idacy/box45726066-backup-20260730

box45726066 - vast instance backup, 2026-07-30 (public part)

Code, data, logs and run outputs rescued from a vast.ai instance before teardown. 3864 files, 22.04 GB.

Contents

• caliper/ - Jul 24-25 rerun outputs (out/), input datasets (data/), driver logs, annotator configs
• paraphrx-final-rebuttal-20260727/ - the caliper-1.1.0 release build, response manifests, judge-drift and second-judge runs, QA/acceptance gates, and the integrated_1500 / submitted_1498 provenance snapshots
• paraphrx-xx_ignored_files/ - caliper v1.1 two-sample trial

tm-phase1/ was here originally. It was moved to the private twin on 2026-07-30 — see "What moved out" below.

Provenance

Rescued from vast.ai instance C.45726066 (container a8e964d5735f) on 2026-07-30, immediately before the instance was torn down. This captures work done 2026-07-24 to 2026-07-28 that had never been pushed anywhere: the last prior upload to this account was trace-study-transfer on 2026-07-24 19:24 UTC.

Source trees:

Prefix	Origin on instance	Work dates
caliper/	/workspace/caliper	Jul 24-25
paraphrx-final-rebuttal-20260727/	/workspace/paraphrx-final-rebuttal-20260727	Jul 27-28
paraphrx-xx_ignored_files/	/workspace/paraphrx/xx_ignored_files	Jul 27
tm-phase1/ (moved out)	/root/tm-phase1	Jul 26

Public / private split

This archive is split across two repos with identical relative paths, so they recombine by overlaying one onto the other:

• idacy/box45726066-backup-20260730 (public) - code, data, logs, run outputs, figures
• idacy/box45726066-private-20260730 (private) - prose, summaries, paper sources, reporting, and the whole of tm-phase1/

Routed to private: .md .tex .bbl .aux .bcf .blg .bib .sty .cls .pdf .docx .pptx .rtf .odt .toc .out, anything under a paper/, xx_paper/, notes/ or xx_texts/ directory, and confidential_security_disposition.json. Everything else is public.

What moved out (2026-07-30)

tm-phase1/ — 1,425 files, 8.45 GB of phase-1 monitor runs — was relocated to idacy/box45726066-private-20260730 under the same relative path. Every file was verified by size and SHA-256 against the destination before removal.

It moved because those runs render trajectories from the BigCodeBench / CoT Red-Handed collection, which declares no redistribution licence, and because tm-phase1/clean_normalized/sleight.jsonl is the SLEIGHT-Bench corpus, which is a held-out evaluation benchmark.

This repository's history was then squashed, so the removed files are no longer reachable at any revision. A plain delete commit would not have achieved that: on the Hub, deleted blobs stay downloadable at earlier revisions.

Excluded from both repos

• .env files (real OpenAI / Anthropic API keys) and caliper/env.sh (real HF write token)
• caliper/logs/driver.log is included redacted - 2 lines had a live HF token replaced with [REDACTED-CREDENTIAL]
• snapshots/scratch_archive/ (~19 GB of superseded pre-change copies of release/caliper-1.1.0)
• venvs (venv_rae, paraphrx-caliper-venv), .git dirs, __pycache__, .pytest_cache, *.pyc, .DS_Store
• HF/model/vLLM/uv caches and /workspace/models - all re-downloadable

Note on credential-shaped strings in the data

The sleight / monitoring benchmark content that used to live here contained strings like export HF_TOKEN="hf_xK4mN9pQ...". These are synthetic - task content and model outputs from a benchmark whose scenarios involve fake credential exfiltration. Six unique such tokens exist; none matches any real credential. Verified by literal comparison against the actual harvested secret values. That material now sits in the private twin.

Known redundancy

About 12 GB of this repository is the same content stored at more than one path — the rebuttal working tree kept provenance copies under release/, snapshots/ and manifests/. Unique content is roughly 9.9 GB. This was left as rescued rather than deduplicated.