ArchRivet 0.4.0 alpha 1 — Claude Code early access

Windows x64 rules-only Claude Code plugin. Reviews eight scoped backend risks, names missing design facts, estimates workload demand, and previews AI-proposed fixes conditionally. No Laya model weights or GPU are required.

Early-access limitations: findings rely on supplied facts and do not verify a deployment or production capacity. A real Claude Code conversation test is pending; bundled MCP protocol and plugin manifest checks pass. Please report redacted examples through GitHub Issues.

Install after publication: download and extract the ZIP, then run claude --plugin-dir C:\path\to\archrivet. See docs/claude-launch.md for a test prompt.

ZIP SHA-256: fa7fc1e845fc14bc3373685e20d43bf7b682b3969238d76139c47bb314a6b175

ArchRivet 0.3.0a2 - checked facts and lighter MCP installation

ArchRivet now asks for confirmation when an assistant's interpretation cannot be verified against the source. A genuine quote alone no longer lets an unrelated extracted fact drive a review. Proposed values and specific questions stay visible; explicit user declarations remain scoped assumptions.

Start with the Rules MCPB or ZIP for the lighter installation: no Torch or Laya dependencies. Choose the separate Laya CPU bundle for the existing optional 322M checkpoint. Both variants use the same archrivet identity; replace the existing installation when switching. Model setup now distinguishes verified files from a loaded worker, supports cancellation/retry, and correctly handles idle restarts.

• Runtime: 0.3.0a2. Desktop bundles: 0.3.0-alpha.1.
• 268 local tests passed, plus actual MCP stdio reviews and before/after comparison for both packaged variants. Laya was tested with the existing weights.
• Eight-check warm CPU review: Rules median 33 ms, p95 41 ms; Laya median 2.39 s, p95 2.73 s across 30 samples per mode. These exclude AI-host writing time and do not establish a speedup over an identically measured previous runtime.
• Rules scored 780/780 on the public synthetic regression set; the existing optional Laya model scored 769/780. This test does not establish real-world accuracy or an advantage from adding Laya.

Connection instructions · Fact verification · Full-review performance samples · Decision baseline

This is an experimental runtime update. Model weights are unchanged; strict model-promotion gates remain unmet. Prose interpretations can require additional confirmation. Actual desktop GUI installation and the second authenticated AI-host test remain pending. Download checksums and package audits are attached; private development datasets are not included.

ArchRivet 0.3 preview — clearer reviews and updated Laya

Icon-enabled MCP bundle

The 0.2.0-alpha.2 desktop bundle adds the ArchRivet icon to the installer and MCP server metadata. The README includes a lightweight animated mark. Host support determines icon display.

Download MCPB · Download ZIP · Setup guide

This branding update retains the same 0.3.0a1 engine, dependency lock and experimental Laya weights. See branding-validation.json and BRANDING-SHA256SUMS.txt. MCP initialization, all seven tools, three rules-mode review outcomes and review comparison passed. Actual Claude Desktop GUI testing remains pending. Branding source. Original assets below remain available.

---

An optional experimental preview with clearer follow-up questions, before/after reviews, and updated Laya 322M weights.

On the same 780 synthetic final-test cases, correct decisions increased from 294/780 (37.7%) to 769/780 (98.6%). Incorrect approvals decreased from 222 to 3. Each version uses its own input format; this is a model-plus-format comparison, not production accuracy. Strict accuracy gates remain unmet. Rules are the default.

Choose a download

• Windows desktop host: .mcpb for a host supporting MCPB 0.4 with the UV runtime. Actual Claude Desktop GUI validation is still pending.
• Codex on Windows: .zip bundle (about 123 KB) plus the setup guide. Initial setup downloads Python dependencies.
• Python developers: .whl or .tar.gz for runtime 0.3.0a1.
• Optional model: archrivet-322m-experimental-v0.2.0-alpha.1.zip (about 964 MB), with complete merged weights, tokenizer, configuration, calibration and licenses. No separate base or adapter needed. Requires the 0.3 runtime and 2,048-token format.

Watch the actual Codex workflow · Full benchmark, failures and reproduction

Warm native CPU decisions had a median of 0.299 s for the preview versus 0.178 s for the old model in a small 16-case test. First loading takes much longer. This release improves tested accuracy, not demonstrated speed. Workload estimates do not certify application capacity.

The model still makes incorrect approvals, including a simple outage-cache case in the packaged test. The combined review preserves the rule contradiction and disagreement. Keep the model optional and inspect the findings. GIFs are edited walkthroughs, not live recordings or timing benchmarks.

The runtime passed 228 tests, six downloader tests, packaged rules/CPU-Laya MCP checks and an actual Codex rules-mode conversation retaining the review summary. Full test inputs and predictions are public. See validation.json and SHA256SUMS.txt. No additional training was started for this release.

ArchRivet 0.2.0a3 — clarification walkthrough

ArchRivet v0.2.0a3 adds a runnable clarification walkthrough. It shows what happens when a design mentions message IDs but leaves duplicate-protection enforcement and atomicity unspecified.

The existing retry rules return insufficient_information for missing facts, flag an explicitly unsafe mechanism, and support fully specified protection under the supplied assumptions. This release makes that behavior easier to try and explain; it does not introduce new review rules or model weights.

What is included

• An incomplete messaging-design example: examples/needs-clarification.json.
• A guide showing how answers change a verdict: docs/clarification-walkthrough.md.
• An assistant prompt for asking focused questions and reviewing confirmed answers again.
• Updated setup guidance, a Python wheel, a source archive containing examples/docs, and SHA-256 checksums.

Install

Use Python 3.11 or 3.12 in a virtual environment. The package is distributed through GitHub Releases, not PyPI.

python -m pip install --upgrade https://github.com/Ujjansh05/archrivet/releases/download/v0.2.0a3/architecture_review_plugin-0.2.0a3-py3-none-any.whl
arch-review --version

Download and extract architecture_review_plugin-0.2.0a3.tar.gz from the assets below to get the examples. From its extracted directory, run:

python -m arch_review review examples/needs-clarification.json --rules-only

Expected: retry safety is insufficient_information; acknowledgement durability and queue overload are supported under the example's declared facts. Review status is partial because model inference is skipped. Capacity remains unknown without benchmark evidence.

No GPU or model download is needed for this example. For assistant integration, follow the MCP guide. The optional complete experimental model remains a separate download.

Validation

• 216 local software tests and Ruff passed.
• A fresh Windows Python 3.12 environment installed the wheel and its core dependencies successfully; pip check passed.
• Four packaged messaging examples and the eight-check backend example returned their expected results outside the source checkout.
• Five clarification variants verified missing, unsafe and supported retry outcomes.
• Publication is gated on all six GitHub CI checks for this release commit.

These checks verify software and packaging behavior, not general model accuracy or production capacity.

Scope

Experimental alpha. Checks use supplied design facts; ArchRivet does not discover undocumented retry paths from application code. Follow-up conversations are guided by the AI host. Automatic interviewing is not implemented. Supported findings do not certify production reliability or capacity.

Original project code is MIT-licensed. Model weights and reference material keep their own licenses and attribution.

ArchRivet 0.2.0a2 — install the experimental merged model

Adds an opt-in, checksum-verified installer for the complete 322M experimental merged checkpoint.

The model includes the base weights, merged adapters, trained head, tokenizer and fitted calibration. Rules remain the default. The wheel and source archive do not bundle weights; download them separately with:

arch-review download-model --variant archrivet-experimental --destination models/archrivet-322m
arch-review configure --checkpoint models/archrivet-322m --backend laya --device cpu --context-tokens 1024

Stop an existing service before changing configuration. See the complete installation guide.

Verification: all 48 existing development/regression decisions survived merging and fresh CPU/CUDA loads. This is packaging fidelity, not improved accuracy: the candidate scored 25/36 synthetic development cases and 4/12 separate regressions. FP32 inference preserves the evaluated calibration. The final test remains untouched.

ArchRivet 0.2.0a1 — backend review alpha

ArchRivet 0.2.0a1 expands the local MCP plugin from three messaging checks to eight scoped backend decisions. An AI host drafts the architecture; ArchRivet returns findings, evidence, missing information and workload calculations.

• Adds transaction atomicity, replica consistency, cache freshness, failover fencing and dependency retry-budget checks.
• Adds schema discovery, background reviews, polling and cancellation. Existing three-check clients remain supported.
• Adds optional local document extraction with source spans, conflict tracking and visible unprocessed sections. The 128k-token intake limit is distinct from each model call's context window.
• Includes frozen diagnostic data, training-only representation variants, executable regression cases and reproducible CPU/CUDA experiments.
• Includes an opt-in Colab/QLoRA workflow; cloud training has not been run.

Use the quick start or eight-check guide. The wheel installs the service; model dependencies and weights are separate. The source archive includes scripts, examples, tests and documentation. SHA-256 checksums are attached. The project is not published on PyPI.

Validation: 200 local tests passed, Ruff and dependency checks passed, and all six GitHub CI jobs passed, covering Windows/Linux with Python 3.11/3.12, the Windows CPU checkpoint adapter, and distribution building. A separate package environment verified all eight schema definitions and the example verdicts.

The rules workflow and one actual Codex MCP host trial were tested. Local RTX 3050 fine-tuning results, failure cases, resource measurements and limitations are published in the experiment report. Best synthetic development accuracy rose from 11/36 to 25/36, but a separate executable regression packet stayed at 4/12. These results do not establish general system-design accuracy or production safety. Experimental model deltas are not included or activated. The final test set remains unevaluated; a second real AI host and independent usefulness evaluation remain pending.

Original project code is MIT-licensed. Upstream models, dependencies and cited materials retain their own terms. No model weights, credentials or local runtime files are included.

ArchRivet 322M — complete experimental model v0.1.0

---

license: apache-2.0
base_model: convaiinnovations/laya-multilingual
library_name: laya
language:

• en
  tags:
• architecture-review
• experimental
• merged-lora

---

ArchRivet 322M — experimental merged model

A complete local decision classifier for backend architecture experiments. The base
Laya weights, trained decision head, merged adapters, tokenizer, encoder configuration
and fitted calibration are included. No separate base-model or adapter download is
needed. Python runtime dependencies are installed separately.

Experimental: 25/36 correct on synthetic development cases, 4/12 on separate
executable-regression cases. This is not validated as a general system-design expert
or a production approval mechanism. It returns choices and probabilities, not generated
architecture explanations. ArchRivet's rules remain the default; using these weights
is an explicit opt-in.

Download the full model
· Plugin release
· Measured results

Install and connect your AI

Use Python 3.11 or 3.12 in a virtual environment. Activate it with
.\.venv\Scripts\Activate.ps1 on PowerShell or source .venv/bin/activate on Linux.
The project is not on PyPI; the URL below installs the GitHub wheel.

python -m venv .venv
# Activate the environment using the command above for your operating system.
python -m pip install torch==2.6.0 --index-url https://download.pytorch.org/whl/cpu
python -m pip install "architecture-review-plugin[model] @ https://github.com/Ujjansh05/archrivet/releases/download/v0.2.0a2/architecture_review_plugin-0.2.0a2-py3-none-any.whl"
arch-review download-model --variant archrivet-experimental --destination models/archrivet-322m
arch-review configure --checkpoint models/archrivet-322m --backend laya --device cpu --context-tokens 1024
arch-review start
arch-review mcp-config

If an ArchRivet service is already running, run arch-review stop before configure.
Keep the same ARCH_REVIEW_HOME across commands. Copy the generated MCP configuration
into your AI host. Windows requires start in an independent terminal before connecting
the host. The host calls get_review_schema, then review_design and get_review.
Rules, evidence and model judgments remain separate; model approvals cannot erase
explicit rule contradictions.

The installer streams the download, checks a pinned SHA256 and file size, validates
archive paths and individual files, then installs the complete checkpoint. It refuses
to overwrite an existing directory and does not change your settings. For manual
installation, download the ZIP and SHA256SUMS.txt from the model release, verify the
archive and extract it into a new folder. Point configure at the folder containing
rl_agent_config.json.

The FP32 weight file is 1,287,653,752 bytes (1.29 GB). Allow several GB for the
temporary archive, extracted model and Python dependencies. CPU inference is supported;
CUDA is optional with a compatible PyTorch installation. Validation used Windows,
Python 3.12, Torch 2.6.0 CPU and CUDA 12.4, and an RTX 3050 6 GB. Runtime memory depends
on input length and concurrency; this does not guarantee fit on every low-memory machine.

Use directly in Python

After installing and downloading as above:

from arch_review.checkpoint import configure_checkpoint_loading

configure_checkpoint_loading()  # Windows-safe loading; a no-op elsewhere.
import laya
import torch

agent = laya.load("models/archrivet-322m", device="cpu")
agent.amp_enabled = False
agent.dtype = torch.float32  # Same precision as validation and calibration.
result = agent.predict(
    "A worker commits a database write, then crashes before acknowledging the queue. "
    "The queue redelivers the message. No uniqueness constraint or deduplication exists.",
    {"retry_safety": {
        "type": "choice",
        "instructions": "Does the design prevent repeated delivery from duplicating business effects?",
        "criteria": {
            "supported": "Explicit facts satisfy the requirement.",
            "contradicted": "Explicit facts establish a violation.",
            "insufficient_information": "Required facts are missing or conflicting."
        }
    }},
    max_len=1024,
    head_max_len=256,
)
print(result["answers"])

Direct calls expose model judgments without the plugin's rules or workload calculations.
Inspect result["usage"] for truncation; do not rely on a truncated answer. Tested context
is 1,024 tokens per decision, including a 256-token question head. Larger document intake
does not enlarge this model window. Direct CUDA callers must also disable autocast as
shown above. The 0.2.0a2 plugin applies the model manifest's FP32 setting automatically.

Training, merge and limitations

• Base: convaiinnovations/laya-multilingual, revision
  1720e3e3357cfe1e281542e223f8273b0890ca34, 321,908,995 parameters.
• Candidate: backend-lora-cuda-6ep-seed29, six epochs, selected epoch six on development
  data. Rank-eight adapters, alpha 16, across 88 linear layers in all 22 encoder layers,
  plus the trained decision head.
• Training: 120 constructed training cases plus 240 representation variants. Variants
  are not new independently verified cases. Eight decisions cover retries, acknowledgements,
  queue overload, atomicity, replica consistency, cache freshness, failover and dependency
  resilience. Specialized examples are English; multilingual accuracy has not been evaluated.
• Merge: W = W_base + (16 / 8) * B @ A, exported in FP32 with ordinary Laya tensor names.
  No adapter wrappers or external encoder-weight downloads are needed.
• Calibration: fitted on 36 separate synthetic calibration cases; still unvalidated.
  Model probability is not the probability that a production system will fail.

No new training was performed for this release. Original base and adapter files were
preserved. Publishing makes the candidate accessible; it does not promote its quality.
The unchanged 4/12 separate regression accuracy and 0/4 correct unknown controls remain
limitations. Changing option order changed answers on 15/36 development cases. The final
48-case test remains unused. Maximum-context and arbitrary prose extraction are unvalidated.

Merge verification

The unmerged candidate and merged checkpoint returned identical labels on 48 existing
inputs: 36 development and 12 executable-regression cases. Fresh CPU and CUDA processes
preserved all 48 labels; maximum reported probability drift was approximately 0.0001.
These verify packaging fidelity, not additional independent accuracy. See the
machine-readable validation.

Reproduction requires the exact base, adapter and original report. Run
scripts/merge_laya_adapter.py --help from the repository. Export into a new directory,
then run --verify-only in fresh CPU and CUDA processes. Retain licenses and checksums
when redistributing. License/model-card packaging is separate from the tensor merge.

License and attribution

The model is distributed under Apache License 2.0, following
Laya's declaration.
Its mmBERT backbone is declared MIT upstream.
The archive contains LICENSE, NOTICE, upstream model cards and license text.
The plugin's separate MIT software license does not replace these model terms.
This is an ArchRivet adaptation; upstream authors do not endorse its judgments.

ArchRivet Desktop MCP — Windows preview

ArchRivet now has a downloadable desktop MCP bundle for a command-free setup in compatible hosts.

Download the Windows x64 preview

Install

1. Download the .mcpb asset below.
2. In a compatible Claude Desktop version, open Settings → Extensions → Advanced settings → Install Extension and select the file.
3. Allow automatic Python/dependency setup. Leave Laya off for rules-based reviews, or enable Enable experimental local Laya model to download the published model.
4. Ask: Use ArchRivet to check my design. Start with setup status and the review schema.

Compatibility: requires a host implementing MCPB 0.4 UV runtime, Windows x64, internet for initial setup, and several GB of available disk space. No terminal commands, user-installed Python or GPU are required in a compatible host. The ~131 KB bundle is a bootstrap package; the CPU dependencies download during setup, and optional model weights are a further ~964 MB download. This is not a standalone .exe installer.

Try a review

Use ArchRivet to review my product-price cache. Prices must never be more than five seconds old. The design can return prices up to 60 seconds old. Behavior during database failures is unspecified. Submit only my stated facts and explain the findings and missing information.

What was checked

• Official MCPB manifest validation; automatic managed-Python and dependency setup.
• The exact downloaded-archive layout in a fresh package environment, with system Python absent from PATH.
• Real MCP schema discovery and flawed, incomplete, and revised cache reviews.
• Actual CPU Laya inference using the published checkpoint.
• First-time model download, checksum verification and extraction; six download/retry tests.
• Explicit package allowlist: no private training data, scraped corpus or credentials.

Preview limitation: the actual Claude Desktop installation/conversation test remains pending. Protocol tests do not prove compatibility with every desktop host. This is a direct GitHub preview, not an approved Claude directory listing.

This reuses plugin 0.2.0a3 and model-v0.1.0. No new model was trained or promoted. The published model remains experimental (25/36 synthetic development cases; 4/12 separate regression cases). Workload calculations estimate demand, not proven production capacity. Rules and model judgments stay distinct.

Reviews run locally; your assistant's own provider and data settings still apply. The bundle includes its wrapper source, dependency lockfile, published engine wheel and licenses. The repository tag points to the existing public source revision; the new desktop wrapper is included in the .mcpb asset.

See INSTALL.md, validation.json and SHA256SUMS.txt below. Report setup issues with your OS and host version at https://github.com/Ujjansh05/archrivet/issues.

v0.1.1 — MIT-licensed experimental preview

This is the MIT-licensed experimental preview of Architecture Review Plugin. Version 0.1.1 adds the owner-selected license to the repository and packages; runtime behavior is unchanged from 0.1.0.

Included

• Local review_architecture MCP tool, loopback HTTP API and CLI.
• Retry safety, acknowledgement durability and queue-overload checks.
• Evidence references, risk consequences, missing information and workload/storage calculations.
• Shared CPU Laya worker with limited concurrency and idle unloading.
• 60 draft evidence records, 12 company subsystem studies and 100 synthetic pilot examples.
• Dataset export, baseline evaluation and specialist-training commands.

Install

Follow the quick start to clone, install CPU inference and generate your MCP configuration. The wheel below installs the Python service; model dependencies and checkpoint download are separate. The source archive includes examples, tests and documentation. SHA-256 checksums are attached.

Validation and limits

36 tests passed locally, plus a smoke check of the installed wheel. All six GitHub CI jobs passed, covering Windows/Linux with Python 3.11/3.12, the Windows CPU checkpoint adapter, and a fresh distribution build. Two real stdio MCP clients shared one local model worker. The recorded CPU benchmark used approximately 2.23 GiB peak aggregate memory and 1.36 seconds for subsequent reviews on the tested Windows machine.

The active model is unmodified base Laya. Specialist fine-tuning, independent expert review and trials with two actual AI hosts remain pending. Base-model evaluation exposed false approvals; deterministic contradictions are preserved by the service. This release does not establish production safety or supported capacity.

Read the validation report for measured results and unfinished work. No model weights, authentication tokens or local runtime files are included.

Original project code, documentation and synthetic examples are distributed under the MIT License. Upstream dependencies, model weights and cited source materials retain their own terms.

v0.1.0 — Experimental preview

Superseded by the MIT-licensed v0.1.1 preview. Use that release for licensed packages.

This is the first experimental preview of Architecture Review Plugin.

Included

• Local review_architecture MCP tool, loopback HTTP API and CLI.
• Retry safety, acknowledgement durability and queue-overload checks.
• Evidence references, risk consequences, missing information and workload/storage calculations.
• Shared CPU Laya worker with limited concurrency and idle unloading.
• 60 draft evidence records, 12 company subsystem studies and 100 synthetic pilot examples.
• Dataset export, baseline evaluation and specialist-training commands.

Install

Follow the quick start to clone, install CPU inference and generate your MCP configuration. The wheel below installs the Python service; model dependencies and checkpoint download are separate. The source archive includes examples, tests and documentation. SHA-256 checksums are attached.

Validation and limits

36 tests passed locally, plus a smoke check of the installed wheel. All six GitHub CI jobs passed, covering Windows/Linux with Python 3.11/3.12, the Windows CPU checkpoint adapter, and a fresh distribution build. Two real stdio MCP clients shared one local model worker. The recorded CPU benchmark used approximately 2.23 GiB peak aggregate memory and 1.36 seconds for subsequent reviews on the tested Windows machine.

The active model is unmodified base Laya. Specialist fine-tuning, independent expert review and trials with two actual AI hosts remain pending. Base-model evaluation exposed false approvals; deterministic contradictions are preserved by the service. This release does not establish production safety or supported capacity.

Read the validation report for measured results and unfinished work. No model weights, authentication tokens or local runtime files are included.

The owner selected MIT licensing after this initial preview; the 0.1.1 distributions include the license. Upstream dependencies and model weights retain their own terms.