Wallet copy for dummies

Plain-English walkthrough of the ogsfrompoly wallet-copy agent, the one that mirrors two hand-picked skilled traders. When it copies. When it refuses. When it leaves. And which numbers you can turn.

A second small agent runs next to the published track record. It mirrors two individual traders — wallets our skill test already flagged, promoted from the same leaderboard the weekly statements are built on. Where cluster copy waits for a crowd of skilled wallets to agree, this one asks a different question: what does one proven trader do when they really mean it?

Think of a very patient apprentice shadowing two masters. It ignores almost everything they do. Then one of them commits serious money to a fresh bet, and it places one small copy of its own — $5, on its own separate account, behind its own kill switch.

Every number on this page is the live setting for the Macro markets as of 2026-08-14, read from config/wallet_copy.yml. Not sure which of the two agents you wanted? The chooser compares them side by side.

Try it — the agent's decision, live

defaults = Macro config, 2026-08-14

Does the leader really mean it? Copy leader-a only once their bet reaches $500 leaders[leader-a].min_leader_notional_usdc Copy leader-b only once their bet reaches $100 leaders[leader-b].min_leader_notional_usdc

Should I copy this at all? Never run more copies of one leader than 2 max_open_copies_per_leader Refuse a leader trading both sides beyond 0.50 max_wallet_two_sided_ratio Skip markets thinner than $1,000 min_liquidity_usdc Skip markets resolving within 1h min_seconds_to_resolution Skip if the price already ran more than 3% staleness_pct …but treat a move smaller than this as noise 2 steps staleness_min_ticks Never pay more per share than $0.95 max_entry_price

How much can I put at risk? Size of one copy $5 trade_size_usdc Never hold more open than $20 exposure_cap_usdc Never let the spendable balance drop below $5 working_capital_floor_usdc

How do I get filled? Walk away rather than overpay by more than 1% slippage_pct …but always give the order at least this much room 2 steps slippage_min_ticks

Incoming signal

leader-a built a $650 position in “US CPI prints below 3% year-over-year in September” at $0.64.

First — would the service even start?

✓ Can my ticket even buy 5 shares at the worst price I allow? $5 → 5.2 shares at $0.97 ≥ $4.85 the service refuses to start — no signal is ever read

Then upstream — does the tracker raise a signal at all?

✓ Has leader-a bet enough to mean it? $650 at peak ≥ $500 no signal is emitted — the agent is never asked

Then the agent's own rails, in order

✓ Am I already copying this leader enough? 0 open for leader-a < 2 the real agent stops here
✓ Do I have room under my cap? $5 open + $5 ≤ $20 the real agent stops here
✓ Is there anyone here to trade with? $5,200 ≥ $1,000 the real agent stops here
✓ Is this market about to end? 6h ≥ 1h the real agent stops here
✓ Has the price already run past the leader? 0.9% (6 steps) ≤ 3% or < 2 steps the real agent stops here
✓ Does this leader trade both sides of this bet? 0.08 ≤ 0.50 the real agent stops here
✓ Is a share too expensive to be worth the downside? $0.644 ≤ $0.95 the real agent stops here
✓ Will this take me below my floor? $23 → $18 ≥ $5 the real agent stops here

COPY $5

Every check passed. It sends a $5 fill-or-kill order at a limit of $0.652 — about 7.7 shares — it fills completely or dies. A killed order gets no retry. Then it holds through small trims and sells the whole leg the moment leader-a has unwound 50% of their peak — or flipped, or closed out.

The real executor stops at the first ✗ in the last list and moves on. We score every check here so you can see what each knob does. Prices are always on the outcome the leader really holds — a position built by selling one side is a position in the other. The engine also runs a row of bookkeeping refusals — is this bet already copied, is the market in scope, is the kill switch on. We leave them out because none of them is a number you can turn.

With JavaScript off this shows our live config against the first example signal. The sliders and signal buttons need JavaScript to recompute.

And when does it leave? Drag the leader's selling

The leader has sold this much of their peak position 30% the signal, not a setting Close everything once the leader has unwound 50% trim_close_fraction

their peak position close line

HOLD

The leader has unwound 30% of their peak — under the 50% line. That is information, not an exit: the agent writes it down and keeps the whole position.

A flip to the other side, a close to zero, a leader removed from the config, or the market resolving — each of those closes the leg whatever this slider says. Only the trim line is a setting.

The whole story, one step at a time

1. It follows two traders, not a crowd

We picked the two wallets it mirrors from our own skill leaderboard — the same measured, classified roster the rest of this site reports on. On this page they are only ever leader-a and leader-b, which are literally their labels in the config. We never publish their addresses. Flag a skilled wallet in public and the whole market front-runs them, and watching them watch us would leak our own positions too.

One signal here is one wallet's actions. There is no quorum to wait for and no crowd to count. That is exactly why every remaining check works differently than on the cluster page.

2. It only moves when the leader really means it

Each leader has a conviction bar (min_leader_notional_usdc). The agent copies a bet only once that leader's own position on it reaches the bar. The bar sits per leader, against that leader's own history, because the same dollar figure does not mean the same thing twice.

Measured over 90 days: leader-a's positions have a median of $285 and a 90th percentile of $1,729. A $500 bar selects their most convinced third, and our $5 copy is about 1% of what they stake. leader-b's median is $5. At a $500 bar this rail would fire roughly once a quarter while reading as an armed leader, so their bar is $100. A knob that looks armed and never fires is worse than no knob.

The bar fires once per fresh bet, at the moment it is first crossed. Scale-ins never re-trigger it. A bet becomes fresh again only after the leader closes it in full — or flips sides.

3. It counts what the leader holds, not what they clicked

On a two-outcome market, buying one side and selling the other are the same opinion. So the agent measures conviction in what the leader ends up holding, at the price of that outcome — not in the words on the trade ticket.

The difference is not small. Selling 500 shares of a nearly-worthless outcome at half a cent collects about $2 of cash. It is also roughly $498 of conviction on the other outcome, and the agent now counts it that way. The entry price follows the held outcome too. Before we fixed that, a position a leader had built by selling could show a price drift of thousands of percent and get refused as stale — a bug that rejected exactly the trades it existed to copy.

4. It refuses a leader who trades both sides

The rail most likely to refuse an entry. Before copying, the agent looks at the leader's own flow on this bet over the last 48 hours and compares buys to sells (smaller ÷ larger, in dollars). Above 0.5 (max_wallet_two_sided_ratio) it refuses. A wallet trading both sides of the same bet is making a market, not expressing a view — and every flip a copier follows is a full, paid round trip. On the cluster agent, one live flip-follow cost $0.25 on a $5 position — about 5% — to end up short what it had just been long.

One honest subtlety: an empty window is "nothing to weigh", not a clean score of zero. When the 48 hours contain no flow to measure, the rail stands down and says it was unmeasured instead of pretending the leader passed. We added the distinction after real skips showed most early readings were exactly this: zeros over empty windows, dressed up as evidence.

5. The usual safety checks

The rest of the entry gate is the same family of questions the cluster agent asks, in the same spirit:

"Is there anyone here to trade with?" — under $1,000 of liquidity (min_liquidity_usdc), skip.
"Is this market about to end?" — under 1 hour to resolution (min_seconds_to_resolution), skip.
"Has the price already run past the leader?" — skip if the ask has moved more than 3% past the leader's effective entry (staleness_pct) and more than two price steps (staleness_min_ticks). A move the market's own grid cannot subdivide is noise, not information.
"Is a share too expensive?" — never pay more than $0.95 (max_entry_price). At the top of the price range a copy risks $5 to win cents, and one bad resolution erases hundreds of wins.
"Am I already copying this leader enough?" — at most 2 open copies per leader (max_open_copies_per_leader), so one leader on a spree cannot eat the whole budget. This rail has no cluster equivalent.
"Do I have room, and a floor?" — at most $20 open in total (exposure_cap_usdc), never spending the account below $5 (working_capital_floor_usdc).

6. What it deliberately does not check

Three rails from the cluster agent are missing here. Honesty about that matters more than symmetry:

No fee rail. The cluster agent's most expensive lesson — the exchange's per-share fee was 77% of its early losses — produced a fee ceiling there that works out to a $0.60 minimum price. This agent has no such rail. A 7¢ outcome its sibling would refuse sails through. We are betting that a leader's own conviction already filters the cheap lottery tickets. That bet is not yet proven.
No reversal window. The cluster agent refuses a crowd whose members just switched sides. Here the two-sided flow rail plays that role for a single wallet.
No gas check at entry. The venue's relayer pays for settlement, and the trading account structurally holds no gas. An entry-path gas rail refused every entry before we removed it. Gas is a payout question: the weekly sweep keeps 2 POL back (gas_reserve_pol).

7. A ticket that cannot exist refuses to boot

The order is the same shape as the cluster agent's: fill-or-kill, at a limit of the current ask plus 1% of room (slippage_pct) but never less than two price steps (slippage_min_ticks), snapped onto the market's own price grid. It fills completely or dies. A killed order gets no retry.

The venue's 5-share minimum works differently — at boot, not per order. The service works out the worst limit price its own rails allow (with a $0.95 ceiling and two steps of room, that is $0.97) and refuses to start if the ticket cannot buy five shares there. Under $4.85 a bet, the agent will not boot. A config that cannot trade should fail once, loudly, at startup — not once per signal forever.

8. It holds through trims, and leaves at half

Exits are where this agent most differs from its sibling. The cluster agent leaves at the first sale by anyone in the crowd. This one is harder to shake out on purpose:

A leader trimming their position is information, not an exit. The agent records it and holds.
Once the leader's cumulative selling reaches 50% of their peak position (trim_close_fraction), the agent closes its whole leg at once. Measured against the peak, not per fill — a leader bleeding out 20% at a time would never trip a per-fill rule. And always the whole leg, never a slice: a proportional sliver can drop under the venue's 5-share minimum and become unsellable.
A leader going to zero, or flipping to the other side, closes the leg immediately. A flip also starts a fresh bet — the agent can close and re-enter on the other side in the same breath, exactly as the leader did.
Remove or disable a leader in the config and the agent closes their legs within a minute. When a market resolves, the agent redeems it.

The shape of this is hard-wired. The trim threshold is the one number in it you can turn — the next step adds two more, and they do not watch the leader at all.

9. And a stop-loss and a take-profit, on its own numbers

Step 8 is entirely about the leader. This is not. About once a minute the agent prices a full exit of every open leg — what the best bid would pay, minus the exit fee, against what the leg cost including the fee it paid going in — and closes it if that reading is down more than auto_close_loss_pct or up more than auto_close_profit_pct.

It matters more here than on the sibling agent. This one has no fee ceiling at entry (step 6), so a cheap outcome the cluster agent would refuse outright can be copied — and a cheap outcome is exactly where the per-share fee bites hardest. A rail that measures what would really come back, rather than where the price is, is the guard that absence leaves room for.

The mechanism is shared code with the cluster agent, so it behaves identically on both: two consecutive readings past the same line before anything happens; an unreadable book skips the leg for that minute and forgets the breach it was holding; the whole leg goes at once, never a slice; and the kill switch does not gate it, because halting entries must never strand an open position.

Both numbers are read at boot, and 0 — or an absent key, which parses the same way — stands that side down. Both are absent today. The rail is deployed on this agent and armed on neither side, so nothing has ever closed this way, and the values we eventually arm will not be published. A live threshold is something a reader could trade against while the leg is still open.

10. A failed exit is queued, not forgotten

An exit order can die just like an entry. Each one gets three attempts, tracked per position in a durable queue — one stuck leg cannot block the others — and if the third fails the agent says so loudly instead of going quiet. A market that resolved but is not redeemed yet sits in the books as unpaid. The books never count cash they do not have.

11. The kill switch stops buying, never selling

At any moment the kill switch halts new entries while exits, retries and settlements keep working — so a halt can never strand an open position. It stops taking risk; it does not abandon it.

12. Its money cannot touch the other agent's

Wallet copy trades a separate account with its own key, its own books, its own alert channel, its own kill switch and its own service. A wallet-copy bug can spend only wallet-copy money. Neither agent can even read the other's records. The one thing they share, on purpose, is the destination of the weekly profit sweep — named in the config as profit_destination behind a destination_allowlist, with a $1 dust threshold (dust_threshold_usdc). We never publish the addresses themselves.

13. Where it is today

The feature went live on 2026-08-11. It spent its first days correctly copying nothing. With bars this high, silence is the expected state, and the operations tooling exists to prove that silence is health rather than breakage. Its first real position followed within days. That is the entire live history so far. Read every number on this page as a version 1: argued from measurement, barely exercised in production.

And when does it sell on its own numbers? Arm the rail

Both sliders start where the running config has them: off. Arm either side and watch the same position play out — the rail reads the book about once a minute and wants the same answer twice before it does anything.

Sell if the position is down more than off auto_close_loss_pct Sell if it is up more than off auto_close_profit_pct

Price path

$5 bought at $0.40 — 12.5 shares, fees included both ways.

what a close would realize rail
min 1 -10.6% inside both lines
min 2 -20.2% inside both lines
min 3 -29.7% inside both lines
min 4 -34.5% inside both lines
min 5 -39.2% inside both lines

Disarmed

Neither threshold is set, which is what the live config says, so this rail never fires. The position leaves the way it always has — when the exit rule above says so, or at resolution.

The rail is deployed on both agents and armed on neither. When we do arm it we will not publish the numbers: a threshold is something a reader could trade against while a position is still open. Everything on this page is the rule, not our settings.

What is hard-wired, and not a setting

The conviction bar fires once per fresh bet. Scale-ins never re-trigger; only a full close or a flip resets it.
The agent counts positions in the held outcome. Buying one side and selling the other are the same opinion, normalised before any rule looks at them.
Exits close the whole leg, always — at the trim threshold, on a flip, on a close-out, on config removal, at resolution, or when the auto-close rail fires. There is no partial close. The rail's two thresholds are numbers you can turn; the two-reading confirmation and the whole-leg close are not.
One leg per bet per leader. The same market copied from both leaders is two independent legs.
Entry is a fill-or-kill order priced on the exchange's grid, and the venue's 5-share minimum is enforced at boot, not per order.
Macro markets only, qualified against the same catalog as everything else on this site.
The kill switch spares exits. Only entries can be halted.

Why split it that way? The same reason as on the cluster page: thresholds are numbers, and numbers are safe to expose. The copying logic itself — what counts as conviction, how exits work — lives in code. One careless line of YAML must never derail the strategy.

What we are not claiming yet

The live history is days old. The feature went live on 2026-08-11 and has opened a single-digit number of positions. We argue every rail on this page from measurements taken before it was armed. Almost none has a meaningful record of firing in production.
There is no fee rail, and that is a bet, not a proof. The cluster agent's fee lesson (77% of its early losses) has no guard here. If the leaders' conviction does not filter cheap outcomes the way we expect, this page will say so — after the books do.
The two-sided rail's early evidence was thin. Most of its first readings were zeros over empty measurement windows. The payload now separates "measured clean" from "nothing to weigh", but the rail has yet to prove its threshold against real traffic.
Redemption is not automatic. A resolved win sits in the books as closed but explicitly unpaid until it is redeemed — the books refuse to count money that has not arrived.
The auto-close rail has never fired. It shipped disarmed on both agents, so step 9 describes a mechanism with no live record at all. Its arithmetic is pinned to the same evaluator the agent runs, which is a different claim from knowing how it behaves against real books.

What this page is not

These are the operating parameters of our own small agent, published in the same spirit as the rest of the track record. They are not advice, not a signal service, and not a claim that mirroring skilled traders will keep working.

We never publish any of the wallets involved — the two leaders, the agent's own trading account, or the payout address. Watch any of them in real time and you see open positions, which is exactly what the disclosure policy exists to prevent.