argus-proof

Post-training LoRA evaluation and optimisation: generate samples from a trained LoRA and score them against the curated dataset it was trained from.

Part of the Argus suite — the stage that closes the loop after argus-forge emits a training config and you train a LoRA:

argus-quarry -> argus-curator -> argus-lens -> argus-forge -> your trainer -> argus-proof
  acquire         curate/export     caption       configs        LoRA           evaluate + optimise

Status: functional end-to-end. The executing verbs (run, score, report, inspect) are live, the server triggers generation+scoring runs and serves the resulting images to the argus-studio /proof review view. Roadmap: argus-proof epic.

Install

uv pip install "argus-proof[cli]"          # CLI
uv pip install "argus-proof[cli,server]"   # + HTTP server for argus-studio
uv pip install "argus-proof[cli,score]"    # + the full scorer stack (torch, insightface, …)

CLI

# Generate a sample grid from a trained LoRA (prompts come from the curator
# export's captions; backend/engine/model dirs from the environment, see
# .env.example — PROOF_BACKEND, COMFYUI_BASE_URL, PROOF_MODELS_DIR):
argus-proof run subject.safetensors ./curated_export \
  --checkpoint sdxl_base.safetensors --seed 1 --seed 2 --seed 3 --out runs

# Score a generated run into a stored EvalReport (identity needs --references,
# a held-out image dir that must NOT overlap the training set):
argus-proof score runs/proof-l00-w00-p000 --references ./holdout

# Browse stored reports / print one run's digest:
argus-proof report
argus-proof report proof-l00-w00-p000 [--json]

# Summarise an export dir (prompt sources) or a run dir (manifest + images):
argus-proof inspect ./curated_export
argus-proof inspect runs/proof-l00-w00-p000

Serve

argus-proof serve --port 8104 --cors   # peer to lens :8100, curator :8101, quarry :8102, forge :8103
argus-proof serve --read-only          # replay/demo mode: serve stored reports, 403 live eval + writes
curl -s localhost:8104/health          # {"status":"ok","service":"argus-proof","version":"...","read_only":false}

Routes (the backend of the argus-studio /proof view):

GET  /exports                           curator export dirs available to evaluate against
GET  /models                            checkpoints + LoRAs under $PROOF_MODELS_DIR
POST /run/stream                        generate + score + store one run (NDJSON progress)
GET  /reports                           stored report digests (run browser)
GET  /report/{run_id}                   full EvalReport            PUT to store one
GET  /report/{run_id}/refined           passing subset, refined ranks first
GET  /report/{run_id}/image/{image_id}  a generated sample (ids only — no path input)
POST /report/{run_id}/hitl              apply a review; recomputes pass-rate + verdict
POST /report/{run_id}/refine            second-pass re-rank (rank: null retracts)

Concurrent reviews of the same run are serialised with a per-run file lock, so a streaming run's report write and a reviewer's HITL save can't drop updates.

Generation backend

Generation is a pluggable backend (argus_proof.backends) so swapping the engine is a config change, not a code change — the ComfyUI adapter ships first:

from pathlib import Path
from argus_proof.backends import get_backend
from argus_proof.backends.base import make_dir_resolver
from argus_proof.backends.workflow import example_template
from argus_proof.models import RunSpec, LoRASpec, SamplingParams

backend = get_backend(
    "comfyui",
    workflow_template=example_template(),          # or workflow.load_template(path)
    resolve_model=make_dir_resolver(Path("~/ComfyUI/models")),
    base_url="http://127.0.0.1:8188",
)
spec = RunSpec(
    run_id="run-1",
    base_checkpoint="sdxl_base.safetensors",
    loras=[LoRASpec(name="subject.safetensors", weight=0.8)],
    sampling=SamplingParams(sampler="dpmpp_2m", scheduler="karras", steps=30,
                            cfg=7.0, clip_skip=2, width=1024, height=1024),
    prompt="a photo of sks person",
    seeds=[1, 2, 3],                               # seed-set: one image per seed
)
result = backend.generate(spec, Path("out/run-1"))  # writes images + manifest.json

The ComfyUI adapter drives a parametric workflow template (an API-format graph with $placeholder values — $base_checkpoint, $positive, $seed, $steps, $lora / $lora_weight, $clip_skip, …), polls for completion, reads back each image's embedded PNGInfo, and emits a RunManifest that pins every checkpoint/LoRA by SHA256 so the run reconstructs exactly. See templates/comfyui_sdxl_lora.json for the shipped example.

More backends. get_backend(name, ...) selects the engine by config (PROOF_BACKEND in .env.example); scoring/report code is unchanged regardless of which produced the run, and the RunManifest records the engine + version:

• diffusers — in-process diffusers SDXL pipeline: deterministic, no external service, weights hashed from disk (pip install "argus-proof[diffusers]").
• a1111 — an AUTOMATIC1111 / SD.Next /sdapi server (checkpoint via override_settings, LoRAs via <lora:…> prompt syntax; models hashed from disk).
• remote — a hosted/cloud endpoint that speaks the proof wire (POST /generate → RunManifest + base64 images). The weights live remotely, so the service supplies the manifest and it's validated at the boundary; a bearer api_key authenticates. Point it at a self-hosted proof-gen service or a thin wrapper in front of Replicate / fal.

The a1111 / remote adapters need no extra (stdlib HTTP); all three reuse the shared manifest + transport helpers and are unit-tested with fakes.

Scoring

Generated images are scored into an EvalReport by a pluggable framework (argus_proof.scoring). Per-image ImageScorers each fill one normalised [0,1] metric (identity / clip_score / aesthetic / preference / safety); a Deduper collapses Monte-Carlo near-duplicates so a cluster counts once toward the pass rate; a DiversityScorer rewards variety. A GateConfig routes each image to auto-pass / auto-fail / needs-HITL on a weighted composite, so humans only rate the borderline band:

from argus_proof.scoring import score_run, ScoreContext
report = score_run(manifest, images, scorers=[...], deduper=..., diversity=...)
report.aggregate.pass_rate   # computed over near-dup groups, not raw frames
report.verdict.passed        # run pass/fail vs GateConfig.run_pass_rate

Concrete scorers live in argus_proof.scoring.scorers, behind the [score] extra, and are lazy-imported (each reports is_available() so the orchestrator skips it when the extra is absent). Shipped:

• dedup + diversity — PhashDeduper, PhashDiversityScorer (perceptual hash, CPU-only)
• identity — IdentityScorer (InsightFace ArcFace cosine vs a held-out reference set)
• quality / adherence — clip_score_scorer() (CLIPScore), pyiqa_scorer() (CLIP-IQA), image_reward_scorer() (ImageReward), each normalising its raw score to [0,1]
• safety — SafetyScorer (NudeNet ensemble, 1 - unsafe); set a safety hard gate to auto-fail unsafe images. safety_tail_aggregate() surfaces the any-hit/min/percentile tail. NudeNet runs as a server over HTTP (PROOF_NUDENET_URL, via argus-cortex's NudeNetClient), never in-process (it is AGPL-3.0); without a URL the scorer is skipped. "Unsafe" is the suite's exposed-nudity set (exposed buttocks, breasts, genitalia, anus), the same as curator's

from argus_proof.scoring import score_run
from argus_proof.scoring.scorers import (
    PhashDeduper, PhashDiversityScorer, IdentityScorer, clip_score_scorer, pyiqa_scorer,
)
report = score_run(
    manifest, images,
    scorers=[IdentityScorer(), clip_score_scorer(), pyiqa_scorer()],
    deduper=PhashDeduper(), diversity=PhashDiversityScorer(),
)

The quality scorers' default [0,1] normalization ranges are placeholders — calibrate lo/hi (e.g. clip_score_scorer(lo=…, hi=…)) against real generations. Heavy backends need pip install "argus-proof[score]"; remote/hosted variants build on argus_cortex.backends.RemoteBackend (point at a service by IP/port). The spine itself is dependency-free and fully tested with fakes.

HITL review & refinement

Reports are stored per-run (argus_proof.reports.ReportStore, a directory of <run_id>.json) and reviewed over the server (peer to the argus-studio /proof view):

POST /report/{run_id}/hitl     # 5-star ratings + reject reasons; recomputes pass-rate + verdict
POST /report/{run_id}/refine   # optional second pass: re-rank the passing subset 1-5 + notes

The refinement stage (argus_proof.refinement) is a finer re-rank of just the images that already passed — a separate layer (ImageScores.refinement) that never overwrites the first-pass hitl_rating/reject_reasons or the run's verdict, so both the original decision and the refined ordering are kept. refined_ranking(report) surfaces the passing subset best-first; refining an image that isn't in the passing subset is refused.

from argus_proof.refinement import RefinementRequest, RefinementImageUpdate, apply_refinement, refined_ranking

refined = apply_refinement(report, RefinementRequest(
    rater="alice", updates=[RefinementImageUpdate(image_id="img-3", rank=5, notes="cleanest hands")],
))
best_first = refined_ranking(refined)   # passing subset, refined re-ranks on top

Policy moderation (optional)

The Phase-2 safety scorer catches nudity; argus_proof.moderation extends it to a Guard-class policy taxonomy (violence / hate / self-harm / weapons / illegal) — over both the generated images and the input prompts / captions, so a toxic prompt is flagged even when its output is clean (pip install "argus-proof[moderation]", Llama Guard 3 Vision):

from argus_proof.moderation import PolicyModerator, moderate_images, moderate_texts

mod = PolicyModerator()                         # default: Llama Guard 3 (lazy, [moderation])
out = moderate_images(image_paths, mod)         # per-category tails over the outputs
inp = moderate_texts(prompt_grid_variants, mod) # ...and over the inputs
out.flagged()                                   # e.g. ["violence", "hate"], worst first
report.scorers.append(mod.provenance("output")) # version-stamp the Guard model + taxonomy

Each category gets a tail view (any-hit / max / 95th percentile — the extremes that matter, not the mean, same rule as safety), combined conservatively across an ensemble (most-unsafe detector wins). Detectors are pluggable and injectable, so the taxonomy/ensemble/tail logic is dependency-free and unit-tested; only the real Llama Guard adapter needs the extra. A reviewer's HITL flag attributes to a category via RejectReason.category. CSAM matching stays a separate policy gate (Thorn Safer / PhotoDNA), not an ML metric here.

CI acceptance gate

Turn "was this LoRA/dataset good enough?" into an automatable yes/no. argus-proof gate evaluates a scored EvalReport against declared thresholds and exits non-zero when rejected, so it drops straight into CI:

argus-proof gate eval_report.json \
  --min-pass-rate 0.75 \
  --min-pass-rate-ci-lower 0.7 \   # Wilson lower bound — a lucky 3/3 won't pass
  --min-identity 0.6 \
  --max-unsafe-rate 0.0            # exit 0 = accepted, 1 = rejected, 2 = unreadable

The pass-rate lower bound uses a Wilson score interval (argus_proof.stats, no scipy), so acceptance is statistically defensible at small N. A configured metric that wasn't measured fails its check rather than passing silently.

Cross-run stats

Per-run reports accumulate into a queryable store so "which checkpoint / LoRA weight / token wins?" is answered with evidence, not vibes (argus_proof.crossrun, pip install "argus-proof[stats]"):

from argus_proof.crossrun import CrossRunStore, run_stats, krippendorff_alpha

store = CrossRunStore("proof_stats.parquet")
store.append(run_stats(manifest, report))          # one tidy row per run (re-append updates)
for cell in store.slice_pass_rate("base_checkpoint"):
    print(cell.value, cell.pass_rate, (cell.ci_low, cell.ci_high))   # pooled pass-rate + Wilson CI

# Comparing A/B experiment arms: attribute each run to its cell, then slice by the arm.
for arm in plan.cells:                             # an ExperimentCell (see the matrix section)
    store.append(run_stats(manifest, report, step_config=arm.step_config, labels=arm.labels))
store.slice_pass_rate("step_config")               # fast vs quality
store.slice_pass_rate("label:caption_strategy")    # florence vs wd14 (an upstream factor)

alpha = krippendorff_alpha([{"alice": 5, "bob": 4}, ...])   # inter-rater reliability

Pass-rate slices carry a Wilson confidence interval, so a lucky 3/3 cell reads as far less certain than 300/400; the store is parquet, keyed by run_id + versions.

Recommendations

The gate says did it pass?; argus_proof.recommend says what to change, and where — mapping weak metrics to the suite stage that owns the fix:

argus-proof recommend eval_report.json --store proof_stats.parquet
# [lens]  unsafe outputs [unsafe_rate 0.04 vs 0.00]: filter/re-caption training data…   (safety first)
# [forge] identity didn't transfer [identity 0.41 vs 0.60]: add/curate more identity images…
# [lens]  prompt adherence low [clip_score 0.32 vs 0.50]: revisit the captioning strategy…
# [grid]  prompt adherence low [clip_score 0.32 vs 0.50]: try different prompt/token combinations…
# [checkpoint] base_checkpoint outcome varies across runs: prefer base_checkpoint='sdxl_v2'…

from argus_proof.recommend import RecommendConfig, recommend
# keep the floors in lock-step with the CI gate so the two can't disagree:
cfg = RecommendConfig.from_acceptance(thresholds)
for rec in recommend(report, config=cfg, store=cross_run_store):   # store optional
    print(rec.stage, rec.metric, rec.value, rec.action)

Safety first, then: low identity/aesthetic → forge (training), low adherence → lens + grid, low diversity → grid, borderline → refine (HITL). With a cross-run store it also surfaces the best checkpoint / LoRA weight — but only when the evidence separates a clear winner (non-overlapping CIs), never on a tie.

A/B experiment matrix

Compare LoRAs across more than one axis at once. An ExperimentMatrix (argus_proof.experiment) declares factors × levels and expands to a cell per base_checkpoint × step_config, each a full grid (LoRA checkpoint × weight × prompt × seed). Cost is aggregated across cells and estimated before launch, with a --max-gpu-hours guardrail that refuses an intractable matrix:

argus-proof experiment matrix.json --export ./curated_export --max-gpu-hours 40
# experiment exp: 4 cells, 32 runs, 96 images
#   [exp-c00sdxl-a-fast] 24 images (8 runs)
#   ...
# est. 0.2 GPU-hours @ 6s/image

from argus_proof.experiment import ExperimentMatrix, StepConfig, expand_experiment
from argus_proof.grid import read_export_prompts
from argus_proof.models import SamplingParams

matrix = ExperimentMatrix(
    base_checkpoints=["sdxl_a.safetensors", "sdxl_b.safetensors"],
    step_configs=[StepConfig(name="quality", sampling=SamplingParams(...))],
    lora_checkpoints=["e10.safetensors", "e20.safetensors"],  # epoch sweep
    lora_weights=[0.8, 1.0],
    seeds=[1, 2, 3],
    labels={"caption_strategy": "florence"},  # upstream factor, carried for cross-run slicing
)
plan = expand_experiment(matrix, read_export_prompts(export_dir), max_gpu_hours=40)
for cell in plan.cells:
    ...  # generate + score each cell.plan; cell.labels feed the cross-run store

Upstream factors (caption strategy, source-image variation) are trained into a LoRA, so proof can't vary them — it lists them in labels, which ride on every cell. Feed a cell's step_config/labels into run_stats(...) and the cross-run store compares the arms directly (slice_pass_rate("step_config"), slice_pass_rate("label:caption_strategy")). For a matrix too large to brute-force, optuna_search() (optional [opt] extra) does sample-efficient search over the same factor levels.

FiftyOne exploration (optional)

A power-user surface over a scored run, complementing the /proof HITL view. argus_proof.explore turns an EvalReport into a FiftyOne dataset — every computed field attached to its image — so you can visualise embeddings (UMAP/t-SNE) to spot mode collapse / clusters / outliers, run the uniqueness/near-dup brain, and triage by tag (pip install "argus-proof[fiftyone]"):

# open the App, then fold the tags you added back into a new report on close
argus-proof explore eval_report.json --images ./run-1/images --umap \
  --ingest reviewed.json --rater alice

from argus_proof.explore import to_fiftyone_dataset, compute_visualization, ingest_from_dataset

ds = to_fiftyone_dataset(report, {"img-1": "run-1/images/img-1.png", ...})
compute_visualization(ds)                       # UMAP embedding viz (needs umap-learn)
report = ingest_from_dataset(ds, report)         # round-trip: fold tags back as ratings/reasons

The round-trip is tag-driven: in the App you add rating:<1-5> / reject:<code> tags (these are the input channel — exported samples carry the scores as fields and only a verdict/refined display tag, so a round-trip never re-ingests the run's own auto-computed rejects). Ingest is authoritative — an image's tags are its full decision, so a rating:5 with no reject: tag un-rejects it — and folds through the same apply_hitl path a review uses (the verdict recomputes identically; pass the original gate to keep non-default thresholds). The mapping (sample_fields/sample_tags/ingest_tags) is dependency-free and unit-tested; only the dataset/brain/App adapters need the extra, and explore.is_available() guards them.

Develop

make install   # venv + editable install with the "dev,server,cli" extras
make test
make lint

CI / Release

• CI runs via the shared argus-ci reusable workflow.
• Release publishes to PyPI (OIDC trusted publishing) and GHCR on v* tags.
• Once PyPI has the version, a GitHub Release is created with that version's CHANGELOG.md section as its notes. Add the section before you tag; the job fails without it.
• Versioning is derived from git tags via hatch-vcs — tag vX.Y.Z to cut a release.

This repo was scaffolded from argus-pkg-template. Run copier update to pull template changes (CI, release, tooling).