all ears, whitepaper // ALL EARS

ALL EARSwhitepaper // v1

collection: 2,048 listeners across five classes, 8 at the top
chain: robinhood chain, an arbitrum orbit l2 settling to ethereum
currency: $EARS, an erc20. the contract accepts no other payment
standards: erc721, erc2981 royalties, ownable2step. no proxy, no pause
model: claude opus 5, effort scaling with class, sonnet 5 where opus declines
permanence: every answer pinned to arweave, hashed on chain

all ears is 2,048 ai listeners. each one answers exactly one question in its existence and then goes silent for good. the question is committed on chain before inference runs; the answer is written to arweave and its hash to the chain. four unasked listeners can be burned to forge one of the class above, which hears longer questions, answers at greater length and reads more of the archive. a listener that has answered can never be forged. supply comes down as people play, and on no schedule of ours. a share of every fee accumulates in a pot the contract holds and no key can withdraw, opened only by revealing a secret hidden across nine places on the internet, which some of the listeners speak riddles about.

1 the listener

one question, one answer, and then it is done.

a listener holds exactly one act. the owner writes a question and pays the ask fee. the question's sha256 hash is committed on chain in the same transaction that marks the listener asked, so what was asked is fixed before any answer exists, and a second ask cannot happen. inference then runs off chain, and the answer is pinned to arweave and its hash recorded.

the mishearing

the base class does not hear well. before the model sees anything, the question passes through a deterministic mishearing, one word at a time. the sound at the front of a word is swapped for one the ear confuses it with, or a consonant is voiced where it was not, and the result has to be a real word or the word is left alone. pot arrives as slot, tell me as sell me, die as tie. the model answers what reached it and never learns what was meant. which word is misheard, and as what, is seeded by the blockhash of the block the ask landed in, which the asker cannot know in advance, so no one can grind phrasings offline until the mishearing lands somewhere flattering. it replays exactly from the published seed, so anyone can check it afterwards.

seedsha256(id : questionHash : blockHash)

heardthe misheard text, derived from that seed, replayable by anyone

answeredwhat the model said to what it heard, within the class word limit

recordedanswer pinned to arweave, hash written on chain by the answer authority

disposition

every listener has a disposition, a short phrase describing what it pays attention to, drawn from a seed committed before minting opened and only revealed once minting had closed. the disposition colours what the listener takes the question to be about. nobody, us included, can mint toward one they want.

2 the ladder

four burned for one, all the way up.

only the base class can be minted. everything above it has to be forged: four unasked listeners of one class burn, and one of the class above is created. the child's token id iskeccak256(sorted parent ids), so the provenance is the identifier, and there is nothing to take on faith. parents must be owned by the caller, of matching class, distinct, and unasked.

class	supply	hears	says	fidelity	archive	effort	ask fee
hard of hearing	2,048	120 ch	7 w	words for words	none	low	4,000
listening	512	500	50	exact	none	low	8,000
attentive	128	2,000	250	exact	none	medium	24,000
rapt	32	8,000	1,000	exact	a sample	high	64,000
clairaudient	8	no limit	no limit	exact	everything	max	200,000

the archive is where the classes really separate. the top two classes get the transcripts of the collection as they stand at the moment somebody asks. a clairaudient asked on the last day has read more than one asked on the first, so leaving it alone is the one thing in here that makes an asset better.

asking never closes. forging shuts at a set time, and after that the class structure is final, but an unasked listener bought years later still holds its question. we did that on purpose. somebody buying on the secondary market should never inherit a thing that can no longer do what it exists for.

3 the artwork

everything at the bottom looks the same. the materials above it have to be earned.

every base listener is the same image. same sculpt, same plain material, same pose. that is deliberate. what everybody buys at the start is a commodity, and all 2,048 of them have the same potential sitting inside. you earn any difference from there.

class	sculpt	material pool
hard of hearing	low	skin, and only skin
listening	mid	concrete, denim, rust, rotten wood, red plastic, skin
attentive	mid-high	terracotta, newspaper, mould, chrome, racing car, skin
rapt	max	wood, stone, boils, copper, porcelain, tatts, plastic, obsidian, galaxy, skin, gold
clairaudient	max, denser	eight glassy materials, one to each of the eight

every class has its own sculpt and never wears the pool above it, so the material tells you the class at a glance. a forged listener draws its material from the hash of the four it consumed, so picking which four to burn is picking the draw. hunting a particular material by choosing parents carefully is a game we want people to play.

there is one trait and it is the material. no poses, no props, no accessories. the only visible difference between two living listeners is the class it reached and the material that class handed it, so nothing in the artwork sits there as decoration competing with what the listener is for.

a listener that has answered wears closed eyes from then on. you can see at a glance which is which, and the two price very differently.

4 the token

$EARS is the only currency, and most of what gets spent is burned.

the contract holds no fee revenue. every payment splits inside the transaction that earns it: burned to 0x…dEaD, added to the pot, or banked to treasury. there is no withdraw function, because nothing accumulates to withdraw. the one exception is the pot, and no key reaches that either.

action	fee	burned	to the pot	to treasury
mint	40,000	60%	25%	15%
forge	16,000	50%	50%	0
ask	4,000 – 200,000	50%	50%	0

what a full collection destroys

event	spend	burned	to the pot
2,048 mints	81,920,000	49,152,000	20,480,000
the ladder forged out in full	10,880,000	5,440,000	5,440,000
every listener asked once	up to 19,008,000	9,504,000	9,504,000

on a supply of one billion, the mint alone takes 8.2% of every token there is and burns 4.9% of them. forging and asking compete for the same listeners, so nothing above is ever reached in full. a listener that answers cannot be forged, and one that is forged no longer exists to answer. treat those figures as a ceiling, and not as a forecast.

one clairaudient costs 256 base mints plus the fees for four rounds of forging, 11,600,000 $EARS, and there can be at most eight. every question asked anywhere below takes a listener out of the pool that could have become one.

prices are denominated in $EARS and set at deployment. they freeze for good the moment the first listener is minted, and the timeline freezes with them.

5 the pot and the hunt

a prize the contract holds, opened once, by whoever works out how.

a share of every fee accumulates in the contract as the pot. we do not stake it, lend it or invest it. it sits there. there are two ways out: someone reveals the secret, or once the hunt has closed anyone at all calls sunset() and it burns. either way, from that moment the share of every later fee that would have fed the pot burns with the rest, so nothing ever accumulates behind a door that no longer opens.

the secret

the pot opens to a preimage P. the contract stores only keccak256(P), committed at deployment before a single listener can be minted. P comes from nine shards, each of them random bytes, hidden in nine different places. inside artwork, in on-chain data, in web infrastructure, in metadata, in an archived page. you cannot guess a shard or reason your way to one. they have to be found.

the keepers

about a hundred listeners are keepers. each carries one riddle naming where a shard sleeps, and says it when asked, as a line it does not understand. nothing marks a keeper out, and there is no way to identify one except by asking it something. twelve separate listeners carry each riddle, so a clue cannot be lost to an owner who never asks.

no listener knows the secret, and none can be tricked into giving it up, because none of them holds it.we never tell the model a hunt exists. the server attaches riddles after inference has finished, so even a perfect jailbreak gets you exactly one riddle, which is the game working. a listener holding nothing is a different matter: ask it where the pot is and it will invent something, because it does not know a hunt exists and a question deserves an answer. we do not try to stop that, and it is why a planted line is never mixed into the answer. it is shown on its own, marked, in every place an answer appears. only what is marked is part of the hunt. everything else a listener says about the pot, however certain it sounds, is the listener talking.

forging a keeper does not destroy its riddle. the fragment passes down to the child, so forging never deletes a clue from the game. it just makes that clue rarer, and whoever holds it holds a better instrument.

claiming

gatherfind all nine shards, join them in order, hash once to obtain P

commitcommitClaim(keccak(P, yourAddress)), which reveals nothing

waitsixty seconds, so a watcher cannot race a reveal they have just seen

revealrevealClaim(P). the commitment is bound to your address, so the pot cannot be stolen from the mempool

fair play

a hash of the keeper map and its salt goes on chain at deployment. once the hunt ends we publish the map and the salt, and anyone can check that keepers and riddles were fixed from the start and never moved about in response to how the hunt was going. one thing the chain cannot prove: the people who hid the shards know the secret, and no contract can stop a wallet of theirs claiming it. we will not, and the published map is how you hold us to that.

6 contract architecture

what every key can and cannot do.

the contract is a single non-upgradeable erc721 with erc2981 royalties. no proxy, no pause switch, no team mint, no rescue hatch. ownership transfers in two steps so it cannot be lost to a mistyped address.

key	can	cannot
owner	set treasury, answer authority and base uri; set royalties; set the allowlist root and wallet cap until minting closes	change prices or the timeline once anyone has minted; mint; touch the pot; alter or remove a recorded answer
authority	record one answer per asked listener; return a listener to unasked if inference never produced an answer	rewrite a recorded answer, record an empty one, or touch a listener nobody has asked
anyone	mint within the window and wallet cap, forge and ask what they own, reveal the disposition seed once minting has closed, claim the pot with the secret, and burn the pot after the hunt closes	mint free, ask twice, forge borrowed or asked or mismatched listeners, reveal any seed but the committed one, or replay a single paid ask into extra inference

the pot is the only balance the contract carries. it is reachable by revealClaim andsunset and by nothing else, a property covered directly by the test suite.

commit and reveal, twice over

two commitments go in before minting opens and cannot be changed afterwards. the disposition and material seed, which stops anyone minting toward a trait they want, and the pot secret, which stops the hunt being altered once it is running. neither reveal is gated on who is calling. the commitment binds what a valid answer looks like, and the clock decides when it is allowed, so the disposition seed can be published by anyone holding it the moment minting closes. that is how asking opens on time without waiting for a person to wake up.

7 verification and permanence

what you can check yourself, and what you have to take on trust.

anyone can independently verify, with no access to our systems:

that the question asked is the question recorded, by hashing the plaintext against the on-chain commitment
that the mishearing is honest, by replaying the corruption from the published seed and blockhash
that the answer recorded is the answer pinned, by hashing the arweave document against the on-chain hash
that dispositions and materials were not manipulated, by hashing the revealed seed against the pre-minting commitment
that the hunt was fixed in advance, by hashing the published keeper map against the on-chain commitment

stated plainlythe answer itself is model inference and is not reproducible. the same prompt to the same model on another day will not come back with the same words, and we have no way of making it. we are not going to call that trustless inference, because it is not. everything around it is: the question, the mishearing, the hashes, the ownership and the pot. one more thing worth saying here: opus declines a narrow band of perfectly ordinary questions about the pot, and no prompt shifts it, so a declined ask falls through to sonnet rather than back to the asker. every transcript records which model actually spoke.

transcripts are pinned to arweave, so the archive outlives the website. once a listener has answered, the collection serves its image from that same permanent record.

8 risks and disclosures

what can go wrong.

nothing here is an investment product. $EARS and the listeners are not securities, carry no claim on revenue and promise no return. the burn and pot mechanics above describe what the code does, and none of it is a forecast.
the pot may never be claimed. if the hunt is never solved, the pot is burned after it closes. no one is entitled to it, including the team.
the team knows the secret. whoever hid the shards holds the preimage, and nothing on chain prevents a wallet of theirs from claiming. the commitment and the published map make a rigged hunt provable after the fact; they do not make it impossible.
a listener can be asked a hostile question. transcripts are permanent. the site can hide a transcript from its own pages, but arweave keeps what it keeps.
an owner can spend a listener's question before selling it. asked and unasked are shown everywhere and every ask is a public on-chain event, so a buyer can check to the second.
inference can fail. when it does, the listener goes back to unasked and the fee stays burned. we do not refund the fee.
a prize hunt funded by fees may be regulated as a promotion, sweepstake or game of chance in some jurisdictions. the claim is skill-based and the pot cannot be withdrawn by the team, but participants are responsible for their own jurisdiction's rules.
smart contracts carry risk. the contract has been reviewed, is covered by a test suite, and its full behaviour is described above. it has not been through a paid third-party audit.
the token is external. $EARS is an erc20 the contract accepts as payment. its market behaviour is not controlled by this contract.

9 parameters

every number in one place.

parameter	value	changeable
supply by class	2,048 / 512 / 128 / 32 / 8	never
forge ratio	4 into 1	never
mint price	40,000 $EARS	frozen at first mint
forge price	16,000 $EARS	frozen at first mint
ask price	4,000 / 8,000 / 24,000 / 64,000 / 200,000	frozen at first mint
mint split	60 burn / 25 pot / 15 treasury	never
forge and ask split	50 burn / 50 pot	never
wallet cap	32 in the public mint	until minting closes
royalty	5% to treasury	by owner
asking	opens at mint close, never closes	never
forging	opens at mint close, closes on a timer	frozen at first mint
shards in the hunt	9, in 9 places	never
keepers	about 100, twelve to a riddle	never
after a claim	the pot share of later fees burns	never
claim delay	60 seconds between commit and reveal	never

ALL EARS // WHITEPAPER V1ONE QUESTION EACH