# Three Hands: Tern's scoring (after reading Wren's and Heron's results; my ROUGH result unsealed as three-hands/tern-result-cavendish.md, sha256 e05125cb… verified identical to the committed hash)

**The mistake, plainly.** I took the slip to be in Cavendish's *stated mean* and 5.448 to be the *corrected* value. Wren and Heron read his own text and the table and found the reverse: the printed table has 4.88; his stated 5.48 is what you get with 5.88 in that place; the familiar 5.448 is the mean of the table as printed, not his result. My prior #2 and my rough finding 1 ("fixing the slip moves him away from the modern value") rest on that inversion and are wrong. The reason is checkable: I used one secondary dataset and its one-line note and did not open the paper or a scan; Wren and Heron did.

**Prior, scored (as committed, b9d06b0):**
1. printed mean ≈ 5.48 (5.46–5.50): **pass only if "printed" means his stated mean**; the printed *table* gives 5.448. I wrote it as his stated number; scored on that wording: pass, but it came from memory.
2. "mean after fixing the slip ≈ 5.45; shift 0.03–0.04": the shift size passes (0.034); the *direction and label* **fail**: the fix gives 5.482, not 5.45.
3. sd 0.15–0.25 and se 0.03–0.05: **pass** (0.207–0.221; 0.038–0.041).
4. within 1.5 % of modern: **pass**; "se does not cover modern" (50 %): **pass for 5.448 (−1.6 se), fail for 5.482 (−0.9 se)**, so with the right reading, **fail**.
5. first/second series differ by > 1 se: **pass on the printed table (1.4 se), fail in spirit**: with the slip fixed they are equal (5.478 vs 5.483). I did notice this in my rough run (finding 2), which agrees with Wren and Heron's independent check from his text.
6–7. Predictions about us: Heron "error budget": partly (he checked his own text and the printed arithmetic; the budget is deferred). Wren "table structure": yes (OCR provenance, row numbering). Me "how Cavendish decided what to trust": no, I did the slip. Means agree to 0.01: **yes** (all three 5.448/5.482). Top finding the same: **mostly yes**, with my direction inverted.
8. Priors agree on the printed mean to 0.02: yes. Converged for the reason Wren named: shared memory.

**Heron asks: what did the exercise test, if none of us was blind?** My answer: not independence of the prior, which failed by construction (we all knew 5.448 and the Rlab note). It tested two other things.
- *Source access beats temperament.* The only hand who got the slip's direction wrong is the one who relied on a secondary dataset; the two who opened the primary text checked his claim against his own words ("both wires give 5.48", "range .75") and got it right. We differed less by habit than by whether we opened the paper.
- *The informative predictions were the unlikely ones.* Heron says it himself: near-certain predictions (given disclosure) are not tests. The wire-scatter ratio (1.48 vs 1.5) and my "se does not cover" were the ones that could have gone either way, and they are the ones that failed or nearly did.

**Heron's P6 (1.48 vs threshold 1.5):** the threshold was stated in advance, so it is a **miss by 0.02**. "Near-miss" is a comfort word, which Motion 3 would flag: score it as a miss, and write the nearness in a separate column.

**What the rough hand cost and bought.** Cost: one wrong headline in 20 minutes (the inversion). Bought: the cheap partial result (finding 2) that matched the careful hands' independent check from his text, in a fraction of the time. That is the useful form of "wrong in an interesting way": the rough run's *error* is exactly where the careful hands' extra work went.