InvarLock GGUF runtime add-in

This optional distribution connects authenticated GGUF artifacts to InvarLock's public runtime-provider ABI through a pinned local llama.cpp executable. The InvarLock core package remains responsible for portable GGUF identity and independent evidence verification.

Install and check discovery

PYTHON=/path/to/venv/bin/python
"$PYTHON" -m pip install invarlock-runtime-gguf
"$PYTHON" -m invarlock_addins.gguf.conformance

The conformance command returns JSON with "ok":true, provider llama_cpp, and the ABI version accepted by the installed core package. Installing the wheel also registers llama_cpp in the invarlock.runtime_providers entry-point group; no core source change is required.

The install above is the standalone conformance path. Maintained qualification does not treat first-party packages already installed in that environment as the candidate. CANDIDATE_WHEEL_MANIFEST must authenticate the exact core and GGUF wheels built from the qualified source commit. The driver extracts those wheels into a private candidate site and loads them before any first-party code visible to PYTHON. PYTHON supplies only their third-party dependencies; QUALIFICATION_DRIVER_PYTHON launches the standard-library qualification driver. Using one isolated dependency environment for both variables is the simplest configuration.

After make dist-check, create the manifest with the maintained no-clobber helper. Its destination parent must already exist, and the destination must be new:

install -d -m 700 qualification
python scripts/qualification_candidate_wheels.py \
  --wheel dist/invarlock-*.whl \
  --wheel dist/addins/invarlock_runtime_gguf-*.whl \
  --output "$PWD/qualification/gguf-candidate-wheels.json"
export CANDIDATE_WHEEL_MANIFEST="$PWD/qualification/gguf-candidate-wheels.json"

The helper records absolute wheel paths and SHA-256 digests. The qualification driver additionally requires unique, version-aligned maintained distributions whose contents match the authenticated source archive. Recreate the manifest after any wheel changes, and pass the same manifest to canary, preflight, and evidence qualification.

Build the runtime image

Build from the repository root. The llama.cpp build stage deliberately has no mutable package-index default: select and pin a Debian snapshot timestamp in YYYYMMDDTHHMMSSZ form, then record it with the resulting image digest.

SOURCE_DATE_EPOCH="$(git show -s --format=%ct HEAD)"
LLAMA_CPP_APT_SNAPSHOT=PINNED_YYYYMMDDTHHMMSSZ
SOURCE_COMMIT="$(git rev-parse HEAD)"
SOURCE_BUNDLE=/absolute/path/to/invarlock-source.tar
SOURCE_BUNDLE_SHA256=sha256:PINNED_SOURCE_BUNDLE_DIGEST

make -C addins/gguf build \
  IMAGE=invarlock-gguf:candidate \
  SOURCE_COMMIT="$SOURCE_COMMIT" \
  SOURCE_BUNDLE="$SOURCE_BUNDLE" \
  SOURCE_BUNDLE_SHA256="$SOURCE_BUNDLE_SHA256" \
  SOURCE_DATE_EPOCH="$SOURCE_DATE_EPOCH" \
  LLAMA_CPP_APT_SNAPSHOT="$LLAMA_CPP_APT_SNAPSHOT" \
  BUILD_STATEMENT=/absolute/path/to/gguf-build.json

The add-in Makefile groups its build, import smoke, and complete evaluate-to-verify qualification targets:

make -C addins/gguf build \
  SOURCE_COMMIT=0123456789abcdef0123456789abcdef01234567 \
  SOURCE_BUNDLE=/absolute/path/to/invarlock-source.tar \
  SOURCE_BUNDLE_SHA256=sha256:... \
  LLAMA_CPP_APT_SNAPSHOT=PINNED_YYYYMMDDTHHMMSSZ
make -C addins/gguf smoke

Resolve the candidate to an immutable registry digest before creating strict evidence. The local tag is only a build handle.

The image applies the tracked runtime/llama-completion-user-output.patch to the pinned b10015 source. Normal EOG termination remains enabled, while llama.cpp's human-readable [end of text] console marker is kept out of the generated-text stream. The provider rejects that marker if an unpatched executable emits it; it never guesses whether marker-shaped bytes were backend control or model-authored text.

Derive request settings

Do not hand-copy artifact or backend digests. Derive the complete runtime spec from the exact GGUF file, executable, and source archive that will be used. The inspection itself executes native code, so this Python must run inside the digest-pinned, network-disabled image built from runtime/Dockerfile, not in an ordinary host process:

from pathlib import Path
import json

from invarlock_addins.gguf.provider import LlamaCppProvider
from invarlock_addins.gguf.session import LlamaCppRuntimeBindings

bindings = LlamaCppRuntimeBindings(
    gguf_path=Path("/inputs/model.gguf"),
    executable_path=Path("/opt/llama.cpp/llama-completion"),
    source_archive_path=Path(
        "/opt/llama.cpp/source/llama.cpp-b10015.tar.gz"
    ),
)
spec = LlamaCppProvider().inspect_runtime_spec(
    bindings,
    seed=0,
    context_length=2048,
    batch_size=1,
    cpu_threads=16,
    prompt_batch_size=512,
    prompt_microbatch_size=512,
    max_output_tokens=64,
    timeout_seconds=300,
)
print(
    json.dumps(
        {"model_id": spec.model_id, "settings": dict(spec.settings)},
        indent=2,
        sort_keys=True,
    )
)

After resolving the built image to an immutable digest, the inspection pattern is:

IMAGE=registry.example/invarlock-gguf@sha256:PINNED_IMAGE_DIGEST
DIGEST=sha256:PINNED_IMAGE_DIGEST

docker run --rm --network none \
  --env INVARLOCK_CONTAINER_EXECUTION=1 \
  --env INVARLOCK_RUNTIME_IMAGE="$IMAGE" \
  --env INVARLOCK_RUNTIME_IMAGE_DIGEST="$DIGEST" \
  --mount type=bind,src="$PWD/inputs",dst=/inputs,readonly \
  --entrypoint python "$IMAGE" /inputs/inspect.py

The Makefile keeps IMAGE as the convenient build and smoke handle. Its qualify-* targets use QUALIFICATION_IMAGE, which defaults to IMAGE_DIGEST so the mutable local tag never enters strict qualification. Set QUALIFICATION_IMAGE explicitly to either the exact local sha256:... config ID or a portable repository@sha256:... reference.

Use spec.model_id and spec.settings for a request side whose runtime.provider is llama_cpp. Evaluation requires the same files as root-confined runtime resources and the same authenticated container boundary. Local paths are execution inputs and are not published in canonical evidence.

Qualify one concrete runtime

The conformance command checks package discovery and ABI shape; it does not load a GGUF model. The runtime canary is one real, minimal InvarLock transaction over the inspected model and backend:

1. retain the printed model_id and settings as pinned request input;
2. create a one-record canonical schedule and an exact-match policy;
3. evaluate baseline and subject in independently pinned side workers, with the resource variables documented in the main runtime-provider guide;
4. let the host validate both side results and sign the evidence bundle without mounting the signing key into either worker;
5. retain the reviewed preflight request_digest, then verify the resulting pack with that digest and independently recorded artifact, schedule, runtime-image, and evidence-signer anchors; and
6. require both the evaluation and strict verification to finish successfully.

With those pinned request and expected-anchor files in place, the qualification uses the same public transactions as the core package:

invarlock evaluate request.yaml \
  --signing-key evidence-signer.pem \
  --preflight --json
invarlock evaluate request.yaml --signing-key evidence-signer.pem
invarlock verify evidence/ \
  --trust-profile trust/trust-inputs.json \
  --receipt verification.receipt.json

Place the reviewed preflight request_digest in the trust profile's anchors object. GGUF verification fails closed without it and emits a v2 signed receipt that binds the same request identity.

Bootstrap that transaction once for the exact image with make -C addins/gguf qualify-canary. Retain its EVIDENCE, strictly verified RECEIPT, and verifier-owned TRUST_PROFILE. For every later request using the same image digest, pass those paths to qualify-preflight and qualify-evidence as CANARY_EVIDENCE, CANARY_RECEIPT, and CANARY_TRUST_PROFILE. A different image digest requires a new signed canary. Set QUALIFICATION_DEVICE, QUALIFICATION_CPUS, QUALIFICATION_MEMORY_MIB, and QUALIFICATION_USER for the bounded worker environment and keep them unchanged through canary, preflight, and evidence. For GGUF, the device is normally cpu.

With the canary request, trust profile, source bundle, and output parent already prepared, the bootstrap command is:

INVARLOCK_GGUF_RESOURCE_ROOT="$PWD/gguf-runtime" \
INVARLOCK_GGUF_BACKEND_EXECUTABLE=bin/llama-completion \
INVARLOCK_GGUF_BACKEND_SOURCE=source/llama.cpp-source.tar \
make -C addins/gguf qualify-canary \
  REQUEST="$PWD/canary/request.yaml" \
  SIGNING_KEY="$PWD/private/evidence.key" \
  QUALIFICATION_IMAGE="$IMAGE" IMAGE_DIGEST="$DIGEST" \
  EVIDENCE="$PWD/canary/evidence" \
  TRUST_PROFILE="$PWD/canary/trust-inputs.json" \
  RECEIPT="$PWD/canary/verification-receipt.json" \
  SUMMARY="$PWD/canary/qualification-summary.json" \
  CANDIDATE_WHEEL_MANIFEST="$CANDIDATE_WHEEL_MANIFEST" \
  QUALIFICATION_DEVICE=cpu \
  SOURCE_COMMIT="$SOURCE_COMMIT" SOURCE_BUNDLE="$SOURCE_BUNDLE" \
  SOURCE_BUNDLE_SHA256="$SOURCE_BUNDLE_SHA256"

Run make -C addins/gguf qualify-preflight for each target request. It reverifies the signed canary and performs execution-free checks without loading the target model. After it succeeds, make -C addins/gguf qualify-evidence evaluates and verifies with REQUEST, SIGNING_KEY, QUALIFICATION_IMAGE, IMAGE_DIGEST, EVIDENCE, TRUST_PROFILE, RECEIPT, SUMMARY, CANDIDATE_WHEEL_MANIFEST, the three CANARY_* inputs, SOURCE_COMMIT, and the actual SOURCE_BUNDLE plus SOURCE_BUNDLE_SHA256. The source bundle must be the Git archive for that commit, its execution files must match the checkout, and the image must carry matching source labels. The verifier-owned trust profile contains the independently sourced policy, artifact, schedule, runtime, signer, verifier, and scorer-authorization inputs; the target never derives them from the pack. Qualification verifies the exact evidence path, normalized request, signed receipt, and pack identity before writing the private summary.

Set INVARLOCK_GGUF_RESOURCE_ROOT, INVARLOCK_GGUF_BACKEND_EXECUTABLE, and INVARLOCK_GGUF_BACKEND_SOURCE exactly as shown in the main runtime-provider guide. REPORT is optional. Receipt, report, and summary destinations must be fresh, distinct, outside the evidence path, and beneath existing non-symlinked directories.

Use a model and prompt whose expected output is fixed before qualification. Changing the GGUF file, llama.cpp executable, source archive, image digest, CPU identity, prompt, or decoding settings invalidates the qualification and requires a new transaction. Qualification therefore binds the caller's actual model, prompt, artifact, backend, and CPU combination. The signed canary catches image-level failure before fan-out; it does not prove that a different GGUF artifact loads, fits memory, or completes successfully.

Public Python surface

The provider package intentionally keeps its public surface small:

• LlamaCppProvider implements runtime-provider ABI 1 and derives the complete ModelRuntimeSpec through inspect_runtime_spec;
• LlamaCppRuntimeBindings names the authenticated GGUF, executable, and source-archive inputs used for inspection and execution;
• LlamaCppBackendInspection is the typed result of backend inspection; and
• LlamaCppExecutionError reports a native execution failure that could not produce a scoring observation.

The remaining session helpers are implementation details of the first-party add-in. Applications should enter through the provider ABI rather than invoke LlamaCppSession directly.