flamehaven-veritas 3.3.0

VERITAS — AI Critique Experimental Report Analysis Framework

VERITAS v3.2.0

AI Critique Experimental Report Analysis Framework

A sovereignty-grade experimental report critique engine.
Implements the VERITAS v3.2 protocol as a fully executable Python package + REST API + CLI.

---

What It Does

Accepts a raw experimental report (text, PDF, DOCX, MD) and produces a structured critique through a 7-phase pipeline, enriched with LOGOS reasoning, HSTA scoring, bibliography analysis, and reproducibility assessment.

Phase	Name	Weight
PRECHECK	Artifact Sufficiency Gate	—
STEP 0	Experiment Classification	—
STEP 1	Claim Integrity	40%
STEP 2	Traceability Audit	30%
STEP 3	Series Continuity	20%
STEP 4	Publication Readiness	10%
STEP 5	Priority Fix Synthesis	—

Output enrichment layers:

Engine	Output	Description
LOGOS IRF-Calc 6D	irf_scores	M/A/D/I/F/P reasoning quality dimensions
BioMedical-Paper-Harvester HSTA	hsta_scores	N/C/T/R bibliometric quality
BibliographyAnalyzer	bibliography_stats	Reference count, formats, year range, quality score
ReproducibilityChecklistExtractor	reproducibility_checklist	8-criterion ARRIVE/CONSORT assessment

---

Workflow

flowchart TD
    A([Document Input<br/>PDF / DOCX / MD / TXT]) --> B[PRECHECK<br/>Artifact Sufficiency Gate]

    B -->|BLOCKED| Z([Halt — insufficient material])
    B -->|FULL / PARTIAL / LIMITED| C[STEP 0<br/>Experiment Classification]

    C --> D[STEP 1<br/>Claim Integrity<br/>40%]
    D --> E[STEP 2<br/>Traceability Audit<br/>30%]
    E --> F[STEP 3<br/>Series Continuity<br/>20%]
    F --> G[STEP 4<br/>Publication Readiness<br/>10%]
    G --> H[STEP 5<br/>Priority Fix Synthesis]

    H --> I([Enrichment Layer])

    I --> J[LOGOS IRF-Calc 6D<br/>Reasoning Quality]
    I --> K[HSTA 4D<br/>Bibliometric Score]
    I --> L[BibliographyAnalyzer<br/>Reference Stats]
    I --> M[ReproducibilityChecklist<br/>8-criterion ARRIVE/CONSORT]

    J & K & L & M --> N[Omega Score<br/>0.0 – 1.0]

    N --> O{Output Format}
    O --> P[Markdown<br/>CLI / Agent]
    O --> Q[DOCX<br/>A4 Professional]
    O --> R[PDF<br/>A4 Print-ready]
    O --> S[LaTeX / TEX<br/>XeLaTeX]
    O --> T[MICA JSON<br/>Skill / Agent Pipeline]

---

Quick Start

pip install flamehaven-veritas

Python API

from veritas import SciExpCritiqueEngine

engine = SciExpCritiqueEngine()
report = engine.critique(report_text)

print(report.precheck.line1)                    # PRECHECK MODE: FULL
print(report.omega_score)                       # 0.8571
print(report.irf_scores.composite)             # LOGOS IRF composite
print(report.bibliography_stats.quality_score) # 0.74

CLI

# Critique from file (output to terminal as Markdown)
veritas critique path/to/report.pdf

# Critique and save formatted report
veritas critique report.pdf --format docx --output report_critique.docx
veritas critique report.pdf --format pdf  --output report_critique.pdf
veritas critique report.pdf --format tex  --output report_critique.tex
veritas critique report.pdf --format md   --output report_critique.md

# Use KU Research Report template
veritas critique report.pdf --template ku --format docx

# Run PRECHECK gate only
veritas precheck report.pdf

# MICA Playbook mode — structured JSON for agent/skill pipelines
veritas critique report.pdf --mica

# Multi-round critique with delta Omega tracking (v2.3+)
veritas critique report.pdf --round 2 --prev report_r1.json

# Batch processing (v2.4+)
veritas batch "*.pdf" --format md --jobs 4 --output-dir results/

# Session memory (v2.5+)
veritas session start
veritas session show

# CR-EP governance (v2.5+)
veritas govern init
veritas govern status

# Peer-review simulation with 3 personas (v3.2+)
veritas review-sim report.pdf --reviewers 3
veritas review-sim report.pdf --reviewers 3 --format md --output sim_result.md

REST API

# Start the server
uvicorn veritas.api.app:app --reload --port 8400

# Submit text
curl -X POST http://localhost:8400/api/v1/critique/text \
  -H "Content-Type: application/json" \
  -d '{"report_text": "...", "template": "bmj", "round_number": 1}'

# Upload a document
curl -X POST http://localhost:8400/api/v1/critique/upload \
  -F "file=@report.pdf" -F "template=bmj"

# Download formatted report
curl -X POST http://localhost:8400/api/v1/critique/download \
  -F "file=@report.pdf" -F "format=docx" -o critique.docx

---

Output Formats

Format	Flag	Description
Markdown	--format md	Structured .md with tables (low token cost)
DOCX	--format docx	A4 professional report (python-docx)
PDF	--format pdf	A4 print-ready (ReportLab)
LaTeX	--format tex	Standalone .tex (XeLaTeX-compatible, optional compile_pdf)

All outputs use either the BMJ Scientific Editing template or the KU Research Report template (--template bmj|ku).

---

API Endpoints

Method	Path	Description
POST	/api/v1/critique/text	Full critique pipeline (JSON body)
POST	/api/v1/critique/upload	Full critique pipeline (file upload)
POST	/api/v1/critique/download	Upload file, receive formatted report
POST	/api/v1/precheck	PRECHECK gate only
POST	/api/v1/classify	STEP 0 classification only
POST	/api/v1/review-sim	Peer-review simulation (v3.2+)
GET	/health	Liveness check
GET	/version	Package version

See docs/api_reference.md for full schema.

---

Enrichment Engines

LOGOS IRF-Calc 6D

Six-dimensional reasoning quality score computed over the critique text:

Dimension	Key	Meaning
Methodic Doubt	M	Systematic uncertainty articulation
Axiom / Hypothesis	A	Central claim falsifiability
Deduction	D	Logical step validity
Induction	I	Evidence generalization quality
Falsification	F	Testability and counter-evidence exposure
Paradigm	P	Framework consistency
Composite	—	Mean of M+A+D+I+F+P; threshold ≥ 0.78 = PASS

HSTA 4D (BioMedical-Paper-Harvester)

Four-dimensional bibliometric score:

Dimension	Key	Meaning
Novelty	N	Unique technical term density
Consistency	C	Contradiction marker absence
Temporality	T	Version / date marker presence
Reproducibility	R	Method detail completeness
Composite	—	Arithmetic mean (N+C+T+R)/4

Bibliography Analysis

Extracted automatically from the reference section of the submitted document:

• Total reference count and format detection (Vancouver / APA / Harvard)
• Year range (oldest → newest)
• Self-citation detection
• Quality score: 0.0–1.0 composite (recency 50% + breadth 50%, −10% if self-cites detected)

Reproducibility Checklist

8-criterion assessment derived from ARRIVE 2.0 / CONSORT 2010 / STROBE / TOP Guidelines:

Code	Criterion
DATA	Open data availability statement
CODE	Code / software availability
PREREG	Pre-registration declaration
POWER	Statistical power / sample size justification
STATS	Statistics description (test, software, version)
BLIND	Blinding / randomization procedure
EXCL	Exclusion criteria stated
CONF	Conflict of interest declaration

---

PRECHECK Modes

FULL     — All artifacts present. Execute STEP 0 through STEP 5 normally.
PARTIAL  — Primary claim evaluable; secondary artifacts missing. Proceed, mark gaps.
LIMITED  — Primary claim partially evaluable. Constrained execution.
BLOCKED  — Insufficient material. Critique halted after PRECHECK.

---

Traceability Classes

The engine uses exactly three traceability terms (no weaker substitutes):

Class	Meaning
traceable	Fully anchored to a measured artifact
partially traceable	Some anchoring present; incomplete
not traceable	No artifact anchor found

---

Evidence Precedence

Conflicting artifacts are resolved by rank:

1. Measured artifact / raw result file
2. Hash manifest / trace log / deviation log
3. Inline figure or table
4. Narrative interpretation
5. Cross-cycle comparison prose

---

MICA Playbook Mode

The CLI supports MICA (Memory Invocation & Context Archive) structured output for agent / skill pipeline integration:

veritas critique report.pdf --mica

Returns a machine-readable JSON payload suitable for direct consumption by AI agents, orchestrators, or downstream skills — without token overhead of formatted prose.

Load the full playbook at session start:

veritas playbook   # prints memory/playbook.md to stdout

---

Peer-Review Simulation (v3.2+)

Simulate a 3-member editorial panel, each applying a different calibration stance:

Persona	CalibrationGate	Bias
strict	Omega ≥ 0.85	Conservative; penalises M/D/F deficits × 1.4
balanced	Omega ≥ 0.78	Neutral; uniform weighting across 6 IRF dimensions
lenient	Omega ≥ 0.70	Liberal; M/D/F penalties reduced to × 0.85

Algorithm:

1. Run the full SciExpCritiqueEngine once → base IRF-6D scores
2. Per persona: apply weighted calibrate_omega(irf, weights) → persona Omega
3. CrossValidator.check_consensus() — consensus reached when spread ≤ 0.30
4. If consensus_omega < 0.60 → DR3Protocol.resolve() applies 0.90 penalty factor
5. Final recommendation: ACCEPT ≥ 0.78 / REVISE ≥ 0.60 / REJECT < 0.60

# CLI — outputs per-reviewer + consensus table + final recommendation
veritas review-sim report.pdf
veritas review-sim report.pdf --reviewers 3 --format md --output peer_review.md

# REST API
curl -X POST http://localhost:8400/api/v1/review-sim \
  -H "Content-Type: application/json" \
  -d '{"report_text": "...", "num_reviewers": 3}'

---

Development

git clone https://github.com/flamehaven01/Flamehaven-Veritas.git
cd Flamehaven-Veritas
pip install -e ".[dev]"

pytest                          # full suite + 80% coverage gate
ruff check src tests            # lint
mypy src                        # type check

---

CI/CD Pipeline

flowchart LR
    A([git push / PR]) --> B{Trigger}

    B -->|push to main<br/>or develop| C[CI — Python 3.10 / 3.11 / 3.12]
    B -->|push tag<br/>v*.*.*| H[Release]

    subgraph CI [CI Jobs — parallel matrix]
        direction TB
        C --> D[🔍 ruff lint]
        C --> E[🎨 ruff format check]
        C --> F[🔬 mypy type check]
        C --> G[🧪 pytest + coverage ≥ 80%]
    end

    D & E & F & G --> Z([✅ CI Green])

    subgraph Release [Release Pipeline]
        direction TB
        H --> I[python -m build]
        I --> J[PyPI Trusted Publisher]
        J --> K[GitHub Release<br/>auto-generated notes]
    end

---

Architecture

See docs/architecture.md.

---

Roadmap

Version	Target	Features
v2.2 ✅	2026 Q2	LOGOS IRF-6D, HSTA 4D, BibliographyAnalyzer, ReproducibilityChecklist, LaTeX output, MICA Playbook
v2.2.1 ✅	2026 Q2	SPAR framework optional import fallback, CI green (159 tests, mypy 0 errors), version string fix
v2.3 ✅	2026 Q2	Multi-round iterative critique (--round N), delta Omega tracking, DriftEngine (JSD/L2)
v2.4 ✅	2026 Q2	Batch processing (veritas batch *.pdf), parallel engine execution, JSON summary index
v2.5 ✅	2026 Q2	MICA persistent session memory, CR-EP governance, BM25+RRF RAG, auto-template selection
v3.2 ✅	2026 Q2	Peer-review simulation (veritas review-sim), 3-persona consensus, DR3 conflict resolution, tabbed Web UI

---

Acknowledgements

VERITAS is built on top of the Flamehaven Sovereign Stack and draws from the following open research frameworks:

Component	Reference
BMJ Scientific Editing Report	BMJ Author Services — Medical Scientific Editing Report template
KU Research Report Template	University of Kuala Lumpur — Research Report Writing Template
ARRIVE 2.0	Percie du Sert et al. (2020) — Animal Research: Reporting of In Vivo Experiments
CONSORT 2010	Schulz et al. (2010) — Consolidated Standards of Reporting Trials
STROBE	von Elm et al. (2007) — Strengthening the Reporting of Observational Studies in Epidemiology
TOP Guidelines	Nosek et al. (2015) — Transparency and Openness Promotion Guidelines
IRF-Calc 6D	Flamehaven LOGOS Engine — internal reasoning quality metric
HSTA 4D	Flamehaven BioMedical-Paper-Harvester — bibliometric scoring framework
MICA v0.2.3	Flamehaven MICA — Memory Invocation & Context Archive for AI agents

---

License

MIT © 2026 Flamehaven