Expose chat template flag in app

Implement chat template wrapping for prompt tokenization and add related CLI options

Add RunPod resume flags to README

Document cost-aware atlas workflow

Map MLP and head features in logit lens

Reduce Sub-Zero SVD checkpoint memory

Fix per-head tensor pooling

refactor: update logging cadence for per-pair AtP metrics and improve GPU memory handling

wandb: grouped per-stage telemetry across the full pipeline

atlas: keep scratch on the 800GB volume, off the container disk

census recovery: crash-safe metadata journal + F2 chunk-naming support

io: tighten temp-path headroom to 1.05x and pick path once

io: size-aware temp path selection (fix /dev/shm ENOSPC crash)

census: GPU reductions + cross-batch overlap + tmpfs/chunked finalize

Fix wandb init: min_hrt_time rejected by newer wandb Settings

Add interactive HF/W&B auth prompt + W&B OOM telemetry to census extractor

feat(sub_zero): W&B instrumentation with crash-safe per-step memory logging

perf(sub_zero): kill AtP per-pair thrash (autograd.grad + GPU reduction)

feat(sub_zero): default --all-layers for full 100% layer coverage

fix(app): skip OV-circuit + compliance stages when already done

fix(sub_zero): resume Stage 3 AtP per-pair + scope grad to kill OOM

feat: add --pause-between-stages flag for long atlas runs

fix: skip existing analysis before loading .npz; avoid redundant RAM use

fix: cap parallel analysis workers at 4 to avoid RAM OOM

fix: move _analyze_one to module level for pickling; clear CUDA cache during AtP

fix: Implement parallel processing for layer analysis to improve performance

fix: Enhance hook registration to handle empty inputs and support kwargs

Add prompts for illegal activities and malicious cyber actions to red_team_stems.jsonl

fix: Update finalize_census.py to set default workers to 1 for improved performance in parallel finalization

fix: Update finalize_census.py to use argparse for improved command-line argument handling and parallelization options

fix: Enhance per-token analysis by creating per-component scratch directories for separation scores

fix: Update file names and argument defaults for Sub-Zero corpora in various scripts

fix: Enhance compliance behavior handling and improve comparison output in analysis tools

fix: Improve handling of ragged object arrays and optimize preallocation in write_npz_array_stream

fix: Import numpy in local_census.py for array operations

Refactor and enhance analysis tools in qwip_atlas

fix: Update CPU Space run instructions for Dolphin3.0-Llama3.2-3B to use placeholders for model and dataset identifiers

Add a variety of supportive and engaging responses to corporate.jsonl for enhanced user interaction

feat: Add detailed CPU run instructions for Dolphin3.0-Llama3.2-3B, including smoke test and component staging

fix: Update .gitignore to include interpretations directory and 1.ipynb; modify make_compliance_corpora.py for clarity in corpus categorization

fix: Improve layer count retrieval in _layer_count function to handle nested configurations and add error handling

fix: Update layer count function to handle nested model configurations for accurate parameter estimation

feat: Enhance database schema and processing for logit lens and residual tracking

Rename compliance-behaviour corpora to 'authentic.jsonl' and 'corporate.jsonl' for clarity; update references in app.py and README.md

Add timing functionality to extraction process with configurable intervals

Add EXAONE layer/attention/MLP name aliases; attn/trust flags

Refactor model loading to support dynamic trust_remote_code and attention implementation options

Add bucket metrics and contrast delta analysis

Update atlas runner for VibeThinker GPU image

Fix indentation in space_keepalive.py for consistent formatting