# Retained AI images and later human trials — 2026-11-01

The archive contains **241 distinct retained 3 MB ROM images**.148 have linked
native CPU test series,244 series in total; every recorded production hash was
found exactly on disk. These are image hashes, not a claim of 241 different AI
behaviors. Normalizing only title120..17F/checksum18E..18F gives 238 byte groups.
147 historical candidate/control images form the review queue, with 147 such
byte-group representatives; the baseline-only image and unlinked diagnostics
remain separately cataloged. No image was changed or human trial started.

[Complete JSON inventory](tools/notes/RETAINED_AI_VARIANTS.json) ·
[CSV inventory](tools/notes/RETAINED_AI_VARIANTS.csv).

Each image retains its exact SHA256, profile names, every independently recorded
series, complete50-game outcomes, home/away counts, native-input protocol and
original production/benchmark/observer/runner/manifest/summary hashes where
present. Original statistics-audit errors remain attached. Four historical
rejection/outcome reports are summarized; invalid early layouts/call sites and
the later failed strength target are not silently promoted to working releases.
The complete assembler-source binding is not invented where it is absent.

An early 43-win/116-goal single batch remains visible. Its contemporaneous record
still required a final release gate and the pool lacks a recorded statistics
audit. It is not stable 39 evidence, the accepted release or a human difficulty
result. The completed accepted Hard base remains the 38-win/132–66 version.

## Orderly human review

Review the original exact image's build, source history, protocol and known
regressions first. These older experiments predate later fixes and menus and
have not each passed the current100-suite matrix. Keep the historical image/results
unchanged. If a compatible current trial requires porting its AI changes, give
that trial its own new hash/version and required regression results; the old
CPU batch does not validate the port. Deliver only one current trial at a time
under the existing fresh-progress and controller-preservation policy, after
other requested work. Record jdc's actual difficulty/passing/shooting/response
feedback before the next. No automatic launch or new strength search.

The following queue prefix sorts by the most goals in a **recorded complete
50-game batch**, purely to locate the goal-heavy images jdc asked to try. This
is not a stability, average performance or human-difficulty ranking. Every
other result for each image remains in the JSON.

| Exact image ID | Archived profile | That batch's wins | Goals for–against |
|---|---|---:|---:|
| sha-fdf88b7f1933 | aim-open16 | 38/50 | 132–66 |
| sha-c35a4bacffe2 | team-flight-bridge160 | 36/50 | 122–68 |
| sha-8727023b8585 | relay-pointcarry | 34/50 | 118–61 |
| sha-4e4c907b06f8 | bridge160 | 37/50 | 117–62 |
| sha-7d733755555e | team-flight | 33/50 | 117–71 |
| sha-1efe7a63cb9a | quickturn-freecross | 36/50 | 116–63 |

The queue is prepared. Human comparisons remain pending until jdc plays them;
no feedback, preferences or successful historical regressions are fabricated.

Finalr5 audit rechecked all241 original catalog image hashes: every image remains
exactly present. This confirms retention, not current-regression qualification
or human preference. See MOMENTUM_LEFT_RESULT.json.