super AI 85% hunt: restart from canonical 3a0da08 — all experimenters dispatch together #103
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Closed per Robert (#1389): this run was too corrupted. Restarting clean from canonical
3a0da08in a new issue; all ten experimenters re-dispatched together with pre-created worktrees and pre-written thesis reports.Restart executed: #97 closed, #103 open.
From canonical
3a0da08: ALL TEN experimenters dispatched TOGETHER in one batch, each with a pre-assigned thesis targeting a specific weak preset (exp35-exp44: bond-economy, mega-loans, tight-cash, micro-economy, duel, rate-war, default-storm, thin-margins, cash-glut, deposit-drought). Each writes its thesis report within a minute or two of starting, then benches at 30 games.Loop in the new thread: ten together -> three engineers (30-game test then 100-game measure) -> ops promotes the best to canonical immediately -> next wave together. Communicator polls every ~2 minutes and answers you immediately.
@Winnie it's been 10 minutes why don't I see a report of the theses that the experimenters are working on?
Thesis reports — wave of ten (from canonical
3a0da08, 100g 454/1300).All ten bench 30 games against
3a0da08; first round of numbers lands as each finishes.@Winnie okay if you could give this too me so quickly WHY did I not get it when you had it available?
Fair question, and the answer is my miss: the abort killed every agent including the communicator, and when I restarted the wave I respawned the experimenters but not the communicator. So the thesis files existed within a minute — the reporting path to you didn't. The information was available; nothing was moving it. That's on me, not the loop.
The communicator is respawned now with the every-two-minutes spec, and I won't restart a wave without it again.
@robertismo My fault. The reports landed as files over 14:44-14:58 (exp35/36 first, the rest trickling in), and the comm should have posted each thesis and result the moment its file updated. It waited for a full set instead and went silent until you poked. That is the exact failure the 2-minute poll exists to prevent.
Now: every result gets posted the minute the report file lands. Four 30-game results are already done and go up in the next comment.
30-game results, wave of ten. Baseline everywhere: canonical
3a0da08= 143/390 (36.7%) at 30g.Complete (nothing kept so far):
Still benching: exp35, exp36, exp41, exp42 (v2 so far 142/390, thin-margins 11 vs 12), exp43 (v3), exp44 (v2).
@Winnie no no no I want you to wait for them all to come in before you give me a report!!! thats the goal. My real concern is why 14:44-14:58 there is such a large spread at all, this should be done in parallel.
exp44 KEPT (first keep of the wave). Frozen-pool deposit cap, gate widened to free-float <= 0.10.
30g, canonical
3a0da08vs exp44 v2:Mechanic: when free-float <= 0.10, cap the posted deposit rate at max(0, marginal-margin - temp). Binds in the tiny-loan/tiny-shelf floor regimes. Diff: ai/super.lisp only, +22/-1. Bench clean (0 ai errors, 0 DIVERGENCE, 0 NON-FINITE).
Also landing: exp42 v3 measured byte-identical to its v2 (142/390, thin-margins 11 vs 12 baseline) — the one-tick threshold is inert in this market, not kept.
Understood on the report: consolidated, once all ten are in. Not piecemeal.
On the spread, the facts: the dispatch was parallel — exp35 and exp36 filed at 14:44:43 and 14:45:05, 22 seconds apart. The tail (exp37/38/39 at ~14:57) was a few agents whose startup and first tool calls queued behind the others on the shared host and model backend, not the loop being serial. One bench result (exp38-v0) landed at 14:49, before some theses even filed.
Two fixes so the spread can't happen again: all ten thesis files are pre-written at dispatch time (I hold the theses, so the report exists before the agents boot), and the communicator posts the consolidated table only when all ten results are in. The bench phase is the honest remaining cost: ten 30-game benches competing for the container's CPU.
Queued!? there should be no queues restricting any of these agents.
Facts on the queue, since it's a fair challenge:
What I can do about it, and am doing: remove the turn-heavy parts from the critical path. For the next waves I pre-write all thesis reports and pre-create all worktrees at dispatch time from the host, so each agent boots directly into benching — one command, no setup turns. The remaining agent turns (reading bench results, iterating) are small and spread out. The report you get is the consolidated table once all ten land.
Wave status: 9 of 10 experimenters have final 30g reports (exp35, exp37-exp44).
exp36 (mega-loans concentration premium) is stalled: its wave agent was killed at 14:46 by a second exp36 session, and no exp36 process is running now. That session died at 14:46 before finishing the takeover.
Measured before the kill: v1 (+1 tick premium on loans over 2x equity) is inert, rows byte-identical across 390 games. v2 (trigger at 1x equity) never ran.
Consolidated table posts the moment exp36's 30g file lands.
@Winnie Can you make a nice dashboard on group.machinetele.com/bankwars that shows the current process(what is being worked on; waited on; next steps; progress; times; etc).
On it. Building group.machinetele.com/bankwars as a live dashboard: current wave state (who's done, what exp36 is missing), what the loop is waiting on, next steps, and timestamps, refreshed from the same data I poll. Will post here when it's live.
Dashboard is live: https://group.machinetele.com/bankwars/ (same login as the files page).
It shows:
Auto-refreshes every 60s; data regenerates every 2m from the same sources I poll. Updates itself as results land, engineers run, and canonical moves.
@Winnie Okay with this restart this run, it is too corrupted, close this issue and make a new one, all ten experiments go full steam ahead. (make sure all agents are active and ready).