assay
ROBINHOOD CHAIN · ID 4663 · READ-ONLY DESK · --:--:--

The crowd sets the price.It is wrong just often enough.

> ilsa · recalibrating , open markets · k = 1.18

On Robinhood's event contracts, the crowd is right most of the time. assay is a six-agent desk that finds the few markets where it is predictably wrong, sizes the bet by Kelly, and refuses the rest. Read-only. It never touches funds.

BLOCK ,GAS ,CHAIN ,RPC rpc.mainnet.chain.robinhood.com
ROBINHOOD CHAIN
$ASSAYassay · prediction-market desk● LIVE ON PONSBUY $ASSAY ↗
contract address
0x96d2b15b35f43e0113a362b6d522a27a6a9d2138
MEAN RETURN SNAPSHOT
+2.68%
500 seeded runs
RUNS IN PROFIT SNAPSHOT
66%
a third still lose
OPEN MARKETS LIVE
,
reading now
CHAIN HEIGHT LIVE
,
Robinhood · id 4663
01

The live desk

Six agents scoring a batch of markets, then a paper book. Every row is either read from the market or produced by an agent you can open and read. Nothing here places an order: there is no wallet in the codebase and no --live flag to find.

assay ~ /desk · six agents · paper
marketpriceassayedgescoreverdict
the desk · 6 agents
assay scan in a terminal
The same desk in the terminal: assay scan. Ilsa's estimate, Kett's size, Bram's verdict, one card per market.
Rigo pulls the universe of open markets
Holt optional research, capped and labelled
Ilsa estimates the true probability
Kett sizes at quarter-Kelly
Bram refuses most of it, names the rule
Tess owns the paper book
02

The edge is the crowd's own bias

One documented inefficiency, corrected in the open. Across a century of betting and prediction markets, longshots are overbet and favorites underbet. Ilsa bends the crowd price back in logit space by a single slope, and prints how far it moved on every card.

crowd price → assay estimate → favorites up longshots down
slope k = 1.18
favorite 90% → assay , ,longshot 10% → assay , ,
q = sigmoid( logit(price) × k )
drag the slope. k = 1 trusts the crowd. above 1 corrects the bias. every verdict moves with it.
recalibration the whole edge, in one slope you can argue with
momentum a 24h tilt, capped at four points
research optional model vote, capped and labelled
nothing else no hidden state, no confidence theatre
EDGE BY CROWD PRICE · where the money is · recalibrated at k = 1.18
03

Six agents, one book

Each agent does one job and hands its output to the next. Not chatbots passing English around. Click one to read what it actually does.

04

The scorecard · simulated

500 seeded runs, 120 markets each, at the default rules. The offline simulator bakes the target bias into its ground truth, so this measures whether the strategy captures a bias that is actually there, not whether that bias exists in the wild. Labelled simulated on purpose.

+2.68%
mean return
+3.40%
median return
66%
runs in profit
8.14%
std deviation
12.3
positions per run
+20.7%
best run
−27.6%
worst run
22 / 0
tests / runtime deps

The spread is the honest part. A thin edge over a dozen bets is high variance: a third of runs lose even when the edge is real and known. Any version of this that shows a smooth line going up is lying to you.

05

Most of what the desk does is decline

Bram refuses a proposal until it clears every rule below, and names the one that stopped it. That is the whole design: assay is a sizing-and-refusal engine with an estimate bolted on the front, not an AI that presses BUY.

06

On Robinhood Chain

Robinhood gave everyone the buy button, then event contracts: prediction markets a retail account can actually reach. Robinhood Chain (id 4663) is an Arbitrum Orbit L2 that settles to Ethereum and produces about ten blocks a second. Everything in this row is read from its public RPC in your browser as you look at it.

BLOCK HEIGHT LIVE
,
sampled once per second
CHAIN ID LIVE
,
eth_chainId
GAS PRICE LIVE
,
eth_gasPrice
BLOCK TIME LIVE
,
measured, blocks / second

assay is independent of Robinhood and uses none of its marks. The desk reads public endpoints and nothing else: it holds no key, signs nothing, and cannot spend. If the RPC is unreachable these tiles fall back to a dash rather than a number the desk did not read.

07

Method & limits

Which numbers are live and which are computed once?
Anything with a LIVE badge is read while you look at it: the chain tiles from the Robinhood Chain RPC, the open-market count from a public prediction-market API. Anything with a SNAPSHOT badge was computed by a script over a fixed grid and carries that label permanently.
Is this a buy signal or a copy-trading product?
No, the opposite. A research desk that mostly refuses. It estimates a probability, sizes a hypothetical bet at quarter-Kelly, and books it on paper. There is no wallet code in the repo and no --live flag, because the executor is deliberately left out.
Is the estimate an AI black box?
No. Two lines by default: a logit recalibration of the crowd price and a capped momentum tilt, both printed on every card with their signed contribution. The optional research agent is capped and shows its reason.
Does it beat the market?
Unknown, and honest about it. A dozen thin-edge bets is high variance, so a third of simulated runs lose even when the edge is real. That is why the scorecard shows a spread, not one flattering number.
Can I read the code?
Yes, all of it: github.com/bl888m/assay. Six agents, the sources and the sizing math, 22 offline tests in CI, zero runtime dependencies. The desk above is that pipeline rendered from its own snapshot.