Two design critics watch the pipeline
Amplifier · Experimental · v8.1.0Amplifier's document-generation recipe puts every generated doc through an automated validation pipeline from v0 to v3. It confirms the checks pass. It never confirmed that a design expert would actually call the result better.
First, is that pipeline even real? Look at what it actually runs.
The recipe's Version Flow runs structural validation first (ensuring sections exist), then parallel content checks, then parallel quality checks — substantial, genuine machinery.
It clearly does a lot. But does doing a lot mean doing good?
Automated checks can confirm sections exist and rules pass. They cannot tell you whether validation quietly cost you clarity while it gained you structure. Validation may not be purely additive — and nothing in the pipeline was watching for that.
To see that, you need a different kind of reviewer.
From Microsoft's design-intelligence bundle come two specialists: a voice-strategist and a layout-architect. Each brings a design expert's judgment the automated checks lack.
The trick is how they were wired in — without touching the document.
The recipe is explicit: the feedback is purely observational — the agents make NO changes to the document. Their prompts say “Do NOT rewrite or fix anything.” Observation, not intervention, is what keeps the before/after honest.
So what exactly do they measure, and when?
Each agent scores the raw v0 output before validation and the final v3 output after it — on named dimensions, persisted to di_v0_feedback.json and di_feedback.json.
Same scale, before and after — which sets up the comparison that matters.
Report section 6.3, “v0 → v3 Comparison (CRITICAL FOR EVALUATION),” shows before/after scores per dimension — so the pipeline can finally reveal that validation is non-additive.
That third question is the whole point.
Shipped EXPERIMENTAL by Mollie Munoz on 2026-01-30 (PR #202), the lesson generalizes: treat your own validation as a hypothesis, not a guarantee. Measure whether it helps — don't assume it.
Faith becomes a measurement you can check.
Data as of: local microsoft/amplifier checkout, HEAD dated 2026-03-02 (commit 676d677).
Feature status: EXPERIMENTAL — introduced in recipe v8.1.0 (2026-01-30).
Primary source: recipes/document-generation.yaml in microsoft/amplifier (3021 lines).
Research performed:
git remote -v — confirmed origin git@github.com:microsoft/amplifier.gitgit log --oneline -- recipes/document-generation.yaml — commits 375aaed and 0cf1d50git show -s --format='%H %ci | %an | %ae' 375aaed — Mollie Munoz, 2026-01-30git show --stat 375aaed — 1 file changed, 503 insertions(+), 5 deletions(-)gh pr view 202 --repo microsoft/amplifier — PR #202 title and author verifiedgrep -n 'agent: "design-intelligence' recipes/document-generation.yaml — voice-strategist, layout-architectgrep -n 'id: "di-' recipes/document-generation.yaml — v0 and v3 feedback stepssed -n '18,36p' and '2935,2946p' — Version Flow and section 6.3wc -l recipes/document-generation.yaml — 3021 linesGaps: The specific directional finding (that validation lifted navigation while costing clarity/readability) could NOT be verified — no di_feedback.json or di_v0_feedback.json run artifacts were found. The recipe scores those named dimensions and section 6.3 explicitly asks whether validation introduced new concerns, so the mechanism is verified, but observed values are not present in any local artifact.
Contributors: Mollie Munoz <mollie.munoz@microsoft.com> (authored PR #202 / commit 375aaed), co-authored-by Amplifier.