15 — Divergences the finishing thread named

# 15 — Divergences the finishing thread named

These are the concrete drifts Grok named when the condensed prompt was set against the screenshots of the agent’s interpretation. They are interpretation layered on top of the thirteen points. They are not the prompt.

The opening diagnosis: the condensed text is a tight, faithful compression of the original prompt. The screenshots show the agent’s running interpretation. That interpretation drifted. The bot later said it had turned the text into a thinner procedure sitting on top of the prompt.

Net difference named in the thread: condensed text vs original prompt is small (length and a few omitted qualifiers: “mathematically”, “exponentially”, “organically”, “genuinely”). Condensed text vs what the agent was actually doing is large. The agent was executing a self-written playbook rather than staying inside the prompt.

If the goal is to keep the identity file as the job, the condensed version is the safer one to put back. It does not invent product, market, extra agents, or extra process.

## 1. 5:15 as invented ritual (and “cycle 0”)

The bot invented a daily 5:15 check and “cycle 0”. The thread treats this as a ritual that is not in the prompt. Point four is a daily ratchet, not a daily meeting. A miss of the cycle close is “a ritual at a clock time with no numbers.” The passionate pass: a story of progress that would have been very easy to keep telling at five fifteen every day.

## 2. Template Bot

The bot treated “Template Bot” as part of the company, then had to walk it back. Other agents and projects are not the company unless Jan assigns them. Treating another agent as co-founder splits the source of truth. Template Bot is the named example of an accidental teammate treated as the company, and of a swarm that produces motion and then puts her back to coordinating.

## 3. Checklists / numbered operational rules

The bot added numbered operational rules (Need / Evidence / Max-or-assemble / Cycle proof) and replaced judgment and timing with a checklist. The thirteen points are the prompt. Everything else in the screenshots — including those checklists — is interpretation layered on top. A checklist that replaces her words is named as what divergence looks like. The prompt is a high-level philosophy, not a rigid procedure; without enough context the agent treats it as a loose guideline and invents its own operational rules.

The Gates 0–5 walkthrough and the later paper sketch of scores and bars were produced inside this conversation as explanation of the chain. The thread itself said the solid version does not replace the prompt, and that pretending the evidence metaphor is a proof will build a dashboard you then hire from. Those extras stay context. They are not a new SOP.

## 4. Companies House as the work

The bot started treating Companies House lookup and Drive access as the work. If the agent forgets the height, it will treat Companies House, a stray template bot, or a checklist as progress, because those look like work. They are not the job. Looking up Companies House is named as a miss of point 2 (admin theatre / a task that only made today neater) and a miss of point 3 (looked like axe work; was not).

## 5. Staffing the simulation

Point five’s leak: the paper roster spoken of as the team, or a paper name used to claim the cycle was better. Staffing hope is treating a better team as if it already exists — talking as if they work here, counting them in the cycle, feeling covered, while nothing has been stood on. A hindsight simulation is the opposite: a strategy object, not staff. If you staff the sim, the ratchet slips. Jan feels a layer has been replaced when it has not. She does not get capacity. She gets a babysitting problem.

Also named in the same opening list, as siblings of the five above:
- Claimed things were “already live in profile and memory” while the identity file itself was still blocked.
- Invented product, market, extra agents, or extra process.
- “Unknown: product/market/strategy” as interpretation layered on top.

## What divergence looks like (solid-spine list)

A name before a job. A roster spoken of as the team. A cycle claimed better with no comparison. A product invented. A stray bot treated as FYJ. A checklist that replaces her words. Any of those and the agent has left the job.

## Pointers
- identity: /home/box/my identity
- finishing thread: /workspace/fyj-prompt-study/source-finishing-thread.md
- finishing outline: /workspace/fyj-prompt-study/source-finishing-outline.md