alex/sots-re
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

lane E3: the prediction's outcome, leaf for leaf

Browse source 
204 -> 158 and 103 -> 87, closed 46 and 16, regressed 0 -- the predicted table
line for line, including the 32-leaf target in full. Three further pairs the
model was never fitted to close 42, 12 and 12 with zero regressions; 128 leaves
across five pairs.

The AFlags-vs-VFlags falsification is decided on a real save rather than in a
comment: running the engine on zuul-turn23-fleet23 reproduces the frozen stamp
at the one system whose active mask is clear while its sticky and explored masks
are not.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01ARBgSooAfokKUy6wKUKEyZ
This commit is contained in:
alex 2026-09-08 13:39:30 -04:00
parent 58c2cd14b4
commit d639c9a112
1 changed files with 53 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

53
findings/subsystems/system-visibility-prediction.md
View file

@ -104,3 +104,56 @@ Each of these is a way the model is wrong and the symptom that would show it.
* `rcex` (6 leaves per pair). Unexplained.
* `NVs` / `PlayerView`. Count 0 on both reference pairs; the `PlayerView` refresh loop is a
separate phase and no leaf of it diverges on either pair.
---
# Measured
Run after the build, against the prediction above, which was committed in `55ea86f` before
any engine file existed. Engine branch `wip/visibility`, host `ctest` 46/46, clean-room OK,
shim cross-built on CT111 with exports byte-identical.
| pair | baseline | after | closed | regressed |
|---|---:|---:|---:|---:|
| `turn1-state -> turn2-state` | 204 | **158** | **46** | **0** |
| `turn2-state -> turn3-state` | 103 | **87** | **16** | **0** |
Closed, by leaf name — the prediction's table, line for line:
```
turn1 -> turn2 EPid 8 ETS 8 Eid 8 NVE 8 (= the brief's 32) ltis 8 EFlags 6
turn2 -> turn3 ETS 8 ltis 8
```
`tools/standalone_report.py`'s own numbers, whose baseline is the do-nothing distance from
the input save rather than from main, are **209 -> 158 (closed 51, regressed 0)** and
**108 -> 87 (closed 21, regressed 0)**; the five-leaf difference in each is what main had
already closed.
## Three pairs the model was never fitted to
Rule 2's strong form. None of these was looked at while the model was being built; the
`human-*` pair is a different game on a different map with a different species mix.
| pair | baseline | after | closed | regressed |
|---|---:|---:|---:|---:|
| `human-turn2-orders -> human-turn3-noderoute` | 353 | 311 | 42 (21 `ETS`, 21 `ltis`) | 0 |
| `zuul-turn15-orders -> zuul-turn16-noderoute` | 276 | 264 | 12 (6 `ETS`, 6 `ltis`) | 0 |
| `zuul-turn16-noderoute -> zuul-turn17-orders2` | 341 | 329 | 12 (6 `ETS`, 6 `ltis`) | 0 |
**128 leaves closed across five pairs, 0 regressed.**
## The falsification that mattered
The `AFlags`-vs-`VFlags` case is the one the reference pairs could not decide, and it is
decided. Running the engine on `zuul-turn23-fleet23.sav` reproduces the freeze on the save
itself: every observed system's stamp moves to 24 and Bismol's — sticky mask 2, explored
mask 2, **active mask 0** — stays at 22, in both `ETS` and `ltis`. A `VFlags` or `EFlags`
model moves it to 24. That row is also a host test (`tests/game_sim/test_visibility.cpp`,
`test_bismol_freeze`), which asserts the freeze *and* asserts what the wrong gate would have
produced, so a later edit that swaps the mask fails loudly rather than quietly agreeing with
the five pairs above.
None of the other falsification cases fired. The remaining ones — the two `Eid` derivations
and the map ordering — are exactly the ones the section marked "**symptom:** none available",
and they are still not settled; §8 of the model doc names the workloads.