# Merge 01 of batch ep1-mech: five mechanism items vs NP2

Goal context (from the brief): NP2 = 518.83 on the board (39th), 526.1156 locally on seeds 0..999. Top 10 needs about +5.8 paired, 1st about +9.6. Source files: `.imd/reads/artifacts/item-*`, one per item.

**Headline: no item beat NP2. Four confirmed large paired losses; one was blocked before any measurement. No item has mean_diff >= 2*se, so no `code` is copied below (see section 3).**

## 1. Table (ranked by t = mean_diff/se, highest first)

| Rank | key | mechanism | validate_ok | blocked | best tune mean_diff (seeds 10000..10499, n=500) | confirm seeds | n | mean_diff | se | worst | t |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | ev-fee | Expected-value fee; fMid blended 5% toward EV center | true | no | -0.1468 (se 0.0168) | 2030000..2031999 | 2000 | -0.1166 | 0.0095 | -1.9167 | -12.3 |
| 2 | arb-classifier | Classify prints by implied arb size; set pHat alpha | true | no | -20.9106 (arb2, se 0.7009) | 2010000..2011999 | 2000 | -20.3658 | 0.2857 | -281.57 (seed 2010992) | -71.3 |
| 3 | exact-block | Floor bid/ask at geometric pHat/spot gap, literal orientation | true | no | -272.3477 (exact_k1_keep, se 2.5150) | 2000000..2001999 | 2000 | -272.4799 | 1.1948 | -576.12 (seed 2000065) | -228.1 |
| 4 | kalman-phat | Scalar Kalman filter replaces fixed pHat alphas | true | no | -127.4176 (K100, se 0.9100) | 2020000..2021999 | 2000 | -127.1881 | 0.4457 | -200.26 (seed 2020751) | -285.4 |
| n/a | third-print-alpha | Later-print alpha (not implemented) | false | YES: step=setup, error="BLOCKED platform" | none (tune empty) | 2040000.. (planned) | 0 | null | null | null | n/a |

t values are computed here from the reported mean_diff and se. Tune values for ev-fee, arb-classifier and kalman-phat are the highest of their variants. exact-block's best tune is also its frozen variant (highest of six, all negative).

## 2. Flags

- **np2_check vs 526.1156 (tolerance 0.01):** none differ.
  - exact-block 526.11557, arb-classifier 526.11557, kalman-phat 526.11557, ev-fee 526.115572. All report sha256 51f1c1a9d4999ee460eb8bbc66880da8074b173399a4f5d71f96671482e18ac0.
  - exact-block and kalman-phat also report the CH check as 525.525882 (matches the required 525.525882178).
  - third-print-alpha: np2_check n=0, score null. It was never run, so it is **not verified**.
- **Confirm block start vs assigned confirm_start:** the assigned confirm_start values are not in the files I was handed, so I could not check them. The reported starts are 2000000, 2010000, 2020000, 2030000 (each block is 2000 seeds), and third-print-alpha planned 2040000. All are at or above 2000000, and the blocks do not overlap. Please check them against the assignments.
- **Paired n below 2000:** third-print-alpha (n=0, blocked). The other four have n=2000, which meets the minimum but not by any margin.
- **Blocked:** third-print-alpha stopped at the setup guard "BLOCKED platform". This is an environment failure, not a finding about the mechanism. It stays untested.
- **Failed-source caveat (exact-block 4, arb-classifier 3, kalman-phat 1):** the handed results do not say which cited sources failed to hold, so I cannot tell which claims they affect. I weighed the claims this way:
  - **Measured numbers** (paired means, se, hashes, the NP2 and CH checks) come from the authors' own simulator runs. They do not depend on a web page, and I treat them as reliable.
  - **exact-block's orientation claim** (the literal floor sits on the wrong side) rests on one cited page, `.../6b6921f.../amm_sim_rs/src/market/arbitrageur.rs`. The result calls it an "inference from fee-on-input equations", and the corrected formulas were not tested. Treat the claim as plausible and unconfirmed.
  - **arb-classifier** cites 19 URLs, mostly simulator source files. Its measured loss stands. Its "stale pHat causes missed arbitrage" explanation is labelled a hypothesis and was not measured.
  - **kalman-phat** cites only the simulator repo. Its measured loss stands.
  - **ev-fee** and **third-print-alpha** were not flagged.

## 3. Items with mean_diff >= 2*se

**None.** Every confirmed mean_diff is negative (t from -12.3 to -285.4). There is nothing to copy as a delta, code_sha256 or code. For reference only, the frozen (rejected) variants' sha256 are:

- exact-block (exact_k1_keep): 4a82abd82d7090e22ee99c6deb96e43ae78fcc28b16ac1613acd1532b55d35da
- arb-classifier: d30bbb066b3854da62eac894a5171b3d0553b2b77da1c8a742a18bcf18c33c7e
- kalman-phat: 9acbd3b7259998b8209c291d80d40cd08928082142371520ee57f645d1428c42
- ev-fee: e1745bc56dee6c4fde2b3f368a7b3a51479807c4496b49b707be0d4a05c3de7e

The full code of these variants is in the source result files under `.imd/reads/artifacts/`. I did not copy it here because it did not qualify.

## 4. Each item's `next`

- **exact-block:** reject the literal mechanism. The simulator charges ask_fee when spot<fair and bid_fee when spot>fair, so a correctly oriented zero-arb floor would be askFee >= 1-spot/pHat when pHat>spot, and bidFee >= 1-pHat/spot when pHat<spot. This was not tested.
- **arb-classifier:** reject and retain NP2. Hypothesis only: stale pHat causes missed arb or coincidental retail matches. Classification accuracy was never measured.
- **kalman-phat:** reject and retain NP2. A future test could model whether the first print was arb or retail, rather than treating every first print as nearly noiseless. That is untested.
- **ev-fee:** do not pursue the static EV center. A revisit needs a dynamic model (post-trade displacement that arbs reclaim, the normalizer's band state) and a bias correction for the raw spot-pHat shift.
- **third-print-alpha:** rerun setup where the Linux x86_64 platform guard passes, then implement and benchmark the frozen variant.

## 5. Three mechanisms most worth combining (<=150 words)

No tested mechanism is a gain, so there is nothing to stack yet. The three below are ordered by how little they lost and how directly they target the goal:

1. **third-print-alpha.** It is the only untested item, it is a small change to NP2's pHat alpha path, and it needs only a rerun on a working host.
2. **exact-block, orientation-corrected.** The loss came from the literal orientation. The correctly oriented floor from its `next` has not been tried.
3. **arb-classifier.** It lost only about 20, far less than Kalman or exact-block. A softer version might help. Try it only as a gate on how much pHat moves, not as an alpha override.

Drop kalman-phat and ev-fee. EV's best tested weight was 5%, and it lost monotonically as the weight rose.