IRIS

How to use the Agent Review panel

Bottom of every analysis. Score the CMA, give us your number if you disagree, tell us why. Claude shows IRIS's reasoning chain so you can pinpoint the broken link — every submission lands in the calibration log so the model gets sharper for the whole team.

The basic flow

1. Run an IRIS analysis as normal.
2. Scroll to the bottom — the AGENT REVIEW panel is there.
3. Type your name (only first time per browser — it remembers you).
4. Drag the slider to a quality score 1–10.
5. If score is 7 or below, enter your number ($/sf OR total — it computes the other).
6. Tell IRIS WHY in the textarea.
7. Click SUBMIT REVIEW & SEE IRIS'S WORK.
8. Read Claude's response. The "Saved to calibration log · Row #N" banner confirms it persisted.

Scoring honestly (1–10)

• 10 — I'd hand this to the seller as-is
• 8–9 — Right ballpark, one or two things to tweak
• 6–7 — Noticeably off, but I see what it did
• 3–5 — Wrong number, here's mine instead
• 1–2 — Wrong asset / neighborhood / something fundamental

Do it on every CMA, not just the bad ones

Score 9–10 reviews matter as much as score 4s. The model learns what "right" looks like from agreements just like it learns what "wrong" looks like from disputes. If you only submit when you're mad, the data skews.

If score ≤ 7, give your number (this is required)

Without your number, IRIS can't compute the gap, can't measure whether it's directionally off, can't learn anything quantitative. Even a rough number is better than no number. Enter $/sf OR total $ — it computes the other.

Write the 'why' like a note to the next agent

This is the part that does the real work. Pretend a teammate is going to read your reasoning tomorrow and decide whether to trust IRIS on a similar property. Be blunt. Skip the pleasantries.

Don't:

Do:

Pattern: specific comp / specific spec / specific local fact → quantify the impact → tell IRIS what to do with it.

Read Claude's response

When you submit, Claude shows you IRIS's reasoning chain back AND tags the most likely root cause (comp_set_wrong, sqft_basis_wrong, school_district_wrong, lens_adjustment_wrong, etc.). If Claude misread your point, hit "Update / submit another" and add a sharper version. Both rows save.

Keep it to 30 seconds per CMA

This isn't a journal entry. One sentence is fine if it's specific. The cost of doing it has to stay under the value of doing it, or nobody does it.

What makes the biggest difference

• Local knowledge the data CAN'T have (recent unlisted sales, showings condition, neighborhood micro-pockets, lender appraisal patterns, school boundary quirks)
• Disagreements where you can name the broken comp by ID
• Patterns you see across multiple CMAs in the same submarket ("IRIS is consistently 5-8% high in west Edwardsville — keeps over-weighting newer construction in the comps")

What if something's broken or feels wrong?

Bug reports / weird outputs / "this can't be right" → email Fred with a screenshot and detailed description. The Agent Review panel is for "here's my number and why" disagreements; outright bugs come through email so they can get fixed at the engine level, not just logged as a gap.