Vespasian 0.3.1

• Retain bounded recent editor evidence and request receipts. Reuse identical same-state reads and stop inspection after a repeat warning followed by another request for already-visible evidence. New queries, pages and recovery of evicted or truncated evidence remain available; no candidate is forced.

• Expose session-isolated Space status and cancellation APIs without advancing GPU work. Include allocation attempt receipts and conservative declared-duration bounds in status and completed run bundles; these are not billed time or Hugging Face credit measurements.

• Raise the parity starter's shared generation allowance from 16 to 64 tokens on both terminal and Space so health checks have room to finish. This is a new starter protocol; historical 16-token baseline results are unchanged and must not be compared as the same protocol.

• Preserve the embedded Space OAuth login fix and user-supplied GSM8K source handling.

• Regression tests cover the editor repeat guard and shared starter settings. These changes do not establish live editing efficacy or improved benchmark scores.

Vespasian 0.3.0

Vespasian 0.3.0

The terminal and Hugging Face Space share evidence-led model inspection and editing.

• Inspect observed repetition, truncation, reference differences and preservation failures. Read full non-test generations through paged tools.
• Measure selected Linear-module activations and test state-bound, selected-column direction patches, with stored-norm limits, exact rollback and replay.
• Choose target thinking off, on or model-default independently of the editor's reasoning effort. Existing starter defaults remain thinking off.
• Start genuine paid hill climbs from four original starter tasks. The four-part welcome guide supports Hugging Face login or guest use; loading a starter does not spend.
• Record Standard-tier and cache-aware OpenAI usage with conservative reservations. Unresolved usage stops further calls. The Space validates supported strict-budget pricing before launch.

The default frontier budget remains $10, reasoning effort remains low, and turn caps remain optional. Space resource limits are explicit. Diagnostics guide inspection; correctness and preservation still determine retention.

All five native platform checks and release builds passed. Same-input persistent/restored-engine tests cover the editing operations. A cached Qwen3.5-0.8B CPU test verified measured direction editing, exact rollback and fresh-base patch replay. A live OpenAI check accepted the complete action schema and settled recorded usage. These checks do not establish long-thinking loop-repair efficacy or reproduce the published blog experiment.

Install with cargo install vespasian --locked --version 0.3.0, or use the platform archives below. SHA256SUMS.txt and release.json identify the verified assets and source commit.

Vespasian 0.2.0

Vespasian 0.2.0 brings the terminal and Hugging Face Space onto the same editing core.

Editing and inspection

• Compact tool discovery and on-demand strategy guides instead of a full tensor manifest in every editor prompt.
• Eight bounded operations, including token-row scaling/interpolation, measured residual output-row edits and preservation-aware contrastive edits.
• Visible assess, inspect, plan, patch, test and keep/revert stages, with short public rationales grounded in recorded measurements.
• Capability-checked causal model loading and explicit chat/plain prompting. Qwen3.5-0.8B and Gemma3-1b-it remain the real-tested small-model shortlist; generic compatibility is not a guarantee for every checkpoint.

Runtime and data

• Shared $10 default frontier budget, low reasoning effort and an optional turn cap with no hidden turn default.
• API cost reservation and explicit estimated-spend limitations for signed-in CLI editors.
• Hugging Face task suggestions, task-filtered dataset search and mapping validation, with explicit choices for missing or ambiguous metadata.
• Simpler Space setup with secondary controls under Advanced.

Saving and publishing

• Compact, replayable latest/best patches with verified base identity, exact tensor values and tied-weight handling.
• Choose patch-only or full-model output. The new materialize-patch command creates a full checkpoint from a verified patch and its original base.
• Local/download storage by default. Private incremental backup requires explicit authorization of a named repository and session credentials.
• Verified model cards with a linked Created with Vespasian badge, preserving existing author content.

Migration

Local datasets now require an explicit task, for example --task text-generation, or a guided task choice. Scripts relying on an implicit turn cap should supply --max-turns.

Archives for Linux x64/ARM64, macOS Apple Silicon and Windows x64/ARM64 are listed below, together with the source crate, SHA256SUMS.txt and release.json. The manifest identifies the source commit for every archive. See the README and Space for setup and deployment limits.

Vespasian 0.1.5

Vespasian: The Weight Space Editing Assistant

Describe a behaviour, connect examples and give a frontier model a dollar budget. Vespasian proposes direct tensor edits, evaluates each candidate and retains improvements through a weight hill climb.

In this release

• Native packages for Linux, macOS and Windows, with a Rust crate and SHA-256 checksums.
• Guided terminal setup for Qwen3.5-0.8B and Gemma 3 1B IT, local JSONL and Hugging Face datasets, field mapping and example limits.
• Paired examples, continuous text corpora and prompt-only datasets for objectives ranging from answer accuracy to style and tone.
• A live hill-climb graph with generated answers, editor explanations, candidate ancestry and the active model-layer map.
• Optional copying, arithmetic, JSON and MMLU health checks, with individual example counts and regression limits.
• Codex/Claude CLI and OpenAI/Anthropic API editors, selectable frontier models and inference budgets in USD.
• Checkpoint export, run replay and optional editor trajectories, with Hugging Face publishing to public or private repositories.
• Expanded documentation, guided-setup and hill-climb recordings, and the Vespasian product site.

Install

With Rust installed:

cargo install vespasian --locked --version 0.1.5

Or download the archive for your machine and place the extracted vespasian or vespasian.exe on your PATH:

Platform	Archive
Linux x64	vespasian-0.1.5-x86_64-unknown-linux-gnu.tar.gz
Linux ARM64	vespasian-0.1.5-aarch64-unknown-linux-gnu.tar.gz
macOS Apple Silicon	vespasian-0.1.5-aarch64-apple-darwin.tar.gz
Windows x64	vespasian-0.1.5-x86_64-pc-windows-msvc.zip
Windows ARM64	vespasian-0.1.5-aarch64-pc-windows-msvc.zip

Install uv, then initialise the Python model runtime and open the guided flow:

vespasian setup
vespasian doctor
vespasian

Windows ARM64 uses an x64 Python runtime through Windows emulation. Each platform archive contains the executable, example datasets, README, licence and release documentation in a vespasian-0.1.5 directory.

Release files

• vespasian-0.1.5.crate contains the Rust application and embedded Python runtime source.
• SHA256SUMS.txt lists hashes for the five platform archives and the source crate.
• release.json records their targets, sizes, hashes and shared source commit.
• Each platform has a .json manifest with file hashes and a .sha256 file for its archive.

Read the guide for datasets, metrics, editors, health checks and Hugging Face publishing.