Pandemonium

Story mode. Written for a first playthrough and kept free of spoilers — it will not tell you what happens here, only what the game is asking of you.

Prepare first. The fights here come one after another with no shop and no chance to change equipment in between, and Bran Bal is behind you.

Bring wind and lightning resistance, and Locomotion if you have it.

Five things to pick up

• Holy Miter — examine the throne in the first room
• 20,007 gil — set the elevator to 3 and ride to the top
• Carabini Mail — the blue teleport pad room; it teaches Auto-Regen
• Elixir — the second blue pad
• Battle Boots — take the south path at the fork

The blue orbs are on a thirty-second timer. Work out where you are going before you step on one.

The fights

Amdusias — Zidane starts alone. Defend until Freya and Amarant turn up, then fight normally.

Abadon — Quina and Steiner, with Zidane joining. White Wind and Mighty Guard do the work.

Shell Dragon — cast Float. It uses Earth Shake and Float turns that off. Dagger arrives partway in.

Silver Dragon — optional, and unusually willing to take Sleep, Slow, Silence and Blind for something its size. Steal the Dragon Mail and the Kaiser Knuckles before you finish it.

Garland — Silence and Blind both land. He casts Stop, which Locomotion blocks. Three things to steal: Battle Boots, Ninja Gear, Dark Gear.

Kuja — almost all magic, so Shell on everybody. Do not cast Reflect — he replies with Flare Star. Steal the Ether, the Carabini Mail and the Light Robe.

Quina eats well here

• Malboro → Bad Breath
• Abadon → Twister
• Shell Dragon → Earth Shake

Three in one dungeon, which does not happen again.

If you want to grind

This is a better place for it than Terra was. Abadon and Shell Dragon give 32,073 each and Malboro 30,579, against 23,801 for a Mover formation — and Bran Bal’s Vaccines are behind you now, so you can buy protection from Virus instead of stealing it.

Movers are here too, two formations of them, behaving exactly as they did in the trees.

Don’t miss Mozme

The moogle in the maze takes a letter and gives you a Kupo Nut.

Game Notes

Pandemonium is the first dungeon in this game that is honestly difficult in a structural way rather than a numerical one.

Everything up to here has let you leave. Ipsen’s Castle had a way out. Oeilvert was one-way but short. Even the Desert Palace, which locks a lot of doors, gives you a moogle and a chance to think.

Here the game takes away the one thing an RPG player reflexively reaches for when a fight goes badly, which is the menu. No shop, no swap, no reconsidering your loadout between rounds. Whatever you walked in with is what you have.

That is a much older kind of design than most of Final Fantasy IX, and it is doing something deliberate: it is the point in the story where the party stops being able to regroup, and the mechanics agree with the plot for once instead of politely ignoring it.

The practical consequence is boring and worth repeating anyway. Spend the gil at Bran Bal. You will not get another chance, and the game will not remind you.

Completionist mode. Full spoilers, in play order, with every permanent loss called out before you reach it. 7 entries below are marked where our sources disagree with each other — we have not picked a winner for you.

Four bosses, two numbers each, ten thousand apart

We have been carrying a question since the second week of this project: a handful of boss HP records where our field data sits exactly 10,000 above what a publisher prints, with no pattern to say which records do it.

Pandemonium hands us four at once — the largest cluster by a distance — and with them the two arguments that make the question answerable.

Every gap is 10,000 on the nose. And look at which way the guides lean: Caves of Narshe matches our field on three and undercuts us on the fourth; Jegged undercuts us on three and matches on the fourth. The two swap sides within a single dungeon.

That is the crucial shape. The camps are not about the publisher. They are about the record.

The control pair that closes the door

Amdusias and Shell Dragon are, in our data, two thirds of an authored set:

Same level, same experience, five hit points between two of them. Whoever built this dungeon built these three as one job.

And the publishers split differently on the two we can compare. Jegged matches our field on Amdusias and undercuts it on Shell Dragon — two records that are, by every other measure, the same record twice.

So the disagreement cannot be about the enemy’s level. It cannot be about its tier, or the dungeon, or whether it is a story boss. It cannot be about magnitude — we already had six clean controls from 8,106 to 29,186, and now a clean and a carrier five hit points apart.

Whatever this is, it operates on individual published records, one at a time.

The new argument: the neighbouring fields are fine

Here is the thing we had not thought to check, and it is the strongest evidence we have produced on this question.

On the very records where the HP is disputed, our steal tables match the publishers exactly.

Three for three, three for three, and three for three — item names, in order, from the same records whose HP field we and a publisher cannot agree on.

If our reader were misaligned on these records, the item names would be garbage. A byte offset that lands 10,000 wrong on an integer does not then land perfectly on nine consecutive item ids. The drops match too — Silver Dragon gives a Wing Edge and three Elixirs in our table, and the Fenrir card, all of which the corpus confirms.

So the discrepancy is isolated to one field on some records, on data we are otherwise reading correctly. That rules out the explanation we have been most worried about since the beginning, which is that we were simply reading the table wrong.

Which side is right — this part is settled

The lower figure is the one you fight. We should have said so plainly on this page the first time and did not, so here it is.

It is settled by arithmetic, on a case far away from this dungeon. Plant Brain stores 10,916 in the same field and is published at 916. You fight it at around level 7 with a party doing roughly 40 to 60 damage a hit. 916 is about twenty turns. 10,916 is about two hundred. Everybody beats Plant Brain in a few minutes, so 916 is the real number and the stored field is not what the engine fights with.

Roughly twenty-five records carry the offset, from Baku early on to the final fights. Our policy across this whole guide is to render the published figure, not our raw field — and where publishers disagree, to show both. That is why the checklist on this page leads with the lower number.

What is genuinely open is the mechanism, and we have narrowed it a long way:

• The lower value appears nowhere in the affected records. For each one we scanned the entire scene blob, not just the monster record, as both 16- and 32-bit values. It is not stored anywhere. Publishers are not reading this field for these enemies — the low numbers come from play, not from a file.
• It is not a second HP field. No offset in any affected record holds stored − 10,000.
• It is not in the game’s code. The constant 10000 appears in nine method bodies in the assembly and none of them is battle logic.
• The affected set is not separable by value, level, disc, tier or story order. Ark at 20,002 is clean; Valia Pira at 22,119 is offset. Lich, Tiamat, Kraken and Maliris are clean while Nova Dragon, Necron and Trance Kuja are offset — and those two groups sit in the same scene family.

One piece of apparent support that is not support

Caves of Narshe matches our raw field on Silver Dragon, Garland and Kuja, and on half a dozen records elsewhere. That is not corroboration and we will not present it as such.

The likeliest reading is that Caves of Narshe extracted the same table we did, by the same kind of process. Two readers of one file agreeing tells you the file was read correctly. It tells you nothing about what the engine does with the value afterwards, which is the entire question.

It inverts how to weight them: on enemy statistics, Caves of Narshe agreeing with us is a signal that a figure came out of the files rather than out of play — which, for this field, is a reason for more caution rather than less.

The playtest just got much cheaper

Our standing plan was to settle this by playing to Armodullahan, an optional fight in a dungeon most players skip.

Garland and Kuja are mandatory. Everybody who finishes Final Fantasy IX fights both, and the disputed gap is 25% of Garland’s total health — twenty per cent of Kuja’s. That is not a subtle difference; it is the difference between a fight that ends and a fight that keeps going for another four rounds.

Garland and Kuja move to the front of the playtest queue, and Armodullahan drops to third.

And the Abadon we cannot reconcile

Both publishers give the Abadon boss as 12,658 HP at level 58. That is not a ±10,000 case — it is our random-encounter Abadon’s record, exactly. Our scripted-fight Abadon is level 54 at 20,926, sitting in the matched triple above.

One of three things is true: the game really does use the level-58 record for the scripted fight and our set-of-three reading is a coincidence; both publishers pulled the bestiary entry rather than the fight; or we have the wrong record. We cannot tell from here and we are not going to pick. It goes on the playtest sheet with the others.

Chests: five for five, three publishers, unanimous

Three publishers list exactly these five and our field data holds exactly these five. No disputes, no phantom sixth.

The 20,007 is worth a note of its own. Most gil awards in this game are computed at runtime — our reader marks them dynamic and we never publish a figure for them. This one is hard-coded in the field script, and all three publishers print it to the digit. It is the largest single gil chest we have read anywhere in the game, and it is a clean, boring validation of the reader on exactly the kind of field that the boss-HP question makes us nervous about.

Two of the pickups are duplicated by the bosses, which is a small joke: Garland carries a second pair of Battle Boots and Kuja carries a second Carabini Mail, both stealable, in the same dungeon that gives you the first of each in a chest.

Quina’s best dungeon in the back half

Our enemy table stores a blue-magic index on each record. We decoded the mapping early and it has now been confirmed three times against publishers — Roulette from Hecteyes, Magic Hammer from Ring Leader, and now Bad Breath from Malboro, which Caves of Narshe states outright.

Which means these two can be trusted as well:

Three blue magics from random encounters in one dungeon. No publisher we consulted lists all three on the Pandemonium page; the Malboro is mentioned and the other two are not. This is the coverage pattern we keep running into — the information exists in bestiaries, and the page where you would act on it does not carry it.

The grinding actually is better here

Pandemonium’s randoms beat Terra’s by a third:

And the practical difference matters more than the numbers. Bran Bal is behind you now, which means Vaccines cost 100 gil instead of a steal attempt. In Terra, Virus could silently zero an hour of grinding and the only defence was luck with the steal menu. Here you can just buy the answer.

Movers are here in two formations, unchanged — same three, same single payer, same Delta Attack. Fandom’s read of the AI is that the move requires all three healthy and standing, so damaging one of the outer pair is enough to switch it off; you do not need the kill.

Mozme, Moorock, and the marker

Two moogles: Mozme in the Maze, who takes a letter and hands over a Kupo Nut, and a second in the Hall.

Both rooms carry the marker we decoded on the previous page — the unnamed key item that turns out to sit at all 45 Mognet moogle rooms in the game and nowhere else. Pandemonium was one of the five two-moogle locations that made the case, and the two rooms our data marks are the two rooms the corpus names.

An unresolved oddity in the Kuja scene

Our data holds two records in the Kuja battle scene: Kuja at level 64, and a Trance Kuja at level 64 with 42,377 HP.

No publisher describes a trance phase in this fight. We do not know whether the record is used here, whether it is staged for later, or whether it is unreferenced data sitting in the scene. We are not going to tell you there is a second phase on the strength of a row in a table. It goes on the list.

Game Notes

The steal tables are the thing we should have checked eighteen areas ago.

We have been treating the boss-HP disagreement as a question about a number: which number is right, what distinguishes the records that disagree, is there a flag somewhere that marks them. We tried level, magnitude, disc, boss-hood and a flags field, and falsified every one of them, one at a time, over five weeks.

What we never did was look at the fields next to the number. And the moment we did — on three bosses, in one dungeon, all disputed — the answer was immediate and it was not about the number at all. Nine item names in a row, in order, correct, on records whose HP nobody can agree on. That tells you more about where the problem lives than any amount of staring at the integers did.

The lesson generalises badly, which is why it is worth writing down. When a single field is in dispute, the cheapest test is not another hypothesis about that field. It is to check whether everything around it is fine. If the neighbours are healthy, the problem is not your alignment, and half the hypotheses you were about to try were never worth trying.

Five weeks, and it took a dungeon that disputed four records at once to make the question small enough to see.