaped format with an MTT-shaped game-end incentive, and the preflop cliff plus postflop texture flip are the equilibrium's way of saying so.

For a GTO-Lab-caliber reader: the findings are testable, specific, and graded authoritative ([T1]) in book-2 Part 6. Every number in this article comes from fresh pulls through two sibling scripts on the Gameplay-AI preview endpoint via strategy_grid_client.py: pull_squid_hero_last.py (aggregate layer + per-hand preflop) and pull_h2_percombo_postflop.py (per-combo CO flop c-bet). The per-combo postflop question — which exact holdings CO c-bets as bluffs on dry K-high short-stacked, and what 13.8bb Jx is doing on T98 100bb val=10 — is named in §4a. The still-scoped-out slice (turn/river resolution, BB-as-hero postflop) is named in the methodology footer.

§6 — Methodology and caveats

Data rail. Every number in this article comes from a fresh solver pull against the Gameplay-AI range-viewer preview endpoint (https://preview.rlserv.aceguardianrl.com/api/strategy_grid), driven by strategy_grid_client.py with SquidConfig.hero_last('BB', val=V, num_players=6). The pull script is pull_squid_hero_last.py in this article's folder; cached responses live at c:/tmp/sg_cache_articles. A preflight at c:/tmp/nick-squid-preflight/preflight.py ran 2026-04-22 and confirmed that Cash (mode="normal") and Squid hero-last (mode="squid" + SquidConfig) return materially distinct outputs at both preflop (aggregate defense 51.8% vs 95.8% at v3) and postflop (K94ss c-bet 34.6% vs 87.6% at v3) — ruling out a KI-14-style endpoint pass-through. Per METHODOLOGY §4c (rail discipline), §4d (prefer fresh runs over lookups), and §4f (rail-divergence policy — fresh endpoint always wins), this article uses the endpoint rail directly and does not route through MCP; book-2 v1.8.0 locked-table values are cross-references, not primary sources.

Cash-baseline provenance (rail-divergence policy). Per METHODOLOGY §4f ("rail-divergence policy — fresh endpoint always wins," established 2026-04-22), all c-bet numbers in this article — including K94ss Cash, K94ss Squid v3, and A94r Cash — cite the fresh preview-endpoint pull, not the book-2 v1.8.0 locked table values. K94ss: Cash 34.6%, Squid v3 87.6%, Cash→v3 delta +53.0pp. A94r: Cash 66.4%. These match the numbers in uri-nlhe-reflexes-to-unlearn (K94ss, co-re-pulled 2026-04-22 per §4f) when both articles triangulate to the same endpoint on the same date. The book-2 v1.8.0 locked tables will re-bake on book-2's own version cycle; cross-article citation integrity is preserved through the shared fresh-pull rail, not through a locked-snapshot handshake. Deviation-log.md D-6 Resolution sub-block (§4f compliance) documents the reconciliation triggered by the c-rail-divergence-methodology-convention finding; prior D-6 Option-B overrides have been reverted.

Measurement conditions. 6-max, 100bb effective. Preflop: BB defending CO 2.5bb open (BvBB single-raised pot), BB-last position. Postflop: CO c-bets after BB checks on an SRP flop. Hero-last state throughout = hero is the only no-squid player (SquidConfig.hero_last("BB", val, num_players=6)).

Metric definitions. Raise% (the canonical single definition applied throughout prose, methodology, and widgets): at a given preflop hand node, Raise% = the solver's reported action_8 frequency + the solver's reported raise 16.2 BB frequency — the sum of all non-call, non-shove raising activity across the two raise-size slots the endpoint returns. Shove% = the solver's all-in 100.0 BB frequency — reported separately when material (88 Squid v3: 0.8%; 77 Squid v3: 0.2%; 99 Squid v3: 0.1%). Call% = call 2.5 BB. Fold% = fold. c-bet% = CO's flop bet frequency as the preflop aggressor facing BB's check. val = the Squid Classic penalty parameter; trained values {1, 2, 3, 5, 10} BB per game-end payout. Cash = mode="normal" on the same endpoint. A prior v1 draft mixed the narrow (raise 16.2 BB only) and broad (action_8 + raise 16.2 BB) definitions across sections; the broad definition above is now applied uniformly.

Headline numbers (all values as used in prose and widgets).

• Preflop BB aggregate defense vs CO 2.5bb: Cash 51.8% → v1 81.4% → v3 95.8% → v5 99.4% → v10 100.0%.
• Preflop BB aggregate raise: Cash 12.6% → v1 23.4% → v3 30.2% → v5 39.3% → v10 56.4%.
• 88 at v3: raise 74.9% / call 24.3% / shove 0.8% / fold 0%. Cash: call 99.8% / raise 0.0% / fold 0.2%.
• 77 at v3: raise 27.0% / call 72.8% / shove 0.2%. Cash: call 99.8% / raise 0.2%.
• 99 at v3: raise 98.4% / call 1.5% / shove 0.1%. Cash: call 99.1% / raise 0.9%.
• KQs at v3: raise 87.5% / call 12.4%. Cash: call 99.2% / raise 0.7%.
• JTs at v3: raise 86.5% / call 13.4%. Cash: call 96.4% / raise 3.5%.
• AA at v3: raise 100.0%. Cash: raise 100.0% (action_8 96.3 + raise 16.2 BB 3.7).
• K94ss c-bet (CO, SRP): Cash 34.6 → v1 75.2 → v3 87.6 → v5 92.6 → v10 96.4 (fresh endpoint pull per METHODOLOGY §4f).
• A94r c-bet: Cash 66.4 → v1 93.4 → v3 98.2 → v5 99.3 → v10 99.7 (fresh endpoint pull per METHODOLOGY §4f).
• K72r c-bet: Cash 87.0 → v1 97.1 → v3 96.6 → v5 96.2 → v10 97.5 (already-saturated ceiling).
• T98 c-bet: Cash 63.8 → v1 78.5 → v3 79.2 → v5 79.8 → v10 85.3 (connected wet, smaller delta).
• 652ss c-bet: Cash 44.3 → v1 86.8 → v3 91.4 → v5 89.7 → v10 86.3.

[T1] vs [AMBIGUOUS] qualifiers preserved. [T1] in book-2 Part 6 is the strongest-evidence verdict grade the research layer awards (source-authoritative, reproducibly measured). [AMBIGUOUS] is the shorthand used where the mechanism direction is trusted but the specific location/threshold remains data-thin. In this article, the broad polarization direction — preflop "raise strong, fold or limp weak with a clean equity floor" and postflop "c-bet frequency inverts on monotone boards because of BB-range-composition effects" — is [T1]. The specific location of the preflop cliff at val=3 (88/77 neighborhood) is [AMBIGUOUS]: it moves with val (toward 99/88 at v1 and toward an across-the-board raise-dominant equilibrium at v10). The article treats the 88/77 neighborhood as approximate, not a hard rule at a specific rank.

Per-combo postflop — landed in v2 (§4a). The initial v1 scan pulled aggregate postflop only. The v2 pass ran pull_h2_percombo_postflop.py against strategy_grid_client.Postflop.cbet with SquidConfig.hero_last("CO", val, num_players=6) — 75 cells (5 boards × 4 vals × 3 stacks, plus 15 cash baselines per board-stack), 0 failures, 81.6s wall time. Output cached at pull_h2_percombo_postflop.out.json alongside the aggregate pull. Two API-surface constraints are logged as scope, not failure: (a) Postflop.cbet builds flop-only — no turn/river resolution after a specific flop line is reachable through this rail; (b) the builder is opener-as-hero only — BB-as-hero postflop is unavailable. Turn/river and BB-as-hero requests are in the pull script's skip_log (three entries). The runnable surface is "CO c-bet on the flop after BB calls a 2.5bb open," which is the street originally charted for the bluff-shove vehicle hypothesis. Scoring rule for §4a cluster 1: at stack=8bb the solver's largest bet label (bet5.5) is effectively a flop shove given remaining stack depth, but the literal allin action field is used only in basis points. Per-combo shove-vehicle counts therefore use max-size-bet-mass — probability mass on the largest-sized bet label per cell — not the nominal allin field. See deviation-log.md D-7 for full API-surface + scoring documentation.

Reproducibility. Every number in this article is keyed to a cached endpoint response at c:/tmp/sg_cache_articles. Rerunning pull_squid_hero_last.py reproduces the aggregate + per-hand preflop dataset in ~30 seconds (warm cache); rerunning pull_h2_percombo_postflop.py reproduces the per-combo postflop dataset in ~82 seconds (cold — per cell includes a full per-combo action strategy payload). Both pull scripts and the endpoint URL + preflight script are committed alongside this article; the output JSON files are gitignored regeneratable artifacts (regenerate locally with an empty cache).

Companion. Readers who want the unabbreviated mechanism spec and the full [T1] / [AMBIGUOUS] research notes find them in book-2 Part 6: Hero-Last: Desperation Polarization. Cross-book references: book-2 §2.3 (state-dependent VPIP), §5.6 (river polarization), Part 5 (postflop aggression by texture).

Authored for status: drafting. [MODEL-ONLY] Solver grading by QuintAI. Nick Petrangelo (GTO Lab) responds inline per coach_relationship_mode: technical-reviewer. Deploy-first per METHODOLOGY §1b — Nick reacts to the live page. Post-deploy changes land in the v1-changelog block at the top of this section on each redeploy.