Skip to content
docxologyPublic

About

Open curriculum on agentic intelligence, governance, and AI safety — organized in parts with Mermaid and Python-rendered figures, sourced from ODNI, CIA, NIST and NATO publications, built by a Python/uv pipeline into a PDF with claim-calibration and reference-integrity audits.

Topics

Resources

Contributing

Security policy

Stars

5 stars

Watchers

0 watching

Forks

Repository files navigation

AGEINT — Agentic Intelligence Curriculum

ci.yml workflow status on main Digital Object Identifier (DOI), concept record: 10.5281/zenodo.20732274 License: Code Apache-2.0, text CC BY 4.0 Requires Python 3.11 or newer

AGEINT Synthetic Tradecraft System Atlas. Bounded agentic assistance (retrieve, reason, tool, memory, handoff) feeds a central synthetic analytic tradecraft core, surrounded by collection and assurance domains such as HUMINT, SIGINT, OSINT, GEOINT, FININT, cyber, cognitive security and governance. Outputs pass a human review gate to products, halts, citations, figures, PDF, metadata and red-team checks, all resting on a source spine and evidence floor.

AGEINT is a modular curriculum for Agentic Intelligence: atlas, library, course, textbook, cookbook, and playbook. It converts a source guide into a semantic manuscript, registry-backed figures, generated BibTeX, and auditable runtime variables. Everything here builds from a standalone checkout of this repository; the one exception is the combined PDF (AGEINT.pdf), which is rendered by a sibling template repository's pandoc/xelatex pipeline rather than by a script in this repo.

The project is deliberately evidence-bounded. It can teach intelligence tradecraft, IC-cycle governance, OSINT/GEOINT integrity, FININT/economic security, declassified intelligence history, agentic-AI governance, cognitive security, counterintelligence, cyber defense, and ICS/OT readiness, but all examples stay educational, synthetic, accountable, defensive, and bounded by legal and human oversight constraints.

Citation and DOI

Friedman, Daniel Ari. AGEINT: Agentic Intelligence — A Modular Atlas, Library, Course, Textbook, Cookbook, and Playbook for Agentic Intelligence (Edition 0.1, 2026). Active Inference Institute. https://github.com/docxology/AGEINT. https://doi.org/10.5281/zenodo.20732274.

The DOI and repository are baked into the rendered title page and citation block of output/pdf/AGEINT_combined.pdf.

Source

  • Runtime curriculum source: data/curriculum/
  • Optional historical guide filename: SIST-Guide-TOC-and-Bibliography-v2.md is recognized when present but is not required for normal builds
  • Source templates: docs/manuscript/templates/*.md
  • Resolved manuscript: output/manuscript/
  • Figure registry and assets: output/figures/figure_registry.json
  • Bibliography: generated to docs/manuscript/references-*.bib and output/manuscript/references-*.bib
  • Scholarship anchors: data/research_anchors/, src/intelligence_content/, and src/manuscript_variables/

Current generated scope

  • Parts: 16
  • Chapters: 51
  • Methods appendices: 9
  • AGEINT patterns: 20
  • Parsed references: 312
  • Curated official/scholarly/professional research anchors: 462
  • Source-quality support anchors: 10
  • Registered figures: 177 (115 Mermaid, 52 Python, 4 historical, 6 AI-generated; square-normalized canvases; current registry has no placeholder plates and every row carries caption, alt text, long-description metadata, PNG metadata, and visual-semantics fields)
  • Non-numbered generated visuals: 1 deterministic Python-rendered title-page cover at output/figures/cover/ageint-cover-synthesis.png and 1 front-matter Evidence Transit Map at output/figures/frontmatter/ageint-evidence-transit-map.png (both live outside the numbered registry)

These counts are rebuild-time measurements: verify with python3 -c "import json;print(json.load(open('output/data/curriculum_outline.json'))['stats'])" or re-run scripts/build_curriculum.py; see TODO.md Verified State for the last measured pass.

Current state and next actions

  • Backlog (canonical, forward-only): TODO.md — its Open section holds the next useful work; tasks.yaml is the taskboard source of truth.
  • Status: output/reports/publication_readiness.json (ok field) and output/reports/current_artifact_evidence.json are the generated status surfaces; re-run the audits in Commands below to refresh them.

Commands

uv run python scripts/build_curriculum.py
AGEINT_REQUIRE_RENDERED_FIGURES=1 uv run python scripts/build_curriculum.py  # strict Mermaid PNGs
uv run python scripts/generate_figures.py
uv run pytest tests/ --cov=src --cov-fail-under=90
uv run python scripts/audit_artifact_evidence.py --write --format markdown
uv run python scripts/audit_scholarship_quality.py --write --format markdown
uv run python scripts/audit_source_metadata.py --write --format markdown
uv run python scripts/audit_agency_source_coverage.py --write --format markdown
uv run python scripts/audit_claim_calibration.py --write --format markdown
uv run python scripts/audit_reference_quality.py --write --format markdown

# Coverage floor: 90% for `src/`; use the test command above for the current measured result.
# Stage 02 pipeline analysis runs build_curriculum.py only (see docs/manuscript/config.yaml analysis.scripts).

# From the sibling template repo root (AGEINT linked under projects/working/AGEINT):
uv run python -m infrastructure.validation.cli markdown projects/working/AGEINT/output/manuscript --repo-root .
uv run python -m infrastructure.validation.cli prerender projects/working/AGEINT/output/manuscript --repo-root .

# PDF render (from the sibling template repo root, AGEINT linked under
# projects/working/AGEINT): re-renders the combined PDF and copies it back
# into this repo — the render/copy stages this repo does not itself carry.
uv run python scripts/maintenance/rerender_working_pdfs.py --project AGEINT

Use --regenerate-source-template-library only when intentionally resetting the neutral source templates from src/manuscript_templates.py.

Architecture

The source manuscript stays small and neutral. src/curriculum.py composes the sharded curriculum, src/build_pipeline.py orchestrates the build, src/manuscript_manifest/ creates semantic output paths and generated section contexts, src/manuscript_variables/ writes runtime variables and BibTeX, src/intelligence_content/ owns verified research profiles, deterministic agency source packs, and three-tier topic lesson frames (keyword → category → synthesis), src/source_metadata.py audits explicit lane/tier metadata for curated and support anchors, and src/figures/ renders all registry-backed figures. Smoke builds may fall back to deterministic text plates when Mermaid/Chrome is unavailable, but the checked-in registry should remain free of placeholders; set AGEINT_REQUIRE_RENDERED_FIGURES=1 to fail instead of falling back. Generated chapter architecture lives in src/manuscript_manifest/: source-guide H1 chapter titles stay stable, while repeated governance, rights, assurance, practice, assessment, capstone, refresh, and source-map material is composed into reader-facing sections at build time. src/scholarship_quality.py audits generated claim-bearing sections for uncited or thin support and flags single-source-family sections as review warnings; src/source_metadata.py audits the 462 curated anchors plus 10 source-quality support anchors so blank lane/tier fields cannot silently fall back while rendered artifacts remain green; src/agency_source_coverage.py audits the 56-anchor official US IC expansion for agency, pack, lane, tier, checked-date, claim-scope, assurance, rights, and profile-routing metadata; and src/claim_calibration.py audits high-risk empirical, statistical, governance, safety, visual, artifact-readiness, and formalism language against source-support strength and boundary wording. src/reference_quality.py audits generated section/module links, Markdown-file links, generic body headings, lesson cross-links, and citation-table context so reader-facing references are both resolvable and informative. The artifact-evidence manifest reports 0 single-source-family claim-bearing review warnings under the current source-strength classifier; the readiness gates treat any such family as a review row, not a blocker.

PDF rendering is external to this repository. The combined PDF (AGEINT.pdf at the repo root, mirrored at output/pdf/AGEINT_combined.pdf) is produced by a shared research-project template's pandoc/xelatex render stage, not by any script inside this repo. scripts/build_curriculum.py and scripts/generate_figures.py produce everything that render step consumes — the semantic manuscript, the figure registry, and the generated BibTeX — but the PDF composition itself runs from the template checkout against this project linked under its projects/working/ tree. Refreshing the PDF is therefore a manual, out-of-band step rather than something this repo's own CI performs; scripts/audit_pdf_quality.py (and the two audit scripts whose aggregate verdict includes it) stay informational in .github/workflows/ci.yml and .github/workflows/manuscript.yml for that reason.

Generated files under output/ are not primary authoring surfaces. Edit data/curriculum/ shards first when the guide is absent; otherwise update templates, manifest/content modules, or source-anchor data, then rebuild. Cross-references are label-backed ([@sec:...], [@fig:...], [@ageintNNN]) so prose points to generated identifiers instead of literal target names. Do not improve section titles or body prose by hand-editing output/manuscript. Generated chapter H3/H4 scaffolds are chapter-specific reader landmarks, while lesson-level navigation uses Learning-path links with unit-map, module, and curriculum-atlas references. Use docs/citation_workflow.md as the canonical recipe for adding, extending, counting, and validating citations.

See docs/output_inventory.md for the generated artifact contract.

Scholarship posture

Perplexity may be used for discovery and second opinions, but the manuscript cites directly verified sources. The current anchor set includes official and standards sources from ODNI, CIA, DIA, NIST, CISA, NSA, Five Eyes partners, IMDA, OECD, ISO, IETF, OASIS, OWASP, MITRE, NATO, the EU, UNESCO, OHCHR, World Bank, ILO, W3C, OpenAPI, Council of Europe, PCLOB, NARA, NRO, NGA, FinCEN, OFAC, FATF, BIS, WIPO, DataCite, C2PA, the Model Context Protocol Project, the Agent2Agent Protocol project, the UK NCSC, the International AI Safety Report, and a policy-scholarship cognitive security source from GCSP.

The v2 source layer adds lane metadata for AI conformity/compliance, education and assessment, public-sector agentic AI, cross-border data/data spaces, human-rights governance, agent interoperability standards, workforce governance, and model/data provenance. The original ageint001 through ageint231 source identities are locked in data/source_identity/; new guide references are append-only after that range and currently extend through ageint312. The deep pass adds accessibility and digital inclusion, procurement/vendor governance, agent incident response, AI red-team assurance, public-sector transparency, and rights-impact privacy review. The evidence package pass adds model card reporting, dataset documentation, algorithmic transparency reporting, records retention/auditability, secure release/change control, risk exception governance, learner support/accommodations, assurance evaluation evidence, and procurement performance monitoring.

The May 24, 2026 source refresh brought the curated anchor set to 172 directly verified official, standards, public-domain, or scholarly sources. It adds Canada, OECD, UN, NIST, and CISA anchors for bounded agentic-AI governance, algorithmic impact assessment, public AI registers, AI incident reporting, NIST Dioptra evaluation, secure AI deployment, OT procurement, OT asset inventory, and definitive OT architecture evidence. Earlier May 22 anchors remain for CISA AI red teaming/TEVV, CISA AI data-security best practices, NIST's critical-infrastructure AI RMF profile concept note, OECD public-sector AI governance, and the NARA 2025 AI compliance plan. A separate 2026-05-22 verified batch raises the curated set to 186 anchors, adding NATO CCDCOE cognitive-warfare research, DARPA Intrinsic Cognitive Security, CSA frameworks (CDR, MAESTRO, securing autonomous agents, NIST agent red-teaming), the OWASP agentic threats-and-mitigations reference, frontier-lab agent research from Anthropic and Google DeepMind, multi-agent prompt-injection and agentic systems-security preprints, the Mandel-Tetlock analytic-debiasing study, and the UNU Macau bounded-agency policy analysis. Two 2026-06-06 internet-citation passes raise the curated set to 202 anchors by adding NIST AI 100-4 synthetic-content transparency, NIST AI 100-5 global AI standards engagement, the U.S. AI Safety Institute/NIST AI 800-1 misuse-risk draft, the International AI Safety Report 2026, MCP specification and security best practices, the A2A protocol specification, NCSC secure-AI-development guidance, MITRE D3FEND, OASIS CSAF, CycloneDX, SPDX, NIST OSCAL, SLSA, in-toto, and Sigstore. Draft status is retained where applicable rather than treated as final policy.

A 2026-06-08 primary-literature pass added the foundational primaries behind the curriculum's theoretical and technical claims: Friston's free-energy-principle and active-inference papers plus the Parr-Pezzulo-Friston textbook (the unifying cognitive framework); the canonical agentic-LLM architecture papers (ReAct, Reflexion, Generative Agents, Chain-of-Thought, Toolformer) and the Wooldridge multi-agent-systems foundations; peer-reviewed HUMINT elicitation science (the CIA MICE-to-RASCLS article and the High-Value Detainee Interrogation Group / rapport-based-interviewing literature); Heuer & Pherson's structured-analytic-techniques reference; the NIST FIPS 197 / 186-5 / 180-4 and SP 800-57 cryptographic standards behind the SIGINT assurance material; and McGuire's 1961 founding inoculation-theory experiment. These upgrade provenance for claims previously anchored to secondary sources without changing any argument.

A 2026-06-09 Data Cards pass raises the curated set to 225 anchors by adding the Data Cards dataset-documentation paper, tightening source-tier language for practitioner and vendor source-guide rows, and replacing the decorative AGEINT boundary plate with a deterministic control matrix.

A 2026-06-10 source-method and active-inference pass raises the curated set to 228 anchors by adding Buckley et al. on active inference, PRISMA-S for source-construction reporting, and Greshake et al. on indirect prompt injection.

A 2026-06-11 internet-backed visualization pass raises the curated set to 232 anchors by adding NIST AI 800-2 benchmark-evaluation draft guidance, OECD's agentic-AI landscape paper, NSA MCP security design guidance, and the Science Advances psychological-inoculation study. It also raises the figure registry to 154 with source-backed standards, evaluation, synthetic-content, data-security, source-refresh, and freshness/coverage visuals.

A 2026-06-11 analytic-tradecraft integration pass raises the curated set to 248 anchors by adding CIA/Kent tradecraft history, warning-intelligence, post-9/11 and Iraq WMD reform/postmortem sources, NATO alternative analysis, SAT evaluation evidence, forecasting calibration, and failure-theory sources. It also raised the then-current figure registry to 163 with source-backed tradecraft schematics and one evidence-derived chart for source-quality boundaries, first-principles claim decomposition, red-team negative controls, evidence ladders, probability/confidence separation, SAT evidence boundaries, and warning/failure feedback.

A 2026-06-14 official US Intelligence Community source-pack pass raises the curated set to 304 anchors by adding 56 directly verified CIA, DIA, ODNI, Intelligence.gov, NSA, NGA, FBI, and NRO public sources. The new shard is functional, not just numerical: data/agency_source_packs.yaml routes named packs into intelligence profiles, source_agency and source_pack metadata travel through reference dictionaries, agency_source_coverage_ok joins the artifact-evidence and publication-readiness gates, and the registry grows to 172 figures with an agency-source coverage dashboard. CIA Studies/CSI professional literature is treated as official public professional context, not automatically as current agency policy or measured AGEINT benchmark evidence.

A 2026-06-15 SI/OPSEC/cognitive-security literature-integration pass adds 36 directly verified scholarly, official, and public-domain sources from the synthetic-intelligence, analytic-tradecraft, OPSEC, cognitive-security, adversarial-ML, AI incident, OSINT leakage, deepfake, disinformation, social engineering, and active-inference discovery inventories. The pasted literature reviews remain discovery inputs only; final manuscript citations point to the verified source URLs in data/research_anchors/intelligence-anchors-305-340.jsonl. The registry grows to 177 figures with a source-backed SI/tradecraft/OPSEC/ cognitive-security convergence map.

A 2026-06-15 SAT literature-integration pass raises the curated set to 367 anchors by adding 27 directly verified scholarly, official, public, and publisher-backed structured-analytic-technique sources after dedupe against Heuer, Heuer/Pherson, CIA tradecraft, ICD 203, RAND, Dhami, Mandel/Tetlock, Barnes/Mandel, Ard, and other existing anchors. The new shard is data/research_anchors/intelligence-anchors-341-367.jsonl; it routes through analytic tradecraft, SAT evidence, analytic cognition, failure postmortem, warning, forecasting-calibration, and analyst-assistance profiles while keeping SAT claims bounded as review artifacts rather than universal bias remedies or autonomous judgment replacements.

A 2026-06-16 citation-expansion pass reviews 106 proposed citation headings, imports 95 directly verified and deduplicated anchors, and records 11 deferred candidates in docs/citation_expansion_2026_06_16.md. The new data/research_anchors/intelligence-anchors-368-417.jsonl and data/research_anchors/intelligence-anchors-418-462.jsonl shards are routed through non-agency research packs in data/research_source_packs.yaml so collection, cyber/ICS, influence/CI, history/legal/SAT, and agentic-AI chapters inherit richer evidence without overloading the agency-source pack contract. Accepted rows use bounded claim scopes and explicitly exclude collection tasking, exploit steps, covert-action guidance, manipulation playbooks, unsafe cyber-physical actions, and live-target procedures.

A 2026-06-12 local verifier-first, scholarship-quality, and Synthetic Analytic Tradecraft orientation pass keeps the same source corpus but raises the figure registry to 168 with the ageint-artifact-evidence-control-loop and ageint-scholarship-triangulation-map control figures and the ageint-synthetic-tradecraft-method-contract and ageint-analysis-validation-matrix method figures plus the ageint-analysis-validation-family-coverage coverage figure. It adds generated evidence manifests that bind the rebuilt manuscript, citation inventory, scholarship source-family audit, figure quality audit, rendered-reference audit, PDF quality/link audit, and stale-output scans into output/reports/current_artifact_evidence.{json,md} and output/reports/scholarship_quality.{json,md}. The current local manifest reports 330 configured generated Markdown files, 16,604 generated Markdown citation occurrences, 177 registered figures plus one non-numbered cover-art PNG, a non-stale 1,858-page ~34 MB PDF, 6,289 PDF URI links, 0 Markdown-file, file:, or launch-action PDF targets, 0 uncited or thin claim-bearing generated files, and six single-source-family claim-bearing review warnings, with passing SAT method and analysis-validation contracts plus source-metadata explicitness and claim-calibration checks. A follow-on RedTeam pass makes the analysis-validation claim classes canonical in src/analysis_validation.py, so the prose protocol, matrix renderer, scholarship audit, and artifact-evidence manifest now fail if a lane such as artifact readiness or reviewer disposition drops out while the heading and figure reference still remain. A second RedTeam follow-up adds a canonical manuscript-family-to-validation-lane map so a new claim-bearing generated family cannot enter the manuscript surface without a matching review lane, evidence signal, failure signal, and registry-backed coverage visual.

A 2026-06-13 metadata-verifier hardening pass keeps the same source corpus but closes the legacy lane/tier fallback dependency. It makes 119 previously blank metadata rows explicit: 109 legacy intelligence anchors now carry source_lane from their existing domain and source_tier from their existing source_type, while 10 source-quality support anchors carry source_quality_spine / source_quality_anchor semantics. The new scripts/audit_source_metadata.py report writes output/reports/source_metadata.{json,md} and is wired into current_artifact_evidence.{json,md} as source_metadata_ok, so a blank lane/tier row or support-anchor semantic mismatch fails local readiness. The same pass raised the then-current registry to 169 figures with ageint-source-metadata-integrity, a Python control-matrix figure for source metadata explicitness and refresh coverage.

A 2026-06-13 cover, abstract, and TOC hardening pass keeps the same source corpus and figure registry count while adding deterministic cover art outside the numbered figure registry. The abstract is now one substantial plaintext Synthetic Analytic Tradecraft contract rather than a graphical-abstract wrapper. The graphical abstract itself is now the registry-backed Python ageint-graphical-abstract Synthetic Tradecraft System Atlas in orientation, not a Mermaid stack or title-page cover. The PDF table of contents is intentionally limited to H1/H2 entries; each module exposes three chapter-specific H2 landmarks (source/profile frame, practice-lens path, assurance handoff), while H3/H4 scaffolds remain in the body for local navigation and cross-reference targets. The compact LaTeX preamble also carries TOC spacing controls so multi-digit section numbers do not collide with headings.

A 2026-06-14 claim-calibration and visual-semantics hardening pass keeps the same source corpus and adds the registry-backed ageint-claim-calibration-and-visual-semantics control figure. The figure registry schema is now 1.5: every row declares semantic role, evidence role, quantitative status, unit, denominator, counting rule, and interpretation limit, and generated PNG metadata embeds the same fields. The new scripts/audit_claim_calibration.py report writes output/reports/claim_calibration.{json,md} and is wired into current_artifact_evidence.{json,md} as claim_calibration_ok, so unsupported proof-language, p-value language, measured-performance claims, unsupported formalisms, or weak-source-only high-risk claims fail local readiness.

A 2026-06-14 abstract/readiness hardening pass makes the Abstract one continuous reader-facing Synthetic Analytic Tradecraft paragraph instead of a multi-block front-matter summary. That paragraph integrates source-quality and research anchor counts inline, shows the learner/reviewer path through evidence packets, negative controls, human review, rollback, refresh triggers, figure semantics, and validator gates, and explicitly states that artifact telemetry is not a benchmark for model capability, learning outcomes, operational effectiveness, statistical significance, or safety performance. This is still a local-readiness claim, not a publication decision.

After this pass, claim calibration reports 9,107 candidate rows, 0 hard fails, 482 boundary-allowed rows, and 5,129 review rows; scholarship quality reports 0 hard fails and keeps six single-source-family rows as non-blocking review warnings.

These counts are rebuild-time measurements, not hand-authored release claims: output/data/curriculum_outline.json carries the curriculum stats, and output/figures/figure_registry.json carries the figure count, hashes, and rendering provenance. Source-anchor rows include checked_as_of metadata and must be refreshed when source URLs, standards versions, legal text, or supported claim scope changes.

Public readiness

AGEINT is locally publication-ready as a preflight packet, not as a public release. The 2026-06-16 publication_readiness report is ok: true: artifact evidence is green, the source-refresh dashboard reports 472 current metadata rows and 0 due-soon/due/stale rows, the artifact manifest has issues: [], the parent tracked-project confidentiality guard passes, the release-surface scan finds 0 private/local path or Markdown-file-link issues, the source/license posture is recorded, and the fresh PDF audit reports 1,858 pages, 6,289 URI links, 0 file actions, and 0 bad link targets. ageint-27 is therefore closed. ageint-m1 remains todo because no public release, push, PR, archive, promotion, publication upload, DOI, or release record has been approved or performed.

Safety posture

Dual-use subjects are represented as history, governance, tabletop exercises, synthetic-data labs, public-domain imagery, source-integrity review, and defensive analysis. Do not add live target lists, operational collection steps, evasion recipes, exploit instructions, persuasion playbooks, or unsafe cyber-physical procedures.

About

Open curriculum on agentic intelligence, governance, and AI safety — organized in parts with Mermaid and Python-rendered figures, sourced from ODNI, CIA, NIST and NATO publications, built by a Python/uv pipeline into a PDF with claim-calibration and reference-integrity audits.

Topics

Resources

Contributing

Security policy

Stars

5 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages