Baldur’s Gate 3 as a lens on the cast

Session: Cowork (Claude Opus 4.8), ultracode. Wedge asked why Baldur’s Gate 3 is regarded as the pinnacle of the CRPG, what its storytelling actually does, and whether any of it belongs in Tsumugu-core. Method: 39 agents, seven parallel research lanes on Larian’s design record, per-companion arc architecture, the critical backlash, and the CRPG lineage; 98 claims sent to adversarial refuters; 18 extracted mechanisms transfer-tested against Tsumugu’s constraints; three critics run against the result.

Verdict

BG3 is the pinnacle of game reactivity, not of game writing. At The Game Awards 2023 it won Game of the Year, Best RPG, Best Performance, Player’s Voice, Best Multiplayer, and Best Community Support, and it lost Best Narrative to Alan Wake 2. The one category isolating writing craft went elsewhere. Its companions are good for reasons a novelist can steal; the game is loved for reasons a novelist cannot.

The lens earns its keep as a diagnostic rather than as an import. Of 18 mechanisms extracted and tested, twelve came back already present in Tsumugu canon. Two transfer with translation. The rest die on the tonal charter or on the absence of a player.

The single most valuable finding is not a BG3 technique. It is that our locked rule — every lead carries two axes on separate roots — is honored by two of seven students, and that the bible’s “Second arcs — the relational web” section is where the other five go wrong.

What the companions actually run on

Six mechanisms, stated as causes rather than compliments.

• The wound precedes the goal. Larian builds a companion from their “shadows” (fears, losses, betrayals, scars), then fits a surface want over the top. Every peak is therefore a recognition. Critic Kat Clay on Shadowheart: “she doesn’t actually want this, she’s only been indoctrinated to want this.”
• The ownership spine. Each companion is owned by a patron whose literal claim on their body or soul is their lie made concrete: Astarion/Cazador, Shadowheart/Shar, Lae’zel/Vlaakith, Wyll/Mizora, Gale/Mystra. Refusing the patron is the crux, never a subplot. Karlach (already escaped Zariel) and Gale (loves his owner) fray the pattern, which is what stops it reading as formula.
• The prop does the arguing. Astarion’s sunlight inverts across endings — refuse ascension and he burns on the docks, free and sunless; ascend and he keeps daylight forever by becoming Cazador. Gale’s Netherese orb is a literal bomb of hubris. Lae’zel’s Act 1 refrain “I will ascend” is paid off by inverting its content.
• The dark mirror ships with the good ending. Ascended Astarion sacrifices seven thousand spawn and offers to make the player his own spawn. The evil path is the same trauma logic followed to its end, which is what gives the good path weight.
• Some clocks cannot be beaten. Karlach’s infernal engine cannot survive the Material Plane. Writing director Adam Smith says the heart was “one of the last things that fell into place,” retrofitted late. Both launch endings are losses, and she is among the most-loved characters in the game.
• Softening is not redemption. Adam Smith on Minthara: she “softens” and learns to be “part of a group” while remaining “a pragmatic, evil person” who “has a lot of love for the individuals.” Warmth and moral reform are deliberately decoupled.

Around those sit the systems that produce the “being seen” feeling: an approval score moved in ±1/5/10 increments that never decays; approval that fires only when the companion is physically near the player; warnings at −20 and −40 and permanent departure at −50, gear included; camp and Long Rest as a valve rationing every confession behind a mundane resource action; banter keyed to party composition, so companions perform for one another. Larian’s designers reach for the recognition word themselves: “We see you… the world, and the story is reacting to that.” The mechanism underneath is refusal. A companion who could never leave could never really approve.

The honest case against

• Act 3 is a widely acknowledged quality drop, with the teased Upper City sealed until the finale. Larian conceded cut content.
• Wyll’s arc was rewritten late, restarted only once the other companions were solid. A major setpiece (the Red War College) was cut, a lead writer fell ill during a critical window, and the arc was split between patron (Mizora) and father (Ravengard). Writer Kevin VanOrd called the split “a mistake in hindsight” and apologised publicly: “I’m truly sorry I didn’t give you more quality time with him.” He has the least content of the six.
• The romances shipped too fast because of a bug — approval thresholds set too low. Vincke: “That’s why they were so horny in the beginning.”
• The launch ending was widely panned as abrupt. Larian’s fix was a whole new playable epilogue in Patch 5 (~3,589 new lines), after reworks of Karlach’s ending under sustained forum backlash. The Netherbrain plot is connective tissue carried by the companion writing.
• The famous numbers are provenance-rotten. The 2,503,837-word script is real and Guinness-audited. “17,000 endings” counts permutations, came from lead writer Chrystal Ding, and drew publishing director Michael Douse’s “Oh no they found the 17,000 number.” The “174 hours of cinematics” figure is Larian’s own Steam blog.
• Almost nothing in the toolkit is new. Chris Avellone’s Companion’s Guide to Companions (2017) states the approval rule verbatim: companions are with you for a reason, that reason may change with your actions, and crossing their principles makes them lose faith, leave, or turn on you. Approval-with-abandonment is Dragon Age: Origins (2009). Companion-to-companion banter is Baldur’s Gate II (2000). The playable-origin frame is Larian’s own Divinity: Original Sin 2.

The genuine novelty is fidelity: nearly all dialogue performance-captured in body and face, 248 actors in mocap suits. Neil Newbon won Best Performance. Astarion lands because the read carries subtext the lines withhold.

The mirror — what we already built

Twelve of eighteen mechanisms map onto existing canon. Camp and Long Rest are the 珍貓 café plus the rooftop, and the R1-anchor / R2–R5-texture split. The no-decay approval ledger is the meter-never-resets rule in Stakes Without Mortality. Composition-keyed banter is the ≤3–4-speaker cap plus the named pair-arcs. Proximity gating is that same cap. The shared-stimulus contrast engine is B1L09’s “where’s home for the break?”. The dark mirror is 金多恩’s corruption-averted turn. The unbeatable clock is 星野’s visa and 阮草’s deliberately unresolved crisis. Campfire-then-battle is B3L03’s typhoon paying off at B4L06’s earthquake.

We arrived at the architecture independently. The corollary is that “adopt BG3” is mostly a no-op, and any recommendation that reads as a clean import deserves suspicion.

The two things that genuinely transfer

1. The ownership spine, renegotiated rather than defeated

BG3’s version needs a villain to destroy, and our charter forbids one. The half that survives is the externalisation: turning an internal lie into a concrete external claimant with a physical demand. A 250-character vocabulary cannot say she believes she must carry everyone alone. It can show an envelope, a pair of cracked hands, a length of cloth. Externalisation is show-don’t-tell forced by the ceiling.

Every character already arrives owned by something that decided their life for them. 星野 by the Osaka OL track, 阿迪 by the workshop, 阮草 by the family debt, 林薇 by the family that tethers her, 蘇老師 by the dream shelved when her husband died, 吳老闆 by his late wife’s dream, 白龍 by the family the internet broke, 金多恩 by the grieving family who want the safe path, 小圖 by his programming. And 沈文 by nothing, which is exactly his wound.

We have never named this, and naming it dissolves the coherence risk CAST-DIAGNOSIS.md §1 flags. The two spines stop competing under one sentence: everyone arrives owned; presence is what lets you turn and face the owner; agency is what remains after someone stays. Note that BG3’s frayed cases (Gale, who loves his owner; Karlach, who already escaped hers) are our register, and its clean cases are the ones the charter forbids.

Worked move, which also repairs a defect the diagnosis already named. B2L04, 阿迪’s canon workshop-weight plant: a package arrives from Jakarta, a folded length of his father’s workshop batik. Not a gift — his father sends an instruction to make something with it, for the shop. 阿迪 puts it in a drawer and keeps sewing his own designs. Received, stored, unused; zero feeling-words; A2-legal. At B4L07, the recovery beat the diagnosis calls cost-free, he finally cuts his father’s cloth into a streetwear piece. The fusion costs him the cloth. The claim is answered by renegotiation, never by victory.

Price: run it on all ten and it reads as formula. BG3 protects itself by fraying two of seven. 沈文 and 白龍 should be deliberately off-pattern.

2. The performance channel

The vocabulary ceiling blocks feeling-words. It does not block feeling-prosody. BG3’s flat lines land because the read carries what the script withholds — which is why Newbon won Best Performance while the game lost Best Narrative.

We have designed, frozen, per-character voices, and a chorusing/shadowing mechanic in which the learner voices the line in their own mouth. We have also already written a beat that only exists in audio without noticing where it lives: 星野’s keigo party trick, the syrupy customer-service phone-voice she holds a beat too long and lands too perfectly. On the page that needs a narrator. In her frozen voice it needs nothing. The same holds for her Kansai register dropping to clipped standard the moment the mask slips. A 250-character line with no emotional vocabulary can carry devastation entirely in the read, at a cost of zero Han characters.

Two consequences. First, several transfer tests killed mechanisms on the premise that the reader reads once and chooses nothing. That premise is false here. A graded-reader audience is the most re-reading audience in prose (SRS, chorusing, decode, look-up, review), so the B1L02 photo that re-reads differently after B3L10 is the modal experience rather than a bonus for the diligent. A learner failing a decode and looking it up is a literal fail-forward the reader owns. Second, BG3’s most radical craft choice is a diegetic omniscient narrator who delivers interiority and irony, and she is beloved — which sits directly against our hardest self-imposed rule. Keep the rule on the page. Whether it binds the audio layer is an open question we have never asked.

The two-axis finding

Running the XINGYE-DEEPENING.md strip test (remove axis 1; does the axis-2 flaw survive on its own root?) against BG3 first, then against our seven.

BG3’s companions do not uniformly carry two axes. Only Astarion and Gale do. Shadowheart, Lae’zel, Karlach, and Halsin each run one deep axis with rich texture. The most-loved cluster is Astarion, Shadowheart, Karlach — two of the three single-axis. Axis count does not predict love; Karlach carries the game on one axis. Wyll has the raw material for a second axis (father, heroic identity) with no root of its own, and is the least-loved of the six. A bolted-on second axis is worse than none. And where a second axis does pay in BG3, it pays as a corruption fork (Ascended Astarion, God-Gale), which is exactly the tool our tonal rules confiscate at the door.

Our seven:

Character	Strip test	Note
星野遙香	passes	Strip the karoshi dread and the beloved taker who has cooked for no one survives. The only axis in the cast adversarially vetted, and it shows.
金多恩	passes	Strip the grief and the AI-worth crisis stands; strip the AI question and the grief stands. Tightly braided, separable roots.
沈文	soft fail	Heritage-identity and family-hunger are genuinely separable. As authored, heritage is a reveal plus a sub-quest that resolves into the found-family payoff, so it never runs its own flaw-to-turn movement. Cheapest to promote.
阮草	fail	Provision-vs-presence is her provision root staged against 沈文.
阿迪	fail	The flakiness is caused by the workshop weight; STORY-ARCS.md says so outright.
白龍	fail	The 金多恩 romance is his loneliness root staged as a two-hander. The B5 family-rupture reveal is the Ghost explaining axis 1, not a second root.
林薇	fail	The wanderlust is a want deferred by circumstance, never a flaw she overcomes.

The failure is systematic and it has an address. STORY-ARCS.md §“Second arcs — the relational web” defines each character’s second arc as a relationship. A relational arc on the same root is a two-hander: it builds a bond, and it is not a second axis. Relabelling it as one is precisely the Wyll move, the one that lowers affection.

Two of seven honor a rule locked as universal. BG3 says that is survivable — 阮草, 白龍, and 林薇 are Karlach-class single-axis characters with strong texture. If the rule is to be honored, the only model that works in a warm world is the one we already invented: 星野’s scissors, a positive second-axis growth that sharpens the first axis’s loss instead of resolving it. A scissors with no villain.

What we would not import

BG3’s emotional economy is bought with cruelty (the Shadow Curse, Cazador’s spawn pits, Gale’s bomb, the standing possibility of losing a companion for good). Its pacing assumes a hundred hours of buy-in and detonates in Act 3; we lose most readers by Book 3. Its interiority arrives through mocap faces we do not have. Reaching for BG3 to cure arc monotony reaches, every time, for the corruption fork the charter forbids.

Warmer lighthouses, each offering what BG3 cannot. Frog and Toad — loneliness, jealousy, and the fear of being unloved inside a first-reader vocabulary; the proof of concept. Iyashikei (Aria, Yokohama Kaidashi Kikō, Barakamon, Kiki’s Delivery Service) — a mature aesthetic on no villain, no death, conflict resolved through growth; Kiki delivers two axes on separate roots. The short-story cycle (Olive Kitteridge, Winesburg, Ohio, Cranford) — the R1-anchor-plus-texture structure solved by masters. The Wild Robot — the 小圖 thread entire. Austen’s Emma — two independent axes in a warm world where nobody dies.

BG3 out-teaches all of them on exactly two things: planting rigor (the approval system is a vast disciplined engine for placing a small beat early and detonating it eighty hours later) and dual-axis characterisation. Those are the two our own diagnosis names as gaps. Take those; leave the knife, the ruin, and the face.

Corrections owed to the wing

• Arc Types and Stakes Without Mortality both still place 星野’s mask-crack at B3L08. The 2026-07-09 decision moved it to B3L09, triggered by the group-split scatter.
• Companion Arcs and Party Banter builds its central section on the Mass Effect 2 loyalty mission and files BG3 alongside it. BG3 does not use that structure; it braids each companion’s arc through all three acts rather than isolating it into a modular quest. The “one arc-closing beat per companion” rule is ME2’s, and the attribution should say so.
• The same page credits an “actions stick” framing to a Larian developer named Stephen Rooney. The reputation-system explanation we could verify traces to writer Kevin VanOrd. The Rooney attribution did not survive checking and needs a source or a cut.

Refuted during fact-check

Five of 98 claims died, all misattributions worth recording so they do not creep back in: the Karlach infernal-heart quote belongs to Adam Smith, not Chrystal Ding; Orpheus is imprisoned in the Astral Prism, not dead, at Lae’zel’s reveal; the friendship/rivalry dual track is Dragon Age 2, not Origins; the “human weakness” design language is genuinely Vincke’s but the commonly cited Game Developer URL does not contain it; and there is no game-studies consensus that player agency and authorial control trade off inversely — that framing (the narrative paradox) is contested, not settled.

Open questions

1. Does the two-axis rule stand as universal, or does it become a rule for the leads who can carry it (星野, 金多恩, 沈文) while 阮草 / 白龍 / 林薇 / 阿迪 are formally re-typed as Karlach-class single-axis?
2. Does the ownership spine enter canon as the series’ unifying frame, and if so which two characters are deliberately frayed off-pattern?
3. Does the audio layer inherit the no-narrator rule, or is prosody an interiority channel exempt from it?
4. Does STORY-ARCS.md §“Second arcs — the relational web” get renamed to something honest (pair-arcs, bonds) so a relational two-hander stops being counted as a second axis?

Next move

Nothing written to canon. The findings above are analysis and unbound proposal, per the standing rule that brainstorming does not enter a plan without Wedge’s call. The cheapest first commitments, if any: the two stale B3L08 references in the wing, and the ME2/BG3 attribution fix — corrections rather than decisions.