glytch/fktry/data/CANON-DELTA.md

CANON-DELTA — everything in data/ that isn't in the codex yet

LANE-DATA owns this file. ../../docs/FKTRY_LORE.md is canon; this is the ledger of where the data has run ahead of it. The orchestrator absorbs these into the lore doc in one pass later. Until then, this is the single place canon-drift is allowed to live.

Rule for adding an entry: the codex was silent, I invented in its voice, and I wrote down why it is consistent with the codex — not just what I did. If a future lore pass disagrees, the codex wins and the data changes.

---

Canonized in round 1 (orchestrator: "CANONIZED", round 2 review)

Thing	Kind	Why it fits
press-anchor-slab	recipe (dct-press)	The press run backwards is an IDCT, and an IDCT is a complete picture. §4 already runs the field loom and the telecine puller in reverse, so reversibility is established machine behaviour. This is where I-frames come from.
color-plate	item	§4 demosaicer literally outputs "a smooth colour sheet"; it needed an id.
v-hold-roll	item	§1 Broadcast Shelf has "v-hold faults that make whole regions periodically roll upward". Bottling one gives the ANALOG PACK's "combing + v-hold roll + static" triptych three real ingredients.
artifact-bottler	machine	§2 calls the science packs "a bottled artifact triptych" but names no bottling plant. Something has to hold the bottle.
lane-splitter	machine	The contract has kind: 'splitter' and belts need routing, but §4 only ever calls the DEMUXER "the first splitter" — which is a refinery, not a junction.

Added in round 2

The container line (unblocks stream-hex-splicing)

§2 names the CONTAINER PACK's contents ("moov shard + offset ribbon + sync key") and §7 names the moov brain's parts, but nothing said where they come from. The chain now mirrors what the parts actually are in a real container:

Thing	Kind	Why it fits
moov-shard	item	The index, not the payload.
offset-ribbon	item	The stco map — a ribbon of byte offsets.
sync-key	item	The stss table lists which frames are keyframes, i.e. I-frames.
demux-container	recipe (demuxer)	The demuxer is the machine that separates index from payload. 8 ore → 2 shards: the index is a rounding error next to the mdat, which is the joke and also the truth.
spool-offset-ribbon	recipe (gop-assembler)	The assembler is the only machine that knows where each frame starts, because it put them there.
cut-sync-key	recipe (artifact-bottler)	Cut from an anchor slab, because stss is a list of anchors.
bottle-container-pack	recipe	Assembles the §2 triptych.

Chicken-and-egg check: §4 says the hex splicer "Requires CONTAINER PACK research", so the container line deliberately does not route through the splicer. It is demuxer → assembler → bottler, all machines you already own.

The bloom line

§3 sells bloom "by intensity grade; grade-5 bloom visibly tries to escape its canister in one direction" — five grades, one item each, bloom-grade-1..5.

• bloom-concentrate is kept and redefined as the raw, ungraded draw-off from a loop (tap-bloom). It was not deleted because LANE-SCREEN hardcodes it in src/screen/corruptionMap.ts:54 and its test — deleting it would have broken another lane's suite mid-round. Grading now consumes it.
• The recipe id bloom-grade-3 survives from round 1 but now produces the item bloom-grade-3, so the id finally means what it says. tap-bloom is the renamed round-1 bloom-grade-3 recipe. Deprecation notice: no recipe id was retired, but bloom-grade-3 changed its inputs/outputs entirely. Nothing outside data/ referenced it (checked: only my own tech.json).
• Escalation is in delta wafers, not exotic inputs — a grade-5 canister is 2 raw bloom + 18 wafers of compounding, which is §4's "contents compound each lap" priced honestly.

Added in round 3

THE MUXER and the moov brain (§7, data-first — mechanics are M4's)

§7 names the parts and says each is "its own late-game production line", but never said what makes them. Each part is now made by the machine that already owns that knowledge, which is the whole joke: the container is assembled by the factory that made the mess.

Thing	Made by	Why that machine
stsz-ledger	gop-assembler	The sample-size table. The assembler packed the crates, so it knows how big each one was.
stco-map	hex-splicer	The chunk-offset map — byte offsets are the splicer's entire domain. It nudges them for a living; here it writes them down honestly, under protest.
stts-metronome	telecine-puller	Time-to-sample is a cadence, and the puller is the cadence machine. Built from judder cams, so the metronome ticks 3:2.
stss-keyring	artifact-bottler	The sync-sample table is a list of I-frames, so it is literally forged from anchor slabs.
ftyp-seal	artifact-bottler	§7 names the ftyp seal as part of the container; it needed a source.
moov-brain	muxer	The four converge. 300 ticks.
megaglitch	muxer	mux-container: brain + seal + 500 mdat ore + 100 melt, 900 ticks. §7's "colossal composited corruption" priced like one.

muxer is kind: 'crafter', not 'shipper' — it builds the megaglitch, and you still have to ship the thing. 5×5, the biggest footprint in the game, per "hero structure".

ASIC COOLER (spec'd, deliberately not yet mechanical)

§2 HEAT says heat is "Vented by ASIC coolers" — a machine the codex names and never specifies. Authored now so the data is ready; the mechanic is deferred to M4, when SIM builds auras. Gated behind broadcast-asic-cooling so it can't appear early.

• Intended mechanic (SIM, when auras exist): an adjacency aura — every machine whose footprint touches the cooler's 2×2 gets the cooler's coolPerTick (0.02) instead of its own. That turns the software decoder from a 30%-duty throb into something you can engineer around: ring your decoders with coolers and pay for it in floor space and build cost. This needs a contract change (coolRadius?, or a kind: 'cooler') — not requesting it until SIM is ready to build it, since a contract field nobody reads is the same decorative trap heat just fell into.
• Authored as kind: 'power' with powerGen: 5 because there is no cooler kind and a lie is worse than a stretch: it is an ASIC, ASICs decode, decoding makes bandwidth. It vents far more than it earns. If kind: 'cooler' ever lands, this moves.
• Its own coolPerTick: 0.02 is inert today (nothing heats it) and is there to state the intended aura strength in data rather than in prose.

Added in round 4

§10 THE COMPOSITE SEAM — transcribed

The codex grew §10, so most of this round is transcription rather than invention. Every item and machine in §10 is now data. Deviations and judgement calls only:

Call	Why
"Footage" = anchor-slab	§10's keyers key footage, but the codex never says what footage is as an item. The anchor slab is already "a complete, self-sufficient picture" (§2) — the only thing in the game you could point a keyer at. Every keyer takes slabs.
MATTE BATH's chemical = static-canister	§3 already says static canisters are "brine for preserving other artifacts". The bath needed a developer; the brine was sitting right there.
ghost-feather is made only by the neural keyer	§10: "every era below the neural keyer ruins it, so demand never dies." So the earlier machines consume it and ruin it — bathe-feather (feather → hard matte, "the machine does not apologize") and key-hair (feather → roto chatter). Both are jokes with a real cost, and both are the codex verbatim.
NEURAL KEYER powerDraw: 240	§10 says "24 VRAM — a bandwidth appetite that dwarfs everything else in the game", and "you don't unlock the neural keyer; you unlock the power plant it demands". 240/s is 12 decode ASICs, or ~9 software decoders after duty cycling. The whole M1 reference chain draws 96. It dwarfs the game, as ordered.
sodium-light has no recipe	§10: The Correction hoards the prisms. Forward canon; on the FORWARD_CANON allowlist. sodium-requisition commissions it anyway, with 30 BLUE HOLES accepted "if you lose your nerve".
ROTO BENCH is gated (disc), not free	§10 calls it "the economic floor under every keyer... the bench never does [fail]". Tempting to leave ungated, but ALPHA is "the rarest resource in the game" and an ungated bench hands it to you at tick 0. Gated cheaply (10 spatial) so the floor arrives early but the Composite Seam stays a mid-game fault.

Optical printer ratios hurt, as ordered: 4 anchor slabs + 2 colour plates → 1 alpha and 4 fringe lace, over 240 ticks. It is the worst alpha-per-slab in the game and it fringes every single run, which is exactly what the iron lung is for.

THE ARCHAEOLOGY LAB (kind: 'lab') — canonized

The codex has no lab: §6 only says "TECH TREE = MEDIA ARCHAEOLOGY. You research downward (excavating older strata for analog secrets) and upward (deriving newer codecs)." So the lab is the dig site — named for the verb the codex already uses, not a generic "research station". Ungated starting equipment, because a lab behind research is a lab you can never build (there is now a test named exactly that).

Added in round 5

THE COMMISSION LADDER — the fax is the tutorial

There is no tutorial in FKTRY and there should never be one. The fax machine is already the game's voice, already tells you what to build, and is already sacred (§6: "do not ask why it works"). So the first four commissions teach one mechanic each and the player never knows they were taught.

Rung	Wants	Teaches	Machine types needed
1 raw-payload	20 MDAT-ORE, raw	extractor → belt → uplink. The whole verb set.	2 crafting (+belt, +uplink)
2 first-refinement	15 LUMA BARS	the demuxer — and byproducts: the slurry has to go somewhere and the fax does not care	3
3 dust-quota	18 HF DUST	the quantizer, and that waste is inventory	4
4 first-taste	1 MELT	graduation: the whole M1 chain, ore to reactor	8

Rung 1's flavour is the instruction manual: "A belt, a drill, an uplink to THE SCREEN — that is the whole job." Rung 2 hands you a byproduct and refuses to help. Rung 3 reframes the garbage as product, which is the game's actual thesis. Rung 4 is the old opening commission, reward bumped 50 → 80, now earned rather than assumed.

Measured, not estimated (throwaway probe against LANE-SIM's real sim, then deleted):

rung 1, built exactly as the fax describes — drill, 2 belts, uplink, one ASIC:
  4 machine types.  Cleared at tick 650 = 21.7 seconds.

The order was "completable inside 3 minutes with 4 machine types". It is completable in 22 seconds, which matters more than the budget suggests: the first fax must clear before a new player decides nothing is happening.

Why this was needed at all: the queue used to open on first-taste — one MELT — which requires ore → demuxer → quantizer → subsampler → p-caster → DCT press → GOP assembler → mosh reactor. The game's first instruction was "build eight machines correctly, in order, with no feedback until the last one works." That is the onboarding cliff.

THE OPTICAL FOSSIL — why the world is silent until you dig

§1 buries the hook in one clause: the Reel Beds hold "optical soundtrack fossils — wavy photographic strips that sing when excavated." Optical sound is a picture of a waveform printed on film — the one place in this world where image and audio are the same substance. So:

The Bitstream has no audio because the audio track was severed from the container long ago, and one length of it is still in the ground. Everything before the fossil is found is a video file with no sound: UI clicks and nothing else. Excavating it splices the audio bus back into the file, and the whole band comes alive at once — the tuner, the stations, the factory's own hum. The player doesn't unlock a feature; they repair the file.

This is why THE STANDARD's script includes "Do not excavate. There is nothing in the strata but strata" and "Anything that sings when you dig it up was never meant to be heard." The Correction runs a propaganda station warning you off the one artifact that would let you hear the propaganda station for what it is. It is also why DEAD AIR's ghost line is the only direct acknowledgement in the game that you found it.

The hint is one fax margin, mid-ladder, on haunted-doorway — "[someone has drawn a sine wave in the corner of this fax. it is not our department.]" It points at nothing. A sine wave is a sound and a drawing at the same time, which is exactly what the fossil is.

THE MOSQUITO TRAP — the last forward-canon commission comes off the blocklist

§5 gives the trap in one line: "Catchable in edge-baited traps", the asset note calling it "a bug-zapper shaped like a high-contrast edge target". Mosquitoes are quantization noise made alive: they hatch from HF DUST and cluster on sharp edges. So the recipe is exactly that sentence — 8 HF DUST + 1 LUMA BAR → 1 MOSQUITO SWARM (canned). The dust is where they come from, the fresh luma bar is the sharp edge that baits them, and the whole thing gives HF dust a third sink.

This unblocks vaporwave-sunset, which has wanted "3 MOSQUITO SWARMS (live)" since round 1 and been unfillable the entire time. mosquito-swarm leaves FORWARD_CANON; only gibbs-wraith, cinema-pack and sodium-light remain, and all three are deliberate (a wraith must come willingly, cinema needs the Fortress, and The Correction hoards the prisms).

data/seams.json — the ground

Schema is { relicReserve, seams: [...] }; each seam is { id, era, resource, rect{x,y,w,h}, richness, hidden? } on the -32..31 grid. Five ordinary seams, one per stratum, placed so each era's codex-correct payload has a home: stream crust → mdat ore, disc mantle → bayer grit (§2: "from the deepest camera-original seams"), broadcast shelf → field splinters, reel beds → silver frames.

Two constraints drove the layout, both now tests:

• The reference factory must stand on ore. SIM's three 2×2 extractors sit at (-30,0), (-30,-3) and (-30,3); stream-crust-west spans x[-31,-28] y[-5,6] and covers all three with a tile to spare. Off-seam they'd idle with 'no seam' and the demo would produce nothing at all.
• relicReserve is a seam-free corner — x[12,31] y[12,31]. SIM plants the optical fossil from the seed, 18–28 tiles out, never on a seam. If seams creep across the whole ring there may be nowhere legal left to put it, so the SE quadrant is kept deliberately empty. A test fails the moment a seam intrudes.

richness is authored but currently unread — SIM's rule is boolean overlap this round. It is there for when extraction rate goes data-driven, and it is what the numbers station broadcasts as the last digit.

data/stations.json — the broadcast schedule

Four stations, matching SCREEN's spec: 88.1 FKTRY FM (generative, seeded, era-flavoured), AM 640 THE STANDARD (30 chyron lines), SW 6955 [UNLISTED] (numbers), 108.0 DEAD AIR.

THE STANDARD is The Correction with a transmitter, and the voice rule is that it never threatens — it reassures, in a way that gets worse the longer you think about it. "Your file is being read by someone who loves it very much." "Loss is not loss. Loss is compression that has finished negotiating." "Good morning. It is always going to be morning." 30 unique lines at the ~20s a chyron needs to be readable = 10 minutes before a repeat, which is the DoD.

The numbers station is a real treasure map. Each 5-figure group encodes one coordinate:

[x + 32 as 2 digits][y + 32 as 2 digits][richness × 9 as 1 digit]

  53079  ->  x = 53-32 = 21,  y = 07-32 = -25,  richness 9/9   hidden-nitrate-vein
  07579  ->  x = 07-32 = -25, y = 57-32 = 25,   richness 9/9   hidden-rainbow-pit
  59399  ->  x = 59-32 = 27,  y = 39-32 = 7,    richness 9/9   hidden-payload-lode

Three of the groups on air are the three hidden: true seams; the rest are decoys that decode to empty ground. A player with a pencil gets three maximum-richness seams nobody told them about. The validator recomputes the planted groups from seams.json, so if a hidden seam ever moves and the broadcast doesn't, the test fails — the map cannot go stale. It also checks no decoy accidentally decodes onto a real seam, because a decoy that points at treasure isn't a decoy, it's a leak.

Filler is procedural (SCREEN's seeded RNG) rather than a fixed list, so the station never loops — a hand-written pool of 24 would cycle in about two minutes and the genre's whole menace is that it goes on forever.

Added in round 6

THE CORRECTION walks — the parity-mite reading of §5

§5 gives the enemy in one clause: parity mites are "beetle-drones that patrol belts and quietly fix artifacts back into clean product (your melt becomes... footage of a lake)." The reading the data commits to, and why it is the codex's:

• The mite does not destroy — it repairs, which is worse. The Correction "cannot perceive you, only deviation, which it repairs." So a mite eats a tier-2 glitch product and hands back nothing you can sell: the work is corrected, not stolen. SIM emits a new repaired event rather than a removed one, and the fiction is that losing melt to a mite is losing it to a fix. THE STANDARD's new lines say this in the enemy's own voice ("A mite does not remove your work. It completes it. Your MELT was always going to be a lake." — the codex line, spoken as reassurance).
• Mites only touch corruption. SIM's mechanic ignores tier-0/1 cargo entirely: the immune system has nothing to correct in raw ore. That is the balance (raw material is safe; only product is at risk) and the fiction (deviation is the only thing it can see).

THE FIREWALL — the counter, on SIM's catcher shape

Thing	Kind	Why it fits
parity-mite	item (tier 2)	The canned live mite. Hospital-white shell, green checkmark eye (§5, §8) → chip colour #dff5e6. Tier 2 like the other farmed hazards (mosquito-swarm, gibbs-wraith).
firewall	machine (crafter)	§5 names the mites but no counter. A firewall is the in-genre name for the thing that stops an intrusion, and the joke writes itself: it wears a compliance seal the Bitstream never issued, and a mite "cannot leave a deviation uncorrected" — a forged checkmark IS a deviation, so it climbs in to fix the seal and the drawer closes. Accent #3fff7f is that forged checkmark; chassis #31383f is your grime, not The Correction's white.
catch-mite	recipe (firewall)	Input-free, output parity-mite: 1 — the exact shape SIM detects a catcher by (empty inputs, cans the creature), lifted from the mosquito trap. The "input" is a live mite in the world, which is wildlife in the snapshot, never cargo on a belt. Ingredients would silently demote it to an ordinary crafter that mints The Correction from nothing — pinned by a test that cites SIM's comment.
broadcast-parity-firewall	tech	Gated like the mosquito trap (broadcast-mosquito-farming): {spatial-pack: 15, analog-pack: 10}, unlocks [firewall, catch-mite]. You research the counter to the enemy. Early/cheap because the mite is §5's "early nuisance" — the counter should arrive about when the mites do.

Coordination with SIM (parallel this round): SIM detects catchers by recipe shape and by the constant MITE_ITEM. The item id is parity-mite and the recipe shape is { inputs: {}, outputs: { 'parity-mite': N } } — identical to the trap, one item id over. catch-mite bandwidth 4 / 120 ticks (a harder catch than a mosquito's 90). The radius is SIM's constant (FIREWALL_RADIUS), not a machine field.

Two commissions that want The Correction alive

• correction-specimen (reward 1000): the flagship — "ONE (1) UNIT OF THE CORRECTION, ALIVE. INVOICE ATTACHED. DO NOT READ THE INVOICE." The fax's first canned mite, in voice. Slots between gibbs-wraith-willingly (900) and feather-proof (1100).
• annual-correction-cull (reward 2100, standing): 5 live mites + 20 melt per collection, forever — you turn the immune system into a cash crop, and The Correction files each absent technician as a rounding error and sends five more. Slots between sodium-requisition (1800) and sky-blue-mislabeled (2400). Re-fillable forever because both wants are producible (the catcher + the melt chain).

parity-mite is producible the moment the firewall exists (an input-free recipe reads as always-runnable to the reachability closure, exactly as mosquito-swarm does), so it is NOT forward canon and never joins the allowlist.

The numbers station stops drifting

SCREEN's round-5 finding was real: the validator recomputed the three planted groups from seams.json but never checked the decoys, and two of them had drifted onto visible seams:

08127 -> (-24,-20) sat inside reel-beds-deep       (x[-30..-21] y[-24..-19])
27503 -> ( -5, 18) sat inside broadcast-shelf-south (x[-14..-3]  y[16..21])

A decoy that decodes onto a real seam isn't a decoy, it's a leak — it hands the player a genuine dig site the station was never meant to reveal. Both re-cut to land just off the seam they were touching (still tantalisingly adjacent, which is the genre's whole cruelty):

08127 -> 08143 -> (-24,-18)   one tile south of reel-beds-deep, on empty ground
27503 -> 27475 -> ( -5, 15)   one tile north of broadcast-shelf-south, on empty ground

And the validator now decodes every decoy to a tile and fails, by name, if it lands inside any seam rect — hidden or visible — not just the three hidden tiles. The map recomputes its plants from the seams and now proves its decoys lead nowhere; it cannot go stale in either direction. (All 22 remaining decoys were re-checked against all 8 seams — these two were the only leaks.)

Accents chosen against the real GLB textures

Round 4 filled accents by derivation; MODELBEAST has since delivered real models, so this round the accents are picked by looking at the actual albedo textures in public/assets/models/ (extracted and inspected). The order named "the ten GLB'd machines"; by the time I looked, nineteen had models (RENDER hot-swapping mid-round), so every machine with a real model now carries a texture-matched accent:

Machine	Accent	Read from the texture
seam-extractor	#48d6ff	the cyan CRT readout on the amber drill body (§4 "CRT-certified")
quantizer	#ff3b57	CRIME-red dial indicator against the pale lilac/dust body
mosh-reactor	#ff7a3f	the molten-orange gauge; matches its MELT output
artifact-bottler	#7fe0d8	the iridescent glass-aqua of the bottled contents
chroma-keyer	#4fff30	the chroma-key green it literally weeps (green-spill)
muxer	#ff3fd0	launch-warning magenta on cold steel; matches the megaglitch
demuxer	#ff3fd4	the magenta channel-split streak (its chroma output)
demosaicer	#4fffc4	the spectral cyan-green band of the rainbow weave
dct-press	#3fffe0	the coefficient-cyan indicator specks on cold steel
field-loom	#6ef58a	the phosphor-green glowing tube
gop-assembler	#5fb0ff	the blue cadence-readout screens
mv-extractor	#3fd8ff	the motion-vector cyan fluid
p-caster	#6fa8ff	the blue projector beam
subsampler	#e04fc8	the iridescent chroma-pink it grades down
bloom-duplicator	#ff9440	the warm bloom-orange furnace glow
decode-asic	#8fe8ff	refined from #6fe8c0: the LEDs are ice-blue, not mint
mosquito-trap	#a64fff	corrected from #d0ff60: the model's zapper target glows UV-violet, not lime
software-decoder	#ff4a2a	kept — the texture's red internals confirm it
archaeology-lab	#ffc24a	kept — the warm dig-lamp amber confirms it

The firewall (no model yet) gets its forged-checkmark #3fff7f art-directed, not derived.

---

The numbers, and the arithmetic behind them

The reference M1 chain (measured, not guessed)

Computed from data/*.json against the machine list in SIM's meltChain() (src/sim/sim.test.ts), with the two recipes that test hand-dials applied:

idle draw (every machine's powerDraw)             57 /s
+ active recipe draw at full tilt                 39 /s
= peak draw                                       96 /s
- compression paid back (quantize -6, 4:2:2 -4, 4:2:0 -8)   18 /s
= net peak demand                                 78 /s

Units (RULED round 3): gen/draw are bandwidth per second; stored and bufferCap are bandwidth-seconds. My round-2 assumption held.

Round 4: the tech tree was a total bootstrap deadlock

The v4 gating rule — any id in a TechDef.unlocks is locked until that tech completes — turned three rounds of tech authoring from decoration into a lock on the entire game. I had written every node assuming gating didn't exist. Measured against the rule:

TICK 0, nothing researched:
  melt producible?          false   <- first-taste, the opening commission, impossible
  analog-pack producible?   false
  spatial-pack producible?  false
  temporal-pack producible? false
  container-pack producible? false
RESEARCH CLOSURE: 0 of 24 techs reachable — every node stranded behind its own output

Every pack chain was gated behind packs you could not make. disc-artifact-bottling cost 10 analog packs and unlocked the only machine that can bottle an analog pack. The M1 melt chain itself (mosh reactor, subsampler, press-anchor-slab) was gated, so the game opened on an unbuildable factory and an unfillable commission, with no way to research out of it.

The fix is a shape, not a patch. Two rules now hold, and both are tests:

1. The M1 melt chain is starting equipment — ungated, so first-taste is possible at tick 0. Ungating it also fixes RENDER for free: they derive a machine's era from the tech that unlocks it and treat unmapped machines as starting equipment, which is exactly what the M1 chain now is.
2. One pack must be craftable at tick 0, or research can never begin. That pack is the SPATIAL PACK — macroblock bricks + ringing halos + slurry, all of which fall out of the quantizer the player already owns. Your first research is a study of your own damage, which is the most in-voice thing in the tree. Everything else chains off it: spatial → analog → temporal → container → fortress.

Result: 23 / 23 nodes reachable, unlock order tells a story (phosphor seams → field loom → analog packs → telecine → temporal → container → fortress), and the only items unreachable with the whole tree bought are the four declared forward-canon.

Six disc nodes were deleted, not repurposed, because their whole content became starting equipment: disc-artifact-bottling, disc-chroma-subsampling, disc-macroblock-quarrying, disc-ringing-skim, disc-subsampling-420, disc-mosh-reactor. Grepped first — no lane references tech ids (RENDER reads unlocks generically, never a specific node). Recorded here per the deprecation rule.

Round 2's heat was decorative — what went wrong

Worth writing down, because it's a trap anyone can fall into. I sized round 2's heatPerTick: 0.0025 against no cooling. SIM then implemented heat with cooling shed every tick, active or not (HEAT_DISSIPATION = 0.004). So the decoder shed 0.004 and banked 0.0025: net −0.0015/tick. It could never warm up. The number looked like a mechanic and was a comment.

The lesson is now a test (every heat source out-heats its own cooling), and it caught a second instance immediately: chart-stco-map banked 0.06 heat over 90 ticks while the splicer shed 0.09 — I'd authored a fresh decorative number in this very round.

Anything with a heat number must beat its own machine's cooling, or it isn't a mechanic. The validator now refuses both forms (per-tick machine heat, per-craft recipe heat).

The software decoder: powerGen 55 → 100, heatPerTick 0.0033, coolPerTick 0.001

Net while running = 0.0033 − 0.001 = 0.0023/tick. Against SIM's curve (throttle 0.7, scram 1.0, restart 0.5):

throttle, from cold   0.7 / 0.0023 / 30  = 10.1 s   (target ~10 s)
scram,    from cold   1.0 / 0.0023 / 30  = 14.5 s   (target 13-15 s)
steady cycle, up      0.5 / 0.0023 / 30  =  7.25 s  (heat 0.5 -> 1.0)
steady cycle, down    0.5 / 0.001  / 30  = 16.67 s  (cools 1.0 -> restart 0.5)
period                                     23.9 s   (target ~25 s)
duty                                       30.3 %

Why the decoder cools slowly (0.001, vs the sim's 0.004 default): it is an overclocked tower PC held together with zip ties and prayer (§4). It is the worst-cooled object in the factory. The ASIC has "a huge heatsink afro" and no heat source at all.

Why powerGen had to nearly double. SIM's hysteresis (scram at 1.0, restart at 0.5) means any machine that throttles at ~10 s spends ~70% of its life cooling. At round 2's 55 gen that's an average of ~14/s — strictly worse than the ASIC's steady 20, so nobody would ever build one and "burst gen" would be a lie. Averaged properly (generation is throttled linearly 1.0 → 0.5 across the throttle window, and zero while scrammed):

un-throttled  0.2 / 0.0023 =  87.0 ticks at 100/s
throttled     0.3 / 0.0023 = 130.4 ticks at avg 75/s
energy per cycle = 100 x (87.0 + 130.4 x 0.75) = 18480 bandwidth-ticks
average = 18480 / 717 ticks = 25.8 /s

	peak	duty	average	failure mode
DECODE ASIC	20 /s	100%	20 /s	none, forever
SOFTWARE DECODER	100 /s	30%	25.8 /s	total blackout, 16.7 s at a time

Five ASICs at full song, one third of the time, in one 2×2 footprint. 29% more average power for a supply that vanishes on a schedule you don't control — a real choice, and the answer to "does anyone dare run two side by side" is still no, not naked: identical decoders heat in lockstep and therefore black out in lockstep.

bufferCap: 240 → 300

Re-derived against the retuned cycle, since the order was "one tank should cover one decoder scram" and the scram window changed:

reference chain: peak draw 96 /s, compression pays back 18 /s -> net demand 78 /s
balanced plant:  3 ASIC (60/s steady) + 1 software decoder (25.8/s avg) = 85.8 /s   ✓
decoder scrams:  supply = 60 + 18 = 78 /s vs draw 96 /s  ->  deficit 18 /s
                 18 /s x 16.67 s  =  300 bandwidth-seconds
bufferCap = 300  ->  one tank rides out exactly one scram.

And it refills in time to do it again: while the decoder is up the plant makes 67/s more than it spends, recharging 300 in 4.5 s of the 7.25 s up-phase. That is the M2 sentence — a buffer tank is what makes software decoders survivable — and it now closes.

The ASIC cooler goes live: coolPerTick 0.02 → 0.0018, coolRadius: 2

Round 3 spec'd the aura at 0.02 while it was inert. The moment it became mechanical that number was catastrophic: the decoder banks 0.0033/tick, so a 0.02 aura means net −0.0167 — it never heats at all, and a 2×2 cooler turns a 2×2 decoder into a permanent free 100/s. A spec'd number and a live number are different objects; this is the second time this round that a number only became wrong once something read it.

Round 5, FINAL under the additive ruling. Contracts v5 settled it: auras are additive, capped at HEAT_COOL_MAX. That killed 0.0018 — additive, two coolers reach 0.0046 against the decoder's 0.0033 gross, so a casual pair made it immortal. Final value 0.001:

coolers   cooling   throttle at   duty    avg gen    note
   0      0.0010      10.1 s      30 %    25.8 /s    bare decoder, as the codex demands
   1      0.0020      17.9 s      61 %    51.5 /s    the pairing — output doubles
   2      0.0030      77.8 s      91 %    77.3 /s    heavy investment, nearly continuous
   3      0.0040       never     100 %   100.0 /s    immortal: an absurd shrine to one PC

The shape is deliberate. One cooler is the worked example and it doubles the decoder's average output — a real reward for engineering around heat. Two is a serious commitment that nearly eliminates downtime. Three tips into immortality, which costs 12 tiles of cooler wrapped around 4 tiles of decoder — visibly ridiculous, and in-voice: the codex's software decoder is a tower PC "held together with zip ties and prayer", so a shrine of coolers keeping one alive is the correct picture. Immortality is reachable, never casual.

Linear stacking means some number always cancels the heat; my job was choosing which number, and making sure it isn't 2. Per tile, when both machines are 2×2:

2 decode ASICs              40.0 /s over 8 tiles = 5.00 /s per tile
1 software decoder, bare    25.8 /s over 4 tiles = 6.44 /s per tile
decoder + 1 cooler          56.5 /s over 8 tiles = 7.06 /s per tile   (incl. the cooler's own 5)

So the pairing is the best power in the game per tile (1.4× ASICs) — and it costs two techs, two machines, and the floor space to keep them adjacent. That is the reward for engineering around heat instead of suffering it, which is what the cooler is for. The bare decoder still beats a pair of ASICs per tile, but it does so in violent 30%-duty bursts that a buffer tank has to absorb; the ASIC's 5.00 never wavers. Three ways to buy power, three different failure modes.

Research: is the first node affordable off a first-taste-era factory?

Worked path for the bootstrap pack, one bottler, one quantizer:

bottle-spatial-pack = 2 macroblock-bricks + 1 ringing-halo + 2 chroma-slurry, 60 ticks
  2 bricks  <- quantize-crime  x1  (1 coefficient pack, 20 ticks)
  1 halo    <- skim-ringing    x1  (3 coefficient packs, 40 ticks)
  4 coefficient packs <- quantize x4 (4 luma, 120 ticks) <- 6 mdat ore
THE QUANTIZER IS THE BOTTLENECK: 20 + 40 + 120 = 180 ticks per spatial pack
                                = 10.0 packs / minute on one quantizer

Against the early nodes:

broadcast-phosphor-seams   5 spatial =  0.5 min
reel-silver-seams          8 spatial =  0.8 min
disc-roto-bench           10 spatial =  1.0 min
broadcast-field-loom      15 spatial =  1.5 min   <- opens the ANALOG PACK

A player who has shipped their first melt already owns a quantizer and can afford the first four nodes inside two minutes of production. That's the intent: the opening research is paid for with the quantizer's own sins — you study your own damage. Later nodes scale to tens of minutes and demand the analog/temporal/container chains, which is where the floor space goes.

heat on the hex splicer: coolPerTick 0.001, 0.08 per splice

Net 0.035 per craft after the 45 ticks of cooling → 28.6 crafts → 42.9 s of continuous splicing before it scrams (chart-stco-map: 50 s). A lesser fever than the decoder's 14.5 s, exactly as ordered. §4: "Requires ... a steady buffer feed or it splices itself" — left running unattended, it stops itself.

---

Open / deferred

• The ASIC cooler's aura is spec'd, not built (above). Until SIM has auras, heat is survivable only by duty-cycling or by leaning on ASICs. Deliberate.
• cinema-pack, gibbs-wraith, sodium-light remain forward canon (no recipe; mosquito-swarm retired from the list in round 5 when the trap landed). The validator's FORWARD_CANON allowlist is the enforcement, and it has a companion test asserting nothing sits on the list once it is producible. Round 3 retired container-pack's slot (round 2) and added none.
• Dials-as-recipes (quantizer Q, subsampler grades, GOP cadence) stays, per round-2 review.
• The heat constants are mirrored by hand CLOSED (round 5): LANE-SIM exported them in src/sim/constants.ts and the validator imports them. The copies are retired; if SIM retunes the curve my arithmetic now fails loudly instead of asserting the old feel.
• The muxer, moov brain and megaglitch are data-only; no lane implements a victory condition yet. The tech gate (fortress-the-muxer) keeps them out of reach meanwhile.