Step 2 · Mental model · Evidence discipline ENPT
Alembic · Blind Planning Visual Course

Evidence discipline

Non-negotiable rules for blind agents: brightdata scrape for URLs, Read full local paths, cite command + line — never metadata-only summaries.

1

The big idea


Blind planning exists to produce comparable independent plans. That only works if every agent follows the same evidence rules — not intuition, not prior session memory, not "probably works like…".

The prompt forbids treating earlier RE docs, gists, courses, or this conversation as ground truth. Injected corpus in PROMPT-CORPUS-EMBEDS.md is the baseline evidence — but Phase 1 still requires fresh forensic RE on binaries and disk.

Four source classes, four methods. Break any rule and the plan is not comparable.

Think of it like… a court trial with chain-of-custody. A witness summary is inadmissible if you cannot produce the exhibit — the scrape file, the full path read, the strings output line.

Forbidden as truth: prior docs/alembic/*.md from other sessions, published gists, reverse-engineering from other agents, WebSearch/WebFetch/MCP Bright Data.

Forbidden phrasing: "probably", "must be", "seems like" without command + output or path + verbatim excerpt.

Corpus version: PROMPT-CORPUS-EMBEDS v1.3.0 — 10 scrapes in corpus-cache/, P0/P1 integral paths, 8 HF image descriptions.

2

Evidence table


SourceRequired methodForbidden
External URLbrightdata scrape <url> -o corpus-cache/<slug>.mdWebSearch, WebFetch, MCP
Local pathRead file entire — cite path + excerptOnly understanding.md, metadata.json, index summary
Installed binarystrings, grep, disk artifactsSpeculation without output line
HF imagesFull visual description in embedsFilename-only mention

Source: PROMPT-BLIND-PLANNING.md § Disciplina de evidência + PROMPT-CORPUS-EMBEDS.md header

3

Commands


brightdata — external evidence (loop-engineering/SKILL.md)
# Scrape official docs into corpus-cache
brightdata scrape https://martinfowler.com/articles/agentic-ai.html \
  -o docs/alembic/corpus-cache/martinfowler-harness.md

brightdata budget   # before bulk scrapes
corpus-cache — pre-mined scrapes (10 URLs)
docs/alembic/corpus-cache/
  langchain-anatomy.md
  openai-harness.md
  martinfowler-harness.md
  arxiv-2604.25850.md    # Agentic Harness Engineering
  arxiv-2604.11378.md    # loops → graphs
  ...
Binary RE — Phase 1 example
strings -a ~/.local/bin/droid | rg 'propose_mission|start_mission_run'
jq '.interactionMode' ~/.factory/sessions/*/*.settings.json
cat ~/.kimi-code/agents/*/wire.jsonl | head
bash docs/alembic/build-corpus-embeds.sh
# Regenerates PROMPT-CORPUS-EMBEDS.md after new scrapes or paths
4

Try it: valid vs invalid evidence


Next: three chords — what the blind agent must reverse-engineer and fuse after evidence is secured.