Colombia conflict

colombia-conflict — CEV final-report knowledge engine

A permanent, queryable knowledge layer distilled from the Colombian Truth Commission's final report Hay Futuro Si Hay Verdad (2022 — 11 tomos / 24 books / ~8,900 pp / ~3.67M words). It carries the report's substantive learnings and insights, and gives agents tools to act on the problematics the report describes — above all, the goal of no repetición (non-repetition).

Architecture (kg / LLM-wiki)

Mirrors the workspace's kg LLM-as-index pattern:

• Substrate (canonical) — references/: the 12 per-volume digests + synthesis.md (master consolidation) + lexicon.md. These are the knowledge pages; the agent reasons over them.
• Full-text grounding — references/fulltext/*.txt.gz: the complete extracted text of all 24 volumes (~3.67M words, gzipped ~7.7 MB). cc.py source greps this for verbatim source passages — the digests summarize, the full text lets you quote the report exactly.
• Source binaries (on demand) — sources/MANIFEST.json lists the 11 downloadable units (10 digital PDFs + the Colombia-adentro ZIP, each with sha256). scripts/fetch_sources.sh pulls + verifies the 385 MB of PDFs when actually needed; they are deliberately not committed.
• Projection (routing) — references/knowledge-index.md: a dense, one-line- per-item catalog auto-generated from data/ + digest headers. Tier-1 retrieval scores against this.
• Structured data — data/: statistics.json (the conflict universe), actors.json (collective-responsibility shares), recommendations.json (67 recs / 8 blocks), concepts.json (the coined lexicon).
• Engine — scripts/cc.py: two-tier load (catalog tier-1 → body-grep tier-2), data queries, the align non-repetition scorer, and the catalog generator. The agent is the query engine — load surfaces the relevant pages; the agent does the reasoning.

When to reach for it

Load this skill whenever a task touches the Colombian armed conflict or its aftermath — peacebuilding, transitional justice, victims and reparation, paramilitarism, the FARC-EP/ELN, drug policy, the land/agrarian question, the enemigo interno, falsos positivos, the UP genocide, social-leader killings, differential harms (ethnic peoples, women/LGBTIQ+, children, exiles, campesinado), territorial/human security, memory, or non-repetition — and whenever you must ground a claim in a sourced statistic, responsibility share, recommendation, or coined concept rather than guessing.

Usage

# Two-tier kg retrieval over the knowledge pages
python3 scripts/cc.py load "desaparición forzada responsabilidad" -n 6

# Query the 67 recommendations
python3 scripts/cc.py rec --theme drug-policy
python3 scripts/cc.py rec --block 6            # the human-security block
python3 scripts/cc.py rec --search "social leaders protection"

# Sourced statistics / actors / lexicon
python3 scripts/cc.py stat homicides
python3 scripts/cc.py stat --all
python3 scripts/cc.py actor paramilitaries
python3 scripts/cc.py concept --search "cuerpo-territorio"

# "Does this proposal advance no repetición?" — score against the roadmap.
# align is LEXICAL overlap + a curated anti-pattern (contrary-stance) check —
# it flags proposals that CONTRADICT a block (e.g. "increase fumigation"), but it
# is not a full stance model: reason about polarity and gaps yourself.
python3 scripts/cc.py align "a program to legalize coca cultivation and support campesinos cocaleros"

# Verbatim grounding — grep the full source text for exact passages
python3 scripts/cc.py source "Bojayá iglesia" -n 3
python3 scripts/cc.py source "responsabilidad del Estado en el paramilitarismo"

# Regenerate the routing catalog (after editing data/ or references/)
python3 scripts/cc.py index            # write
python3 scripts/cc.py index --check    # CI: exit 1 if stale

# Fetch the source PDFs (385 MB) on demand, sha256-verified
scripts/fetch_sources.sh [DEST_DIR]
python3 scripts/build_manifest.py      # regenerate the manifest from a local corpus

Recommended agent flow

1. load <topic> to pull the relevant knowledge-page sections into context.
2. Read the cited references/ page(s) for the full, faithful detail.
3. Use rec / stat / actor / concept for precise sourced facts.
4. source <topic> when you need to quote the report verbatim (grounding a claim in the exact source prose, not the digest summary).
5. For a proposed intervention, run align to see which recommendation blocks it advances — then reason about gaps the report would flag (impunity, structural causes, differential harms, terceros civiles).

The substance (one screen)

• Thesis — Hay futuro si hay verdad. The war (≈9M victims, ~90% civilian) persists because structural causes were never confronted; the demand is No Matar and sacar las armas de la política.
• Causal web (entramado) — land/agrarian question · restricted democracy + enemigo interno · selective State (más territorio que Estado) · narco + the war on drugs · impunity · economic model + terceros civiles.
• Responsibility (collective) — paramilitaries lead homicide/disappearance; FARC-EP lead kidnapping/recruitment; the State owns falsos positivos and bears responsibility for paramilitarism. "A mayor poder, mayor responsabilidad."
• Differential architecture — ethnic peoples, women/LGBTIQ+, children, exiles and territories get their own volumes; a single national narrative would reproduce the erasure that enabled the war.
• Telos — 67 recommendations / 8 blocks: full Accord implementation, ELN negotiation, integral reparation + memory, inclusive democracy, drug-war → legal regulation, breaking impunity, human security, agrarian/territorial reform, a culture of peace.

Provenance & honesty

• Source: comisiondelaverdad.co, CC BY-NC-SA 4.0. Statistics carry their registry (RUV / CNMH / JEP-CEV-HRDAG) and both figures where the report prints two. recommendations.json is encoded at block granularity (faithful to the synthesis, not a verbatim transcription of all 67 articles).
• This is collective (ethical-political) responsibility as the CEV framed it — not individual criminal guilt (the JEP's domain).
• The full extracted text (all 24 volumes, gzipped) is bundled under references/fulltext/ for verbatim grounding. The 385 MB of source PDFs are not committed — they are fetched + sha256-verified on demand via scripts/fetch_sources.sh (sources/MANIFEST.json).

References

• references/synthesis.md — the master consolidated synthesis.
• references/digests/*.md — 12 per-volume sourced digests.
• references/lexicon.md — the coined-concepts glossary.
• references/knowledge-index.md — the tier-1 routing catalog (generated).
• references/fulltext/*.txt.gz — full extracted text of all 24 volumes (verbatim grounding).
• sources/MANIFEST.json + scripts/fetch_sources.sh — on-demand PDF retrieval.