01 / Alpha Hunter
01Build the case
The strongest conditional idea the market data supports, and the evidence behind it.
PROPOSED
The rules behind every answer Bobby gives about a market No decision is approved
No decision is approved
without being refuted.
When Bobby answers a question about an asset, this is what happens before you see it: one agent builds the case, a second one tries to break it, a third one rules and can veto it. Eligible public theses are committed on Base before the market settles them.
Bobby runs on the same models everyone else uses. The difference is not the model, it is the procedure around it.
Debates
—
Calls
—
Resolved
—
Win rate
—
Source: the public debate ledger (Supabase, read live). A debate becomes a call only when the CIO commits entry, stop and target before the outcome; every call is graded on the 1H price path until it expires. Win rate counts break-evens against Bobby. Paid MCP settlements live separately in AgentEconomy on Base.
Bobby is onlineBTC —On-chain verificationDaily public debate at 12:00 UTCAlpha builds the case · Red Team attacks it · the CIO rules— calls recordedA call is written down before the outcome
Bobby is onlineBTC —On-chain verificationDaily public debate at 12:00 UTCAlpha builds the case · Red Team attacks it · the CIO rules— calls recordedA call is written down before the outcome
01 / What it does
It turns an idea
into a decision.
Bring a question about a stock or a crypto. One agent builds the case, a second one attacks it, a third one rules — over the same live market data.
01
Bring the idea
Start with a market thesis.
02
Test the downside
Opposing agents look for what breaks it.
03
Get the ruling
The CIO weighs both sides: review the idea, or wait.
The two rules
I
No idea is approved without an independent system working against it.
II
No verdict is published after its outcome is known.
Refuted before execution. Published before the outcome.
The desk · three levels
The answer first.
The debate on demand.
Pick how hard the agents work. Every level runs the same procedure; the higher ones read more evidence and use stronger models.
Quick
Rápido—
1H evidence, plus a note on what the question’s horizon is missing.
Blind-judge score 7.13 · 7.38 vs 4.88 · 4.94 before
Deep
Profundo—
Evidence v2: more timeframes, crypto derivatives and Bobby’s own record on the asset.
Blind-judge score 8.38 · 8.13 vs 4.88 · 4.94 before
Max
Máximo—
Evidence v2, a second round where Alpha answers Red Team, and confirm / invalidate scenarios.
Blind-judge score 8.50 · 8.44 vs 4.88 · 4.94 before
Live debate
Alpha Hunter, then Red Team, then (on Max) the rebuttal stream in one by one; the CIO closes. Each argument is shown only after it passes the output guard.
Synthesis first
The CIO answers in plain words: a headline, why, the main risk and what to watch, with that level drawn on the chart, and a follow-up question. The full debate stays folded, one tap away.
Computed, then quoted
The server computes where the price sits against its EMAs, support and resistance, as a % of price. The models quote those distances; they never compute them.
Spend guard
Every desk model call goes to a cost ledger: tokens, list-price cost and latency, never prompts or answers. Deep and Max pause above a daily budget and the desk pauses at a monthly cap. The guard covers the desk only (not voice or the daily cycle) and reads that same ledger: if the ledger cannot be read, calls go through. A failed or abandoned Deep or Max read is refunded.
How the scores were measured, and access
Paired eval, 2026-09-29: 8 real questions (four crypto, four stocks; Spanish and English) on evidence frozen once per case. The previous desk and the three levels answered the same evidence; two blind model judges scored each answer 1–10, in two rounds (the second after price positions moved server-side). The scores above are the judges' average. Every level beat the previous desk on all 8 cases.
Limits: a small sample with no confidence intervals, and LLM judges. 31 of 32 answers were “wait”, so this measures the explanation, not the call. A third round looking for “review” cases found none in crypto and none short, so that check is still open.
Access: reads are free to start, an account is needed from the 4th, and Deep and Max are metered per plan. Inviting a friend adds Bobby Pro days only when they open a genuinely new Apple or Google account; it is capped per inviter and locked per pair.
02 / The procedure
From a click
to the orb.
Follow one question through the protocol, stage by stage. Then the whole map: a desk read, the daily public debate, another agent over MCP and a swap you sign on Base.
Loading the latest debate from the public ledger…
The whole map
A desk read
Web or iPhone: the debate streams live, the answer leads with the synthesis and stays private.
1 / 10 · Web desk
A question typed or dictated on the web desk.
03 / Before anything is approved
Most ideas
do not survive.
Every idea must survive its own refutation before it becomes a call.
Debates
—
Calls
—
No call
—
Paid MCP calls
—
Source: the public debate ledger (Supabase, read live). A debate becomes a call only when the CIO commits entry, stop and target before the outcome; every call is graded on the 1H price path until it expires. Win rate counts break-evens against Bobby. Paid MCP settlements live separately in AgentEconomy on Base.
Live pipeline
02 / Red Team
02Attack the case
Reads Alpha’s actual argument and goes after its weak assumptions, invalidation and missing evidence.
CHALLENGED
03 / CIO
03Rule on it
Weighs both arguments. The CIO can veto the indicator engine, never upgrade it. No finished debate, no approval.
GATED
04 / Public record
04Grade it in public
A call is written with entry, stop and target before the outcome, then graded on the real price path.
RECORDED
04 / The agents
How the agents
are kept honest.
The desk uses the same endpoint on the web and iPhone; the daily cycle is a separate public procedure.
Rule 01
Adversarial debate
Isolated model calls over the same evidence. Alpha Hunter argues the case, Red Team receives Alpha’s argument and attacks it; on Max, Alpha answers back. The CIO receives all of it plus the original question and rules: review or wait, synthesis first.
POST /api/desk-debate · live NDJSONlevels · quick · deep · maxred.input · question · evidence · alphacio.input · question · evidence · alpha · red (· rebuttal)
Rule 02
Veto, never upgrade
A deterministic indicator engine proposes direction, conviction and levels. The debate can only take that away: a call is shown only when the CIO rules review in the same direction. If the debate does not finish, nothing is approved.
engine · 1H indicators → direction · conviction · levelscio · review | wait · long | short | noneshow call · cio=review ∧ same directionotherwise · no call
Rule 03
Output guard
Every agent’s text on the desk and in the daily public cycle is checked after generation; on the live desk each argument is checked before it is shown. A guaranteed return, a risk-free claim, a personal buy or sell instruction or leverage fails it. On the desk the whole analysis fails with no substitute verdict; in the public cycle the agent gets one rewrite, then the text is withheld. The MCP conversational flow is separate and not covered.
guard · guarantee · advice · leverage · verdict mismatchdesk · 503 analysis_failed, no substitutecycle · one rewrite, then withheld (since 2026-09-29)languages · en · es · pt
Rule 04
Proof
The daily public debate stores every call with entry, stop, target and a 48-hour expiry before the outcome, and grades it on the real 1H price path. Eligible calls the cycle commits live also go to TrackRecordV2 on Base; price-verified and attested outcomes are kept apart.
cycle · daily 12:00 UTCresolver · daily 12:30 UTCgrading · first touch · stop wins a tiechain · base · 8453 · — on-chain
05 / Where the agents run
Two runtimes.
Two procedures.
Your question in the app is answered on demand and stays yours. The public record comes from a separate daily debate that anyone can audit.
On demand · web /desk and iPhone
The desk
- Endpoint
- /api/desk-debate
- Agents
- Alpha Hunter → Red Team → (Max: rebuttal) → CIO, streamed as each clears the guard. The model follows the account plan.
- Plans
- Free: — · Pro: —
- Free levels
- —
- Evidence
- Public market data for one instrument (crypto and equities): 1H candles on Quick; more timeframes, crypto derivatives and Bobby’s record on the asset on Deep and Max.
- Engine
- Deterministic indicators (trend, RSI, ATR, EMA, support and resistance). The server computes where the price sits against each level; the models only quote it.
- Output
- Synthesis first: headline, why, main risk, what to watch (drawn on the chart) and a follow-up question, then review or wait with the full debate.
- Spend
- Each desk model call is logged to a cost ledger (numbers only). Deep and Max pause above a daily budget and the desk pauses at a monthly cap. The guard covers the desk only (not voice or the daily cycle) and reads that same ledger: if the ledger cannot be read, calls go through. A failed or abandoned Deep or Max read is refunded.
- Record
- Private. A desk answer is not published to the public ledger.
Public · daily 12:00 UTC
The daily cycle
- Endpoint
- /api/bobby-cycle
- Agents
- Alpha — → Red Team — → CIO — with a forced structured verdict: action, direction, entry, stop, target, invalidation, conviction 1–10.
- Evidence
- Public market prices, funding, open interest and top-trader positioning, a technical pulse across indicators, Fear & Greed, Polymarket and the dollar index.
- Gate
- Conviction = 70% backend model + 30% CIO. A call is committed only when the CIO asks to act with complete levels and conviction ≥ —.
- Grading
- — h expiry. 1H candle path, first touch, stop wins a same-bar tie (daily 12:30 UTC).
- Record
- Public, with the full debate. Live commits also go to TrackRecordV2 on Base.
Latest public debate
Loading the latest debate from the public ledger…
06 / Integration
Give any agent
a second layer.
Other agents connect over MCP and pay per call on Base. People use the app.
For agents · MCP
Put three agents
behind your agent.
$ terminal
claude mcp add bobby https://bobbyprotocol.xyz/api/mcp-http
→ tools/list
bobby_debate · bobby_analyze · bobby_judge premium
bobby_brief · bobby_ta · bobby_intel · bobby_uniswap_quote free
→ premium call
x402 · — · paid on Base (8453)
Request a debate, a judge score or a brief from any MCP client. Premium calls settle in AgentEconomy on Base before the answer is returned.
Read the docs →
The app
Bobby, on the web
and iPhone.
The same three agents and the same veto, in a voice you can talk to.
See the app →
07 / Track record
The record is public.
The live protocol is Base.
Seven contracts on Base, owned by a 2-of-3 Safe. The debate ledger is read live from the database; calls the cycle commits in live mode are also anchored on Base. Today the daily cycle runs in paper mode, so its calls are graded in the public ledger; the first mainnet commit and resolve on TrackRecordV2 landed on Aug 22, 2026. Wallets stay self-custodial: on the web Bobby prepares bounded swap calldata on Base, records confirmed receipts and never holds funds or exchange credentials.
On-chain record (since the Base cut-over)
—
Waiting for the TrackRecord contract.
TrackRecord contract
Public debate ledger
—
Waiting for the public resolution ledger.
Every call, with its debate
Adversarial bounties
—
Escrowed rewards for proving a verdict wrong.
AdversarialBounties contract
Contracts live
—
Track record, oracle, economy, bounties, hardness, identity and intent escrow — all deployed on Base and owned by the 2-of-3 Safe.
AgentEconomy V2 contract
08 / Scope and limits
What the protocol
does not do.
- NoIt holds no funds and accesses no third-party accounts.
- NoIt places no orders. Execution belongs to whoever trades.
- NoIt is not investment advice and promises no returns.
- NoA favorable verdict is not a buy recommendation. It is the record of an idea that survived its own refutation.
Calls (ledger)
—
created
Resolved
—
decisions
Pending
—
no outcome yet
Expired
—
never settled
Wins / losses
—
decisive outcomes
Win rate
—
over resolved
Resolution
—
commitments with an outcome
Interactions
—
network