The Tool That Built Itself

One evening, one non-coder, a self-correcting PR reviewer

The blast radius

A one-line dependency add can smuggle a whole system past a diff-only review

The Change Advisor's own VISION.md warns that "a 1-line dependency addition can bring in 68 files, 9 tools, 3 hooks, and 200MB of dependencies." The clean PR is the disguise.

And this wasn't hypothetical — it happened on a real PR.

68

files behind one line

9

tools

3

hooks

200MB

of dependencies

The real PR

The one line wired a persistent-memory module into a project built on visible evidence chains

VISION.md names the case: "Tested on a real PR (engram-lite memory system integration)." The module is kenotron-ms/engram-lite — "Lightweight persistent vector memory for AI agents" — created 2026-03-04, its main tree holding 67 files.

Both are good engineering. Together, they are incoherent.

The module — kenotron-ms/engram-lite, an invisible-capture memory system
The host — a project whose principle is visible evidence chains
The scan — a diff-only review waves the one line straight through

The conflict a diff can't see

Invisible capture meeting visible evidence chains poisons the memory

VISION.md line 47: "A memory system designed for invisible capture arrives at a project built on visible evidence chains — both are good engineering, together they're incoherent." Junk test sessions would be silently ingested as real knowledge.

A mechanical scan misses this entirely — so who catches it?

Silent capture ingests every session, no matter its quality
Junk in — test, competitive, and throwaway sessions look identical to real work
Self-poisoning — noise becomes "real knowledge" the system now trusts

The person in the loop

The one holding the product context was a self-described non-technical PM who wrote no code

Chris Park (cpark4x) — described in the source repo's own deck as "a non-technical PM who wrote no code" — was sole author of all 12 commits to cpark4x/amplifier-change-advisor. His product knowledge, not a code review, is what made the poisoning obvious.

But product context alone wasn't enough either — it took a pairing.

The leverage

Product context paired with two AI executive lenses surfaced what neither would catch alone

The recipe analyzes through two lenses: "the CTO asks HOW, the CPO asks WHY and FOR WHOM." VISION.md line 254 records that the discussion "surfaced critical insights the automated analysis alone would not have caught."

This became a repeatable method — so he wrote its principle down.

CTO Lens

Technical Excellence and Risk — asks HOW

CPO Lens

User Value and Product Strategy — asks WHY and FOR WHOM

Principle, written first

A 306-line VISION.md declared the tool never recommends APPROVE or REJECT — only DISCUSS

Commit f38af38d added docs/01-vision/VISION.md with +306. Line 202: "The tool never recommends APPROVE or REJECT as a final answer. It recommends DISCUSS." Line 16: "The tool doesn't decide — it makes you a better decision-maker."

Then he pointed that principle at his own recipe.

+306

additions in VISION.md (commit f38af38d)

DISCUSS

the only verdict the tool may give

It corrected itself

Reviewed against its own vision, the recipe caught its own contradiction

The recipe was emitting APPROVE/REJECT — contradicting the vision's "never decides." Commit 772050f7 fixed it: "verdict output always DISCUSS, never APPROVE/REJECT," touching only recipes/pr-review.yaml (+2/-2) to align with Principle 4, "The Decision-Maker Decides."

The tool that refuses to decide proved its principle on itself.

# commit 772050f7
# recipes/pr-review.yaml (+2/-2)

- Verdict: APPROVE / REJECT
+ Verdict: DISCUSS

What to keep

In one evening and six commits, the process became the product

Six commits on 2026-03-05, from 21:45 to 23:56 UTC, by one author: the recipe, a 306-line vision, the verdict fix, executive lenses, a backlog, and a session deck. The AI didn't replace judgment — it gave judgment somewhere to land.

Shipped and public: installable via amplifier run --bundle git+github.com/cpark4x/amplifier-change-advisor@main.

9907bc36 — initial PR-review recipe (+202)
f38af38d — rename + 306-line VISION.md
772050f7 — verdict fixed to DISCUSS-only
d89e8549 / d66ca328 / 5f924329 — lenses, backlog, story deck

Sources

Research Methodology

Data as of: 2026-03-05 session; repo verified through latest commit 2026-03-15.

Feature status: Shipped / public and evolving (v2.0.0 documented 2026-03-11).

Primary source: the public repos cpark4x/amplifier-change-advisor and kenotron-ms/engram-lite, re-derived from live gh API output by an independent checker.

Commands run:

gh repo view cpark4x/amplifier-change-advisor — description & README
gh api repos/cpark4x/amplifier-change-advisor/commits/f38af38d --jq '.files[]' — VISION.md +306
gh api .../commits/772050f7 --jq '.commit.message, (.files[])' — verdict fix (+2/-2)
gh api repos/cpark4x/amplifier-change-advisor/commits?per_page=100 — commit order & authors
gh repo view kenotron-ms/engram-lite --json name,description,createdAt,pushedAt
gh api repos/kenotron-ms/engram-lite/git/trees/main?recursive=1 --jq '[.tree[]|select(.type=="blob")]|length' — 67 blobs
grep -niE '68 file|9 tool|3 hook|200 ?MB|invisible|evidence chain|APPROVE|REJECT|DISCUSS' VISION.md pr-review.yaml

Gaps: "68 files / 9 tools / 3 hooks / 200MB" are session-time blast-radius figures documented in VISION.md (current main tree has 67 blobs); the "non-technical PM" label is Chris Park's own self-characterization in the source deck; "4 executive lenses" was accurate at session date but COO/CFO were removed in v2.0.0, leaving CTO/CPO in the current recipe.

Primary contributor: Chris Park (cpark4x) — sole author of all 12 repo commits.

STATUS: SHIPPED / PUBLIC