Real commands run against a real, unmodified checkout of
github.com/plausible/analytics (analytics/) to produce every raw input file
in this directory. Every grep baseline and EKOS response here is real output
from these exact commands, run in this order.

--- Q1: "What columns does the imported_browsers table have?" ---

# Baseline 1 -- best-case targeted grep: the agent already knows this is the
# right file and the right line before searching. Not realistic, included
# for honesty (it's the one case grep wins).
$ grep -A 15 "CREATE TABLE plausible_events_db.imported_browsers" \
    priv/ingest_repo/structure.sql
  -> saved as q1-grep-targeted.txt

# Baseline 2 -- realistic grep: the agent does not know which file has the
# real schema, so it searches the whole repo with context lines, the way an
# agentic coding tool actually would.
$ grep -rln "imported_browsers" --include="*" . | grep -v '^\./\.ekos'
  -> 32 real files matched
$ grep -rn -C3 "imported_browsers" --include="*" . | grep -v '^\./\.ekos'
  -> saved as q1-grep-realistic-repo-wide.txt

# Baseline 3 -- naive: the agent found the one correct file (structure.sql,
# via baseline 2's grep) and just reads the whole thing.
$ wc -l priv/ingest_repo/structure.sql   # 366 lines
$ wc -c priv/ingest_repo/structure.sql   # 11,962 bytes
  -> tokenized directly from the real repo file, not copied here

# EKOS side -- the same two-step "I don't know the id yet" reality: search,
# then fetch state for the match, over real MCP stdio JSON-RPC.
$ ekos mcp serve --workspace analytics/
  -> tools/call ekos_search {"query":"imported_browsers"}
     saved as q1-ekos-search-response.json
  -> tools/call ekos_state {"id":"<id from the search result above>"}
     saved as q1-ekos-state-response.json

--- Q2: "What tables exist in the Postgres schema?" ---

# grep/naive baseline -- there is no grep pattern that cleanly extracts
# "every CREATE TABLE name" without also matching migration files, test
# fixtures, and ALTER TABLE statements naming the same tables; the realistic
# way an agent answers this from raw source is reading the whole dump.
$ wc -l priv/repo/structure.sql   # 2,738 lines
$ wc -c priv/repo/structure.sql   # 78,793 bytes
  -> tokenized directly from the real repo file, not copied here

# EKOS side -- one structured query against the already-compiled ledger.
$ ekos ekl "FIND Object WHERE kind = 'Table' AND name CONTAINS 'public'"
  -> saved as q2-ekos-ekl-list.txt (42 real tables)

--- Token counting ---

$ python3 count-tokens.py
  -> saved as count-tokens-output.txt

tiktoken's cl100k_base encoding (GPT-4-class, a standard reference
tokenizer) was used throughout -- not a hand-rolled words/4 heuristic.