From 2b402da32a9b4886958715e28cdf988259c49388 Mon Sep 17 00:00:00 2001 From: Alexey Rybalchenko Date: Thu, 11 Jun 2026 10:44:33 +0200 Subject: [PATCH] adapt curriculum stage C docs to reflect the current code action --- README.md | 2 +- curriculum_argument.md | 4 ++-- curriculum_setup.md | 4 ++-- 3 files changed, 5 insertions(+), 5 deletions(-) diff --git a/README.md b/README.md index a5911c4..d286bef 100644 --- a/README.md +++ b/README.md @@ -39,7 +39,7 @@ Current intended sequence: 1. Stage A: flat arrivals + logic prices. 2. Stage B: high-load flat arrivals + logic prices. -3. Stage C: expensive-half-heavy or bursty arrivals + logic prices. +3. Stage C: bursty arrivals + logic prices. 4. Stage D: main arrivals + logic prices. 5. Stage E: main arrivals + noisy logic prices. 6. Stage F: main arrivals + real prices. diff --git a/curriculum_argument.md b/curriculum_argument.md index 1b8c744..b6972f9 100644 --- a/curriculum_argument.md +++ b/curriculum_argument.md @@ -107,9 +107,9 @@ This stage isolates the phase rule itself. With deterministic 12-hour expensive This removes excess slack. If the cheap half-day can clear everything trivially, the agent does not need to learn disciplined backlog management. Increasing load forces the price-timing decision to matter. -### Stage C: expensive-half-heavy or bursty arrivals + logic prices +### Stage C: bursty arrivals + logic prices -This stage specifically targets the trickle policy. Concentrating or bursting arrivals into the expensive half creates a sharper test: the agent must wait under pressure, then ramp aggressively in the cheap window. +This stage tests queue-spike robustness. A small per-hour probability of a burst injects sudden load spikes on top of base arrivals. Because the burst check is price-blind - applied with equal probability across all 336 hours - spikes are distributed roughly evenly between expensive and cheap periods. The agent must absorb unpredictable surges without destabilizing, while still maintaining the defer-then-clear pattern it learned in Stage B. ### Stage D: main arrivals + logic prices diff --git a/curriculum_setup.md b/curriculum_setup.md index c37eab6..8767b40 100644 --- a/curriculum_setup.md +++ b/curriculum_setup.md @@ -160,9 +160,9 @@ Steps: 1M # next stage continues with --continue-existing-only ``` -## Stage C: Expensive-Half-Heavy or Bursty Arrivals + Logic Prices +## Stage C: Bursty Arrivals + Logic Prices -Goal: force deferral under pressure, then aggressive clearing in cheap hours. +Goal: test queue-spike robustness while preserving the defer-then-clear pattern. Steps: 1M