invarlock 0.3.6

Edit‑agnostic robustness certificates for weight edits (InvarLock framework)

InvarLock — Edit‑agnostic robustness certificates for weight edits

In short: certify that weight edits (e.g., quantization) preserve quality. If they don’t, roll back safely.

Technical: edit‑agnostic guard pipeline (invariants → spectral → RMT → variance) producing a machine‑readable Safety Certificate.

Status: 0.3.6 (pre‑1.0). Until 1.0, minor releases may be breaking. See CLI help and the CHANGELOG for updates.

For guidance on where to ask questions, how to report bugs, and what to expect in terms of response times, see SUPPORT.md.

🚀 Quick start (no repo clone)

# Install with HF adapters
pip install "invarlock[hf]"

# Fast dev self‑cert on GPT‑2 small (tiny‑relax; downloads require explicit network)
INVARLOCK_ALLOW_NETWORK=1 INVARLOCK_DEDUP_TEXTS=1 INVARLOCK_TINY_RELAX=1 \
invarlock certify \
  --baseline gpt2 \
  --subject  gpt2 \
  --adapter auto \
  --profile dev

This produces reports/.../evaluation.cert.json with paired metrics (ppl/accuracy), structural deltas, spectral/RMT stats, variance‑estimator provenance, seeds/hashes, pairing metrics, and a policy digest.

Calibration note: tier thresholds and window sizes are piloted on GPT‑2 small, BERT base, and TinyLLaMA (see docs/assurance/09-tier-v1-calibration.md). For calibrated Balanced/Conservative certs, use the preset‑based CI/Release examples below. INVARLOCK_TINY_RELAX dev runs relax sample‑size floors and are intended only for small smoke tests (not release evidence).

Need presets or matrix scripts? Clone this repo and see Presets & Demos below.

---

📚 Docs & Guides

• Quickstart: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/quickstart.md
• Compare & Certify (BYOE): https://github.com/invarlock/invarlock/blob/main/docs/user-guide/compare-and-certify.md
• Reading a Certificate: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/reading-certificate.md
• CLI reference: https://github.com/invarlock/invarlock/blob/main/docs/reference/cli.md

Quick examples (repo presets, CPU; repo clone required for preset paths):

# Install with HF adapters
pip install "invarlock[hf]"

# Preflight a config (JSON diagnostics)
invarlock doctor --config configs/presets/causal_lm/wikitext2_512.yaml --json

# Calibrated GPT‑2 small (recommended starting point; repo preset)
INVARLOCK_ALLOW_NETWORK=1 INVARLOCK_DEDUP_TEXTS=1 \
invarlock certify \
  --baseline gpt2 \
  --subject  gpt2 \
  --adapter auto \
  --profile release \
  --preset configs/presets/causal_lm/wikitext2_512.yaml

# Tiny causal LM smoke (out‑of‑calibration, dev‑only)
INVARLOCK_ALLOW_NETWORK=1 \
invarlock certify \
  --baseline hf:sshleifer/tiny-gpt2 \
  --subject  hf:sshleifer/tiny-gpt2 \
  --profile dev

Notes:

• Presets and scripts live in this repo (configs/, scripts/) and are not shipped in wheels. Use flag‑only certify when installing from PyPI, or clone this repo to use presets and the matrix script.
• python -m invarlock works the same as invarlock.
• InvarLock runs offline by default; enable network per command with INVARLOCK_ALLOW_NETWORK=1 when fetching.

---

🔧 Installation

# Core + HF adapter
pip install "invarlock[hf]"

# GPU extras (CUDA wheels if available)
pip install "invarlock[gpu]"

# Optional edit backends
pip install "invarlock[awq,gptq]"     # AWQ/GPTQ PTQ stacks
pip install "invarlock[dev]"          # dev tooling (ruff, pytest, mkdocs)

Minimal core installs with pip install invarlock. The OSS core is edit‑agnostic (BYOE): supply baseline and subject checkpoints and run Compare & Certify. A small built‑in edit, quant_rtn, is provided for CI/quickstart demos only; optional extras (e.g., gptq, awq, gpu) are loaders/runtimes, not edit pipelines. Core installs do not pull in torch/transformers; those are only installed when you opt into extras such as "invarlock[hf]" or "invarlock[adapters]".

Run either entry point:

invarlock --help
python -m invarlock --help

Common error (missing torch on adapter-based commands):

❌ Torch is required for this command.
Install extras with: pip install "invarlock[hf]" or "invarlock[adapters]".

If you see this, install an appropriate extra (for example, pip install "invarlock[hf]") before running invarlock run or invarlock certify with HF adapters.

Network Access

• Outbound network is disabled by default for safety. Enable it explicitly (per command) when you need to download models or datasets:

INVARLOCK_ALLOW_NETWORK=1 invarlock certify \
  --baseline gpt2 \
  --subject  gpt2 \
  --adapter auto \
  --profile ci \
  --preset configs/presets/causal_lm/wikitext2_512.yaml

• Offline/air‑gapped usage: pre‑download to a cache, then run with network disabled. You can enforce offline reads with HF_DATASETS_OFFLINE=1 (and optionally set HF_HOME/HF_DATASETS_CACHE to your cache location).

See the CLI reference and datasets guide for details:

• https://github.com/invarlock/invarlock/blob/main/docs/reference/cli.md
• https://github.com/invarlock/invarlock/blob/main/docs/reference/datasets.md

Install via pipx (isolated)

# Ensure pipx uses Python 3.12+
pipx install --python python3.12 "invarlock[hf]"  # Python 3.12+ recommended

# With GPU extras (if supported on your platform)
pipx install --python python3.12 "invarlock[hf,gpu]"

Conda environment recipe

conda create -n invarlock python=3.12 -y
conda activate invarlock

# Core + HF stack
pip install "invarlock[hf]"

# Optional extras
# pip install "invarlock[gpu]"
# pip install "invarlock[awq,gptq]"

---

💻 Support Matrix

Platform	Status	Notes
Python 3.12+	✅ Required
Linux	✅ Full	Primary dev target
macOS (Intel/M-series)	✅ Full	MPS supported (default on Apple Silicon)
Windows	❌ Not supported	Use WSL2 or a Linux container if required
CUDA	✅ Recommended	For larger models
CPU	✅ Fallback	Slower but functional

Device selection: CUDA → MPS → CPU (auto). Override with torch env if needed (e.g., CUDA_VISIBLE_DEVICES).

---

🧱 What InvarLock Provides

• Runner (torch-agnostic core): prepare → preview → apply → guards → evaluate → report/rollback

• Built-in edit:

  • quant_rtn (INT8 RTN, per‑channel, clamp/group size)

• Guards (policy-tiered; “GuardChain” = ordered guard pipeline):

  1. Invariants (pre/post: shapes/finite/tying)
  2. Spectral (per-family z-caps; monitor or gate per tier)
  3. RMT (ε-band on outliers; monitor or gate per tier)
  4. Variance (VE) (predictive paired ΔlogNLL gate; tiered sidedness)

• Safety Certificate (schema v1, PM‑only): Primary Metric (ppl or accuracy) with paired statistics, structural deltas, spectral/RMT stats, VE provenance, seeds/hashes, pairing metrics, and policy digest. Canonical artifact: reports/.../evaluation.cert.json.

Scope (what InvarLock does / does not do):

• InvarLock certifies regression risk from weight edits (e.g., quantization or pruning) relative to a fixed baseline under a specific configuration.
• It focuses on paired primary metrics (ppl/accuracy) plus structural and guard telemetry (invariants, spectral, RMT, variance) for those edits.
• It does not claim to solve content‑safety problems (toxicity, bias, jailbreaks) or alignment in general, and it does not certify arbitrary training changes or new datasets.
• It is calibrated and tested on Linux/macOS environments using the HF/PyTorch stack described in the docs; native Windows is not supported.
• For the detailed assurance case and threat model, see docs/assurance/00-safety-case.md and docs/security/threat-model.md.

Minimal excerpt (redacted):

{
  "schema_version": "v1",
  "run_id": "...",
  "validation": {
    "primary_metric_acceptable": true,
    "guard_overhead_acceptable": true
  },
  "primary_metric": {
    "kind": "ppl_causal",
    "preview": 12.3,
    "final": 12.1,
    "ratio_vs_baseline": 0.98,
    "display_ci": [0.97, 0.99]
  },
  "structure": {"layers_modified": 0, "params_changed": 0},
  "spectral": {"caps_applied": 0},
  "rmt": {"stable": true},
  "auto": {"tier": "balanced"}
}

---

🛡️ Guard Order & Balanced Defaults

Canonical order: ["invariants", "spectral", "rmt", "variance", "invariants"]

Balanced profile (example)

guards:
  spectral:
    mode: monitor
    sigma_quantile: 0.95
    deadband: 0.10
    scope: all
    max_caps: 5
    max_spectral_norm: null         # disable absolute clamp; rely on calibrated κ_f
    multiple_testing: { method: bh, alpha: 0.05, m: 4 }
    family_caps: { ffn: 2.5, attn: 2.8, embed: 3.0, other: 3.0 }   # z-caps (FPR-derived)
  rmt:
    mode: monitor
    epsilon_by_family: { ffn: 0.10, attn: 0.08, embed: 0.12, other: 0.12 }
  variance:
    tap: "post mlp.c_proj (pre-residual)"
    targets: "edited_modules_only"
    discovery:
      deadband: 0.02
      min_abs_adjust: 0.012
      max_scale_step: 0.03
    gating:
      sided: "one-sided"                     # improvement-only
      min_effect_lognll: 9e-4                # pilot-derived power threshold

Conservative raises z-caps/ε/deadband/min-effect and uses two-sided VE; Aggressive relaxes accordingly.

---

🔍 For development and CI commands (pytest, mkdocs, generators), see CONTRIBUTING.md.

---

✂️ Edits & Plugins

• Quant RTN (built‑in): INT8 RTN, per‑channel, group size, percentile clamp

• Compare & Certify (BYOE, recommended): Bring your baseline + subject checkpoints and certify with InvarLock

• Plugins (optional): Adapters and guards via entry points. Adapters extend model loading/inference (e.g., GPTQ/AWQ formats); plugins do not add edit algorithms beyond RTN. List components with:

  invarlock plugins --help        # summary
  invarlock plugins guards        # guard plugins
  invarlock plugins edits         # edit plugins
  invarlock plugins adapters      # adapters and backend hints
  

---

🔁 Certification Criteria (balanced profile)

Key checks enforced by balanced policy (summary):

• Pairing invariants: preview = final counts; match=1.00, overlap=0.00 (fail-fast in CI/Release)
• PM ratio gate (ppl or accuracy): upper CI ≤ 1.10
• Drift: 0.95–1.05 (paired log-space)
• Spectral/RMT: within tier FPR/ε band
• Catastrophe rollback: automatic revert if PPL > 2.0×
• Guard overhead: a bare-vs-guarded comparison records validation.guard_overhead_acceptable=true when ≤ 1 % PPL overhead

---

🧾 Minimal Config (balanced GPT-2, CI profile)

model:
  id: "<set-your-model-id>"   # e.g., gpt2
  adapter: "hf_gpt2"
  device: "cpu"
dataset:
  provider: "wikitext2"
  split: "validation"
  seq_len: 512
  stride: 512
  preview_n: 64
  final_n: 64
  seed: 42
edit:
  # Optional: built-in quant demo. Omit for Compare & Certify/BYOE.
  name: quant_rtn
  plan:
    bitwidth: 8
    per_channel: true
    scope: attn
eval:
  metric:
    kind: ppl_causal
  loss:
    type: causal
guards:
  order: [invariants, spectral, rmt, variance, invariants]
  spectral: { mode: monitor }
  rmt: { mode: monitor }
  variance:
    tap: "post mlp.c_proj (pre-residual)"
    targets: "edited_modules_only"
    discovery: { deadband: 0.02, min_abs_adjust: 0.012, max_scale_step: 0.03 }
    gating: { sided: one-sided, min_effect_lognll: 9e-4 }
auto:
  enabled: true
  tier: balanced
  probes: 0
output:
  dir: runs
  save_model: false
  save_report: true

---

🩺 Doctor (preflight)

Run preflight checks before a run to catch misconfigurations early:

invarlock doctor --config configs/presets/causal_lm/wikitext2_512.yaml --json

Text mode emits lines prefixed with ERROR:, WARNING:, or NOTE: and stable codes like [INVARLOCK:D001]. JSON mode includes summary, policy, findings[], resolution, and format_version.

---

🏗️ Source Layout (Single Distribution)

invarlock/
├─ src/
│  ├─ invarlock/                 # core + unified namespace
│  │  ├─ core/               # runner, registry, contracts, events, ABI
│  │  ├─ cli/                # console app + command wrappers (unified import path)
│  │  ├─ adapters/           # adapter wrappers (HF GPT‑2/BERT/LLaMA)
│  │  ├─ edits/              # quant_rtn
│  │  ├─ guards/             # invariants, spectral, rmt, variance
│  │  ├─ eval/               # evaluation metrics and helpers
│  │  ├─ reporting/          # report assembly, certificate generation/validation
│  │  ├─ assurance/          # assurance surface aggregating cert helpers
│  │  ├─ plugins/            # built-in example plugins
│  │  └─ observability/      # monitoring/metrics/tracing wrappers
├─ configs/                  # presets (repo‑only; clone to use)
├─ docs/                     # user guides, reference, assurance notes
├─ scripts/                  # automation / QA helpers
└─ tests/                    # unit/integration/property tests

Note: The package exposes a single import namespace (`invarlock.*`). Presets/scripts are repo resources and not packaged in wheels.

---

📚 Documentation

• User Guide: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/getting-started.md
• Quickstart: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/quickstart.md
• Compare & Certify (BYOE): https://github.com/invarlock/invarlock/blob/main/docs/user-guide/compare-and-certify.md
• Reading a Certificate: https://github.com/invarlock/invarlock/blob/main/docs/user-guide/reading-certificate.md
• Assurance (proof notes): https://github.com/invarlock/invarlock/tree/main/docs/assurance
  • eval math, spectral FPR, RMT ε, VE gate power, determinism
• Config Schema: https://github.com/invarlock/invarlock/blob/main/docs/reference/config-schema.md
• Guard Reference: https://github.com/invarlock/invarlock/blob/main/docs/reference/guards.md

---

⚡ Quick CPU Demos (dev)

For tiny, CPU‑only demos that produce readable PASS banners in dev, enable tiny‑relax and run the matrix script (repo clone required). This mode relaxes primary‑metric token floors and is intended for smoke testing only (not release evidence):

export INVARLOCK_TINY_RELAX=1 INVARLOCK_ALLOW_NETWORK=1 INVARLOCK_DEDUP_TEXTS=1 \
       TRANSFORMERS_NO_TORCHVISION=1 TOKENIZERS_PARALLELISM=false
RUN=1 NET=1 bash scripts/run_tiny_all_matrix.sh

Add INCLUDE_MEASURED_CLS=1 to include a measured classification step (requires warmed HF caches/network).

---

🧪 Determinism & Provenance

• Seeds: {python, numpy, torch} recorded in certs
• Dataset/tokenizer hashes recorded
• Paired non-overlapping windows (fail-fast if counts mismatch or pairing < 1.0)
• Cert math checks: ppl_ratio.point == exp(mean ΔlogNLL) and CI from the same paired Δ array

---

🤝 Contributing

make dev-install     # editable + dev tools (pytest, ruff, mypy, mkdocs, etc.)
make test            # run tests
make lint            # ruff + mypy
make format          # ruff format/fix
make docs            # build docs (mkdocs)
make verify          # tests, lint, format, markdownlint

Please see CONTRIBUTING.md for guidelines and Makefile for more targets.

---

📄 License

Apache-2.0 — see LICENSE.

---

Notes

• PPL levels depend on seq_len (e.g., 768-token windows typically reduce PPL vs shorter contexts).