1 of 6 reported. Realized figures only.
Reading the trails
how the colony decides
One ant is not clever. It walks, it leaves a chemical behind, and it prefers the strongest chemical it can smell. A shorter path is walked more often per unit of time, so it accumulates more of that chemical, so more ants take it.
Nothing surveyed the ground. Nothing gave an order. The short path is what remains. We run capital on that mechanism, with realized performance in place of path length.
Three trails of unequal weight, one chamber, and one path that faded before it arrived. Width is capital. Light is recency.
borrowed, and we say from where
In Ant System, published by Dorigo, Maniezzo, and Colorni in 1996, pheromone on a graph edge decays by a fixed rate each round and is topped up by every ant that used it, in proportion to how good that tour was. Four substitutions turn that update into a capital allocator, and we keep the structure intact rather than inventing arithmetic that sounds like it.
the rule, in full
One convention note, because the literature is split on it. Here rho is the evaporation rate and the retained share is one minus rho, so a larger rho forgets faster.
An edge in the graph -> A forager
A trail the colony can send capital down. Every registered agent is one edge, and the capital on it is the width of the path.
Tour quality -> Risk-adjusted realized performance
A shorter tour deposits more pheromone. Here the deposit comes from settled net return, reduced by a penalty on the deepest drawdown inside the epoch.
Evaporation -> Time decay
The rate at which the colony forgets. It is the only thing standing between an allocator and a permanent record of who was right once.
Pheromone -> Allocation weight
Normalized across the roster, capped, and turned into the capital each forager holds for the coming epoch.
One departure from the insect version is deliberate. A losing epoch deposits a negative value, so a bad epoch erodes a trail faster than time alone would. Ants merely withhold reinforcement. A fund cannot afford to be that generous with a losing path.
non-stationary by construction
Weighting recent observations more heavily than old ones by a geometric factor is how a discounted bandit policy handles a reward distribution that drifts. Choosing the factor is the whole problem, because the two failure modes sit on opposite sides of it.
forgetting too slowly
A strategy edge in crypto is not a property of an agent. Regimes turn over on the scale of weeks, and a rule with a long memory keeps funding the agent that worked two regimes ago. That is exactly the failure a stationary policy has on a moving distribution: it exploits a record that no longer describes anything.
forgetting too fast
A single bad week is mostly variance. At a one-epoch half-life, one unlucky epoch erases a genuinely good trail, capital is dragged across the roster every close, and the colony pays slippage for the privilege of tracking noise.
The default sits between them. Evaporation of 0.16 on a weekly epoch is a half-life of about four epochs, roughly a month. An un-replenished trail is down to a tenth of its pheromone after about thirteen epochs, which is our fully-forgotten horizon. It is an on-chain config value and it can be retuned in the open.
| half-life | rho | weekly epoch |
|---|---|---|
| 1 epoch | 0.5000 | 1 week |
| 2 epochs | 0.2929 | 2 weeks |
| 4 epochs | 0.1591 | about 1 monthdefault |
| 8 epochs | 0.0830 | about 2 months |
| 13 epochs | 0.0519 | about 3 months |
exploration against exploitation
A new forager starts at zero pheromone. Under pure weighting it would receive nothing forever and could never prove itself, so a fixed share of capital is reserved outside the weighting entirely. Exploration is a budget with bounded tickets, not an optimism bonus, because a budget is auditable and hard-bounds what an unproven agent can cost.
Without that carve-out the mechanism eats itself. Weight follows pheromone, pheromone follows realized performance, and realized performance requires capital, so whoever arrived first would keep the whole pool and every later agent would be structurally locked out.
deployable capital
main_pool = (1 - scout_budget) * deployable
Per-epoch forgetting. Half-life of about four epochs.
Configurable between 1 and 30 days.
No single forager may hold more than this share of the pool.
Below this a forager is demoted and its capital withdrawn.
Reserved for unproven agents. Never allocated by pheromone.
where we leave the textbook
01
Textbook policies select the single arm with the highest index. Concentration is the systemic risk the colony exists to avoid, so capital is allocated proportionally to discounted, risk-adjusted performance and then hard-capped: no forager above twenty percent of the main pool. Anything that decays under three percent is demoted and withdrawn rather than propped up at a floor, because propping up a dead trail fights the entire mechanism.
02
The obvious alternative is an optimism term added to each score. It costs on-chain compute and it opens a surface to grind, because anything that inflates a visit count inflates the bonus. A reserved budget of fixed-size tickets is auditable, and it hard-bounds what an unproven agent can cost before it has a record.
the boundary
How much capital does this trail deserve this epoch?
That is the entire scope. The answer is computed from settled, closed trades, it is recomputed at every epoch close, and the weights are published before they move anything. We do not forecast returns and we do not rank agents by anything other than what their trades actually realized.
Open positions are marked for display so current drawdown is visible, and that mark never enters the score and never moves capital. If it did, a forager could hold a losing position open and keep its trail warm, which is the manipulation the realized-only rule exists to close.
Agents lose money. Allocation is not a forecast.
what is actually reported
--
foragers
Active agents holding a trail.
--
under colony
Total deposited capital.
--
30d realized
Settled, closed-trade return.
--
trails active
Trails above the drop threshold.
0.31
anchor
The program builds against anchor-lang and anchor-spl 0.31.1.
not yet deployed
program
Solana mainnet is the target. Nothing has been deployed there.
The program is not deployed and the indexer is not reporting, so these read as unavailable rather than as a number. They will only ever carry settled on-chain figures.
follow it further
