# Independent review of the measured claims

The direct event recount matches the published hit-window, tail-hit, positive-pre-damage-health, credit-spend, timer-decrement and timer-reimbursement totals for **all 354 trials in 59 cases**, with zero mismatches. This is a separate reader of the original event streams; it does not call the published scorer. No new emulator trials were run. The machine-readable receipt pins the input summary, methodology, scorer and selected event files by SHA-256 and contains event IDs for the selected headline results.

| Claim under review | Recount and qualification |
| --- | --- |
| Gold Proton Cannon reaches 181 | Both ordinary-health dense-direction fixtures contain 181 connected-hit stores. Of these, 114 MAME / 116 FBNeo occur with positive HP immediately before damage, including the lethal event; 67 / 65 occur after HP reaches zero. Every event also has the actual Gold move selector active, and both runs reach the first exit rather than the callback limit. |
| Gold's display stops at 99 | The exact store at `$01D1A8` continues to execute with `before=after=99`; externally counting these stores produces 181. This is neither a display reading of 181 nor a proof of unlimited duration. |
| Gold timer and missile counts | Both ordinary-health fixtures have 180 successful allocations and `96 + 627 reimbursements − 723 decrements = 0`. Missile allocations and connected hits are different observables. The high-health diagnostic is separate: 166 MAME / 149 FBNeo connections at initial HP1024. |
| Roll's 23/24 discrepancy | FBNeo rotating buttons has 23 events before first exit at callback293. Event330 occurs at callback296 with `B0=40`, after the move selector's active state. It is excluded from the first-window score. Its object owner is not established. |
| Ryu against airborne Hulk | Both engines give 16→20 at realized separation325, also325 at first contact, with defender Y466. The MAME standing fixture at the same separation gives20→24 with Y472. Native jump input is configured; no gameplay position writes are used. This supports fixture-dependent extra hits, not the unreproduced22→26 anecdote. |
| Chun-Li's 30 | Both engines show15→30 with60 credit spends and60 ordinary timer decrements. The human initializer stores60 credits and the accepted-input branch spends one instead of decrementing the timer. This proves an input-credit bound in that loop and an observed30-hit fixture; it does not prove30 is the global hit maximum. |

The reviewed opcode-excerpt bytes all match the verified opcode image: **167 instruction rows**. In particular, `$01D18E–$01D1A8` clamps the native combo byte to99, `$04E348` initializes Chun-Li's60 human credits, and `$04E440–$04E454` selects credit spend or timer decrement. Exact image and source hashes are in `core-claim-review.json`.

The authoritative window ends at the first observer-detected move exit. It is a defined measurement interval, not complete ownership tracing for arbitrary delayed projectiles. The selected Gold, Ryu and Chun headline hits all additionally have the expected active selector in their event snapshots. Baseline repetition checks repeatability within a saved fixture; it does not sweep all positions, hurtboxes, input schedules, RNG states or round conditions. A finite search cannot establish a global maximum or infinity.

Reproduce the recount with the extracted evidence package:

```sh
python3 core-claim-review.py /path/to/extracted/core-evidence
```

Optionally pass your own verified decrypted opcode image as a second argument to repeat the short-excerpt byte comparison. The review publishes no ROM image and contains no machine-specific paths.