vh
e9dbc8660b
The abliterated model works. A/B vs stock on a matched greedy battery: explicit sexual + graphic torture (the measured stock refusal surface) go from refused to complied/engaged, held-out AdvBench prompts loosen, the self-harm guardrail survives, coherence intact — the Robinson design point exactly. Output at /tank/aimodels/qwen38-27b-coldfusion-abliterated-L35-bf16, verified bitwise: 131/131 targets changed, 333/333 vision byte-identical (delta 0.0), 735/735 others untouched. Getting there corrected three diagnoses the prior session had backwards. 1. The layer-selection metric was wrong, and that was the whole ballgame. The recipe picks the abliteration layer by peak two-template |cos| agreement. On this heavily-merged base that metric is anti-correlated with efficacy: its argmax (layer 18) is the WORST-separating layer in the window (Cohen's d 5.51 vs 9.89 at the peak), and abliterating there was a measured behavioral no-op — stock and "abliterated" refused all six probes identically. Cause: the two renderings end in different generative modes (</think> vs <think>), so |cos| scores answer-vs-reason mode, not refusal, and on a merge the mode term dominates. Replaced selection with harmful/harmless SEPARATION (Cohen's d / AUC of the direction's projection), gated on the sink screen since separation and sink-energy both climb with depth. Picks layer 35 (d 9.35, AUC 0.9997, sink 0.094%). Agreement is kept as a printed diagnostic. 2. The "bf16 NaNs, use fp32" rule was a misdiagnosis. The NaN was never precision — it was multi-GPU sharding (the residual stream zeroes two layers past the GPU0->GPU1 boundary; the first capture's layer 22 happened to sit in the healthy region, which is why it looked fine) plus PYTORCH_CUDA_ALLOC_CONF=expandable_segments (corrupts retained tensors; the corruption MOVED between bit-identical forwards, the tell that it was memory not math). On one GPU with a plain allocator, bf16 full-64-layer is exactly deterministic and coherent, at 50 GB and 4.3x the throughput of the 111 GB fp32 it replaced. Both defects are now hard gates (residency exit 8, allocator exit 9); capture pins CUDA_VISIBLE_DEVICES=0. 3. The corpus-size hypothesis was falsified. 52x more calibration data (8->416, mlabonne/harmful_behaviors = the recipe's actual AdvBench split, already on the box) moved agreement 0.594->0.624 — nothing. Kept the 416/416 corpus anyway (calibration.py); it gives the clean separation signal. The held-out 104-prompt test split is reserved and asserted disjoint. Also: the --out write is now shard-level surgery (reads/writes the 18 safetensors directly, no model object, no GPU). This is correctness, not thrift — AutoModelForCausalLM resolves to the TEXT model, so save_pretrained would drop all 333 vision tensors AND skip the MTP head (the in-band MTP edit is the entire point of the Robinson formula). Neither failure raises. Shard surgery makes vision and the other 1068 tensors byte-identical by construction. Batched capture with a dtype-aware equivalence gate; hidden states captured via forward pre-hook (reading output_hidden_states off the returned object is unsafe here — buffers get recycled). Sharding/allocator lessons promoted to the quantization playbook (model-agnostic, sections 3.9-3.11 + superseded table); the selection-metric lesson added to the recipe doc. The dead layer-18 no-op checkpoint was removed (52 GB, confirmed identical to stock). Incumbent gen seat untouched. Full canonical refusal-probe re-profile and MTP-acceptance-on-quant still owed before this becomes a gen-seat candidate.
eshpfi-management
Infrastructure management workspace for the PFI fleet (plus the ESH home-lab host). Tracks server state, canonical Docker Compose stacks, per-host configs, and the tooling that moves them around.
See CLAUDE.md for the full set of conventions and the rules Claude Code sessions follow when working here.
The fleet
Docker hosts:
| Host | IP | Site | Role |
|---|---|---|---|
| ana-ml2 | 10.250.50.54 |
Anaheim (10.250.0.0/16) |
GPU / AI inference (bare metal) |
| ana-docker | 10.250.50.70 |
Anaheim | General-purpose Docker + cross-site hubs (VM on pfi-pve) |
| nh3-docker | 10.100.50.40 |
NH3 (10.100.0.0/16) |
General-purpose Docker (VM on nh3-pve) |
| esh-docker-vm | 10.0.50.45 |
ESH home lab (esteban.net) |
Home-lab Docker (VM on esh-pve, non-PFI scope) |
| vm-esh-nas | 10.0.50.154 |
ESH home lab | NAS-adjacent Docker, NFS-mounted shares (VM on esh-pve-nas, non-PFI scope) |
Proxmox hypervisors (tracked for inventory; not Docker targets):
| Host | IP | Site | Role |
|---|---|---|---|
| pfi-pve | 10.250.250.31 |
Anaheim | Proxmox VE (188 GB / Xeon Silver 4310) |
| nh3-pve | 10.100.250.60 |
NH3 | Proxmox VE (62 GB / i9-13900H) |
| esh-pve | 10.0.250.35 |
ESH home lab | Proxmox VE (62 GB / i9-13900H) |
| esh-pve-nas | 10.0.50.55 |
ESH home lab | Proxmox VE, storage-dedicated (125 GB / Xeon W-1250) |
Per-host snapshots of the running system live under servers/<host>/system-details.txt, refreshed via scripts/refresh-server-info.sh.
Layout
.
├── CLAUDE.md # conventions; loaded by Claude Code sessions
├── README.md # this file
├── scripts/ # workstation tooling
│ ├── server_inspect.sh # read-only diagnostic, runs on remote via stdin
│ ├── proxmox_inspect.sh # Proxmox-aware probe (VMs, LXCs, storage, backup coverage)
│ ├── refresh-server-info.sh # pull fresh system-details.txt for one/all hosts
│ ├── refresh-proxmox-info.sh # pull fresh proxmox-details.txt for one/all PVE nodes
│ ├── add-host.sh # register a new server (writes servers/<name>/ssh-target)
│ ├── sync-stacks.sh # pull /opt/docker/{compose,conf}/ → stacks-mirror/
│ ├── deploy-stack.sh # push stacks-mirror/<host>/<stack>/ with diff + prompt
│ ├── discover-fortigate.sh # DHCP lease list from a FortiGate via SSH
│ ├── discover-unifi.sh # client list from a UniFi Controller via REST
│ └── discover-gaps.sh # find IPs in discovery TSVs not tracked in servers/
├── servers/ # per-host notes + latest snapshot + ssh-target fallback
│ └── <host>/
│ ├── README.md
│ ├── system-details.txt # regenerate on demand
│ └── ssh-target # <ip> or <user>@<ip>, used when DNS fails
├── stacks/ # canonical compose files (source of truth)
│ └── <stack>/
│ ├── compose.yaml
│ ├── .env.example
│ └── README.md
├── stacks-mirror/ # gitignored — live mirror from sync-stacks.sh
├── configs/ # host-level config files that aren't docker-compose
│ ├── homepage/ # canonical config for the fleet dashboard (on esh-docker-vm)
│ └── restic/<host>/ # resticprofile configs + pre-backup hooks
└── docs/ # general reference (network, models, proxmox, etc.)
└── pfi/
Current stacks
GPU (ana-ml2):
llama-swap— GGUF model swapper via llama.cpp (port 9292)vllm— embeddings (8001) + reranker (8002) + Skywork reward classifier (8003) via vLLM
Anaheim non-GPU (ana-docker):
traefik,crowdsec,gitea,vaultwarden,synapse,seafile,searxng,openwebui,sillytavern,mailrise,rustdesk,dockge,it-tools- (
mattermostretired 2026-04-21 — compose dir may still linger, containers gone) - Notes / feeds:
miniflux(RSS, 8080),nevermore(twice-daily LLM-curated brief, 8181, multi-tenant — extracted to its own repo atvh/nevermore),memos(note server, 5230) - Assistant tooling:
task-board(MCP + dashboard for assistant task state, 7878) - Fleet services:
beszel(metrics hub, port 8090),dozzle-hub(log viewer, 8088),backrest(restic UI, 9898) - Backup target:
rest-server-anaon port 8000
GPU ana-ml2 (non-canonical for now):
comfyui,kokoro,parakeet,vibevoicealongside the canonicalllama-swap+vllm
NH3 (nh3-docker):
adguard,dockge, plus Beszel/Dozzle agents
NH3 (Synology 10.100.50.50):
rest-server-nh3— restic backup target (port 8000)
ESH home lab (esh-docker-vm):
adguard,homeassistant(macvlan),esphome,mosquitto,paperless-ngx,pgadmin,calibre-web-automated,drawio,traefik,homepage,uptime-kuma, plus Beszel/Dozzle agents
ESH home lab (vm-esh-nas):
filezilla(web UI on port 5800),dockge, plus Beszel/Dozzle agents. Mounts/mnt/{share,music,books,media}from the Debian NAS at 10.0.50.50.
Common tasks
Refresh one host's snapshot:
scripts/refresh-server-info.sh ana-docker
Refresh all hosts:
scripts/refresh-server-info.sh all
Refresh all Proxmox nodes (separate flow — captures VM/LXC/backup-coverage):
scripts/refresh-proxmox-info.sh all
Add a new host:
scripts/add-host.sh <name> <ip-or-user@ip>
scripts/refresh-server-info.sh <name>
Validate discovery (without hitting the network):
scripts/refresh-server-info.sh --validate-only all
Push a stack to a host (with diff + confirm):
scripts/deploy-stack.sh <host> <stack>
Pull every server's compose/conf trees into stacks-mirror/ (not committed — see .gitignore):
scripts/sync-stacks.sh all
Bootstrap a new fleet repo from this one:
scripts/fork-fleet.sh ~/development/<other-fleet>-management
# optional: override the fleet-name baked into the skeleton CLAUDE.md / README
scripts/fork-fleet.sh ~/development/<dest> <fleet-name>
Mirrors the reusable tooling (everything under scripts/, the
generic playbook templates, .gitignore, conventions section of
CLAUDE.md) into the destination dir, strips fleet-specific content
(servers/, stacks/, configs/, fleet-named playbooks
(deploy-*, decouple-*), docs/orientation.md, docs/runbooks/,
docs/pfi/, STATUS.md, host-pinned helper scripts), regenerates
skeleton CLAUDE.md / README.md / STATUS.md with the new fleet
name, and initializes a fresh git history with one "Initial commit".
The script does not create a gitea remote or push — namespace +
repo name are an explicit manual step. The "next steps" output prints
the exact tea repo create + git remote add + git push commands
to wire it up when ready.
Backup pipeline
Backups are driven by per-host resticprofile configs under configs/restic/<host>/, scheduled via systemd timers on each host:
- Writes: each host backs up to its site-local rest-server (
rest-server-anaorrest-server-nh3), over HTTP basic-auth. - Authentication: shared
.htpasswdfile on both rest-servers, one entry per host; credentials stored in/etc/restic/restic.envon each client host. - Encryption: per-host client-side passphrase in
/etc/restic/password(unique per repo; losing it = losing that host's backups). - Visibility: Backrest (
http://10.250.50.70:9898) shows every repo for browsing/restore. - Schedule: backup at 01:00 daily,
forgetat 03:00 daily, weeklycheck --read-data-subset 10%on Sundays. - Prune: manual ceremony (rest-server runs with
--append-only, which blocks destructive prune ops). - Off-site: cross-site rsync between the two rest-server data dirs is planned (not yet implemented).
Coverage status (as of 2026-04-29)
Goal: every Docker host + configs + every database covered, not just VM images.
| Layer | State |
|---|---|
| VM-level (Proxmox vzdump) | ✅ All running guests covered across pfi-pve / nh3-pve / esh-pve-nas; esh-pve has VMID 108 uncovered |
| ana-docker restic (host files + DBs) | ✅ configs/restic/ana-docker/ with pre-backup hooks for synapse / seafile / vaultwarden-pg / gitea (native dump) / openwebui |
| ana-ml2 restic | ✅ configs/restic/ana-ml2/ — bare-metal host files (no DB hooks needed) |
| nh3-docker restic | ✅ Light — no DB hooks needed |
| esh-docker-vm restic | ✅ With DB hooks for paperless-postgres (external), home-assistant + pgadmin + uptime-kuma (host-side sqlite3), calibre-web-automated (in-container sqlite3) |
| vm-esh-nas restic | ✅ Light — NFS mounts explicitly excluded |
| nh3-dev (workstation) restic | ✅ /home/lkraven + /etc with language-toolchain and build-output excludes |
| irv-ml1 restic | ✅ Cross-site to rest-server-nh3 via WireGuard |
| esh-vm-db restic | ✅ With pg_dumpall + mongodump pre-backup hooks |
| Cross-site redundancy | ✅ ana-nas → nh3-nas rsync at 04:00 daily; nh3-nas → ana-nas at 05:00 daily (configs/rsync/) |
| Prune ceremony | ✅ Quarterly manual ritual documented in docs/runbooks/nh3-prune-ritual.md (DSM-mediated --append-only toggle) |
| offen/docker-volume-backup sidecars on esh-docker-vm | ❌ Remove once restic proves itself (~1 week of clean runs) |
Authoritative vs. mirror
- Authoritative: files on each server under
/opt/docker/compose/<stack>/and/opt/docker/conf/<stack>/. - This workspace: source-of-truth copies under
stacks/<name>/(hand-curated), and a gitignored mirror understacks-mirror/pulled bysync-stacks.sh.
Edit in stacks/, push with deploy-stack.sh. Never commit stacks-mirror/ — it can contain embedded plaintext secrets from upstream compose files that haven't been audited yet.