diff --git a/README.md b/README.md
index aacdf50..de2cdd8 100644
--- a/README.md
+++ b/README.md
@@ -29,6 +29,7 @@ Current test rig: 1x B200 (180 GB) pod — connection details in `pod-info.txt`.
   (SGLang / vLLM-Omni / diffusers), concurrency model, RunPod serverless
 - `docs/BENCHMARKS.md` — measured results on the B200
 - `docs/CONCURRENCY.md` — concurrency model + cost-per-video math across GPU tiers
+- `docs/MITIGATIONS.md` — per-machine-class OOM map and required ComfyUI flags/patches
 
 ## Quick start
 
diff --git a/docs/MITIGATIONS.md b/docs/MITIGATIONS.md
new file mode 100644
index 0000000..44747cc
--- /dev/null
+++ b/docs/MITIGATIONS.md
@@ -0,0 +1,66 @@
+# MiniMax H3 — per-machine-class mitigations & OOM map
+
+Compiled 2026-08-06 from measured failures and fixes across five pods.
+Companion to `docs/BENCHMARKS.md` (runtimes) and `bench/failures.py` (stats).
+
+## The two resources that fail, and their fixes
+
+**1. Container RAM (cgroup cap — check `/sys/fs/cgroup/memory/memory.limit_in_bytes`,
+cgroup v1 on RunPod consumer pods, or `/sys/fs/cgroup/memory.max` for v2).**
+ComfyUI stages every loaded model's weights in host RAM ("prepared for dynamic
+VRAM loading"). Task-switching between fl2va and ref2va stacks WITHOUT eviction
+stages TE (15 GB) + two DiTs (20 GB each) + overhead ≈ 60 GB+. Exceeding the
+cap = silent SIGKILL, no traceback, "server died" mid-run.
+Fix: `--cache-none` (one model staged at a time) + `--disable-pinned-memory`
+(staging stays pageable/reclaimable). Cost: weights reload per run — measured
+~2.3× slowdown on the RTX 4000 (551 s vs an est. ~240 s), ~1.4× on the 4090.
+
+**2. VRAM, in two distinct phases:**
+- *TE vision encode* (multi-ref): budget bug, fixed by ComfyUI **PR #15316**
+  (apply: `git fetch origin pull/15316/head:pr15316 && git checkout pr15316 --
+  comfy/text_encoders/minimax.py`) + `--reserve-vram 4`–`5`. Ref pre-downscale
+  (`--ref-downscale 0.2` in our harness) also clears this phase alone.
+- *DiT sampling with ref tokens / long clips / large canvas*: needs the
+  activations to fit beside the (partially loaded) 20 GB DiT. On 24 GB this is
+  the hard wall: 15 s clips, 1344×768, and multi-ref all exceeded it pre-patch.
+  `--lowvram` made it WORSE (crashes) — do not use for H3.
+
+Hygiene for all capped machines: `PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True`
+(set automatically by `start-comfyui.sh`), torch ≥ 2.5 (`enable_gqa`), and for
+Blackwell (B200/5090/RTX PRO 4000) torch ≥ 2.7 with cu128 wheels.
+
+## Machine-class matrix (all measured, not theoretical)
+
+| Pod class            | VRAM   | RAM cap | Required config                                              | What works                                        | What fails                    |
+|----------------------|--------|---------|--------------------------------------------------------------|---------------------------------------------------|-------------------------------|
+| B200 ($6.94/hr)      | 180 GB | ~3 TB   | none (defaults)                                              | everything incl. 9-ref, 15 s, full-res, bf16      | nothing observed (96/96)      |
+| RTX 5090 ($0.99/hr)  | 32 GB  | 124 GB  | #15316 patch + `--reserve-vram 4`                            | multi-ref up to 8 full-size refs confirmed; 9-ref + max sizing under test | TBD (matrix running) |
+| RTX 4090 ($0.74/hr)  | 24 GB  | 62 GB   | #15316 patch + `--reserve-vram 5` + `--cache-none` + `--disable-pinned-memory` | t2v/i2v ≤10 s @ ≤864×480, 1-ref ref2v; multi-ref under re-test with patch | 15 s clips, 1344×768, multi-ref (pre-patch) |
+| RTX PRO 4000 ($0.57/hr) | 24 GB | 29 GB | same as 4090 (mandatory even for t2v)                        | t2v/i2v, 1-ref (551–588 s — uneconomical)          | multi-ref untested post-fix; 15 s died |
+
+Key procurement insight: **the container RAM cap, not the GPU, decides the
+mitigation burden.** A 24 GB card in a ≥96 GB-RAM container would skip
+`--cache-none` and its ~1.4–2.3× tax. When renting, prefer pod configs with
+RAM ≥ 4× VRAM.
+
+## Reference-image envelope (node schema + measured)
+
+- The node accepts up to **9 reference images** + 3 reference videos (each
+  with optional soundtrack) + 3 standalone audio clips — "12 refs" = 9 images
+  + 3 videos.
+- `match` sizing (refs scaled down to canvas area) is the default and the
+  cheap path; `max` (2048 px short edge) multiplies sampling cost ~9× at
+  8 refs and is the most VRAM-hungry config in the whole matrix.
+- Ref count scaling on B200: +4 s per extra ref (match). Counts 1→9 and the
+  ref-resolution ladder (0.1 MP → native) are being measured on all three
+  GPU classes right now (`--suite refmatrix`).
+
+## Open items
+
+- 5090: 9-ref and 8-ref-`max` verdicts (running).
+- 4090 with full stack (#15316 + cache-none + reserve 5): multi-ref verdict
+  (running).
+- RTX PRO 4000: dropped from fleet consideration on economics; would need
+  re-test only if its price class matters.
+- Upstream: file the multi-ref OOM issue referencing #15316 (user decision —
+  outward-facing).