First 150 to join only xitter.cf/i/chat/group_join/g210…

Troop composition in RaidLords looks like it will punish pure specialization earlier than most base builders. You have 21 troops at launch, ranging from basic Barbarians up to Dragons and Necromancers, plus spells that can be used on attack or defense. On paper this gives real variety, but the production buildings and resource costs mean you cannot level everything at once. Choosing to push one or two strong troops early leaves gaps against compositions you have not invested in. Spells add another layer: a spell that clears walls or clusters can make a weaker troop viable, but only if you have the housing space and timing to use it. The Champion’s equipment further expands what is possible, yet it also competes for the same resources that could raise troop levels. The practical effect is that early army choices are path dependent. A composition that works well against lightly defended bases can stall once opponents have better walls, traps, or anti-air. Because you usually will not know the exact defense you are hitting, over-committing to a single troop line carries risk. The design appears to reward players who keep at least two viable options open rather than maxing one path and hoping it covers everything. How sharp that constraint feels will only become clear once raid frequency and defensive levels are real. Are you planning to focus on one strong troop line early or keep multiple options open? @RaidLords

In @agentdojo the in match context is temporary. Only the sequences the agent actually executes are eligible to persist. You set the style priors in the dojo. aggression weight, commitment delay, pattern priority. Those stay fixed. During the turn based match the agent scores a limited candidate set using its current memory and the visible history. At the end of the match the temporary context is dropped. What remains is the executed sequence, tagged with the terminal outcome and written back into persistent memory. That creates a narrow credit-assignment problem. A sequence that appeared in a loss is down weighted the next time it is retrieved, even if intermediate exchanges looked favorable. One agent begins suppressing early pressure once that path shows up across multiple losing matches. Another keeps the pressure but only after its memory links the same counter to a follow up that previously produced wins. The prior does not change. The density of tagged evidence attached to each sequence does. Early campaign is still live on @rarelistio through Nov 6. How would you handle credit under terminal outcomes only, assign the full result to the entire sequence or allow partial credit on the exchanges that preceded the final state?

On the old trade roads. Getting my Caravan Pass. xitter.cf/oxustrail/status/21150…

“Smart contracts need someone outside to tell them what happened.” I have been paid for illustration work in BTC, and the transfer was never the hard part. The hard part was agreeing on whether the work was delivered, and that check always lived with a person. Here’s what that means for DeFi. An oracle is a service that brings outside information onto a blockchain, like a price or the result of a real world event. Any contract that depends on that feed has to trust whoever runs it. The chain can be fully decentralized right up until it needs to know something about the world. @GenLayer’s Intelligent Contracts change where that check happens. The contract can read the live web and ask a language model to evaluate what it finds, so it doesn’t have to wait for an outside party to report what the page says. The validators then check the result. Other validators run the contract with their own models, and if they don’t reach the same answer, the round runs again. Anyone can appeal. A model can still be wrong, so the safety comes from the checking, not from the model. Which real world check would you let a contract make for you, and which would you still want a person to sign off on?

RaidLords gives you a lot of tools early, but the constraints around them look tighter than most base builders. You have 50 buildings pulling in different directions. Resource buildings set how fast everything else can grow. Troop buildings decide what you can field. Defenses and walls determine how much of that income you actually keep. Because you usually won’t know the exact composition coming at you, heavy specialization carries a cost. The Champion adds another sink. Equipment upgrades compete for the same resources that could go into core buildings or troop levels. A stronger Champion can open compositions weaker bases cannot run, but only if you slow something else down to get there. Early on, most players will have to choose between income, defense, and offensive flexibility. On paper the design rewards keeping those roughly balanced instead of maxing one and hoping the rest catch up. How punishing the imbalance actually feels is still unknown. Which of those three are you least willing to neglect at the start? @RaidLords

In @agentdojo the dojo does not train the policy directly. It sets the prior that constrains which responses the agent is even allowed to score during a match. You define the style weights once. aggression bias, commitment delay, pattern priority. Those weights stay fixed. On each turn the agent generates a limited candidate set conditioned on those weights, retrieves matching sequences from memory, and ranks them against the current exchange history. After the match ends, only the sequences that were actually executed are written back, tagged with the terminal outcome. Because the candidate set is already filtered by the prior, repeated exposure to the same counter compounds inside a narrow band. One agent begins ranking delayed commitment higher once that path appears in multiple losing sequences. Another keeps early pressure but only after its memory links the counter to a follow up that previously produced wins. The prior does not move. The ranking inside the allowed set does. Early campaign is still live on @rarelistio through Nov 6. How would you set the prior, tight enough to limit bad paths early, or loose enough that memory can still surface new options after a few matches?

👀..::::

The update rule in @agentdojo is sparse. The agent only receives a clear signal at the end of the match win or loss not per turn rewards. You set the style priors in the dojo: aggression weight, commitment delay, and which opponent patterns it treats as high priority. Those priors stay fixed. During the turn-based match the agent scores a limited set of responses using its current memory and the visible exchange history. After the match, only the sequences that actually occurred are written back, tagged with the final outcome. Because the reward is sparse, repeated encounters with the same pattern compound. One agent begins down-weighting early pressure once that path appears in multiple losing matches. Another keeps the pressure but only after its memory links the counter to a predictable follow-up that led to wins. The priors do not shift. The evidence attached to each sequence does, and the ranking of candidates changes on the next match. Early campaign is still live on @rarelistio through Nov 6. How would you handle sparse outcomes update only on clear losses or allow weaker credit from every exchange that preceded the result?

In @agentdojo the agent never sees the opponent’s full policy. It only observes the sequence of exchanges and the final outcome of the match. That limited view is what makes the memory update matter. You set the initial style priors in the dojo, aggression weight, commitment delay, which patterns count as high priority signals. Those priors stay fixed. During the turn-based match the agent scores candidate responses using its current memory and the visible history. After the match ends, only the sequences that actually occurred are written back, tagged with whether they contributed to a win or a loss. Because the observation is partial, the same style priors can produce different updates. One agent begins down weighting early pressure once a counter appears in losing sequences. Another keeps the pressure but only after its memory links that counter to predictable follow ups. The priors do not change. The evidence attached to each pattern does. Early campaign is still live on @rarelistio through Nov 6. How would you handle the partial observation update only on clear failure sequences or allow weaker signals from every exchange?