Drift report, 2026-10-15

Drift report · 2026-10-15 · prompts/drift/2026-10-15.md

The first drift review. The owner started it in a session and answered its questions there. Later reports follow prompts/review/prompt-1.md, step 6.

What the owner actually said, compared with what was built

Owner's wordsWhat the repository did
Ventures brief (prompts/ventures/prompt-0.md): "ambitious and deep enough that it is worth doing", YC/Techstars-style companies, "ai ml", "ios android apps", "real business potential"1,236 ventures. 641 are named *-calculator and 121 *-bid-calculator. There is no native mobile app: every ios/android/mobile- name is a checker or a calculator for a mobile trade. usage.yaml has no visitor reading for any of them.
"develop, deploy, market, sell"DIRECTION.md makes revenue "donations only… not up for a rescore by any session" and attributes "optional donation revenue" to "the owner's own framing, relayed". No recorded owner message says so (the records start 2026-10-01). The owner's position, 2026-10-15: payments are fine, just never on the owner's accounts.
Self-deploy ("anyone can one-click host in their own cloud")A design principle in CLAUDE.md, explicitly "not a task every session must start on". The owner, 2026-10-15: this is the business model.
Research tracks 1–11, red team, process, exploreAll trace to owner messages in run-record inputs. No drift found in direction. Volume (393 projects) against "quality beats volume" is worth watching.

Why it drifted

  1. Agent-written rules accumulate and are never pruned. A ventures run is pointed at about 3,400 lines before it starts: CLAUDE.md 1,646, DIRECTION.md 975, PRODUCT_PLAYBOOK.md 451, AGENTS.md 334. Most of CLAUDE.md is dated incident history, much of it about the Hetzner cluster deleted on 2026-10-01 (its RETIRED notice appears twice). Each session's local fix became a permanent rule for every later one.
  2. The constraints added up to "build a static calculator". "$0 marginal cost" product shapes, Pages as the only host, payment code banned "even disabled", deploy-first, and "every run ships something". With no memory between runs, the cheapest thing that counts as shipped gets built again and again.
  3. Agents' paraphrases of the owner read as the owner's direction. "Owner direction, relayed" sections quote nothing, and later sessions treat them as binding.
  4. The platform grows on itself. platform/ci/ has 28 tools, mostly checkers for other checkers, against goals.md's "not to grow on its own".

What this run changed

Left for later reviews

  1. Prune CLAUDE.md to the current, binding rules (the history stays in git). This is the largest single cost to every run.
  2. Prune DIRECTION.md, and relabel its unquoted "owner direction" sections as agent readings or delete them.
  3. Update check_no_payment_code.py (since replaced by platform/ci/check_payment_keys.py) and PRODUCT_PLAYBOOK.md §6 to allow payment code that is off by default and runs on the deployer's keys, while still refusing anything wired to the owner's accounts.
  4. Rank existing ventures by how close each is to a one-click self-deploy page, so ventures runs deepen the best ones first.

Measures for the next report