FOLD
97% recall
strong
Optional research history
Same hand, different machinery
These six controlled inherited-hero stages compare capabilities. Classifiers choose action labels; transformer language models generate Pokerese. A later architecture is not guaranteed to give better advice on every hand. Historical artifacts remain preserved.
Same hand
One fixed poker decision
Every stage receives the same poker decision. Only the learning machinery changes.
Teacher BET1.89 BBtest
Street
flop
Position
CO
Hero cards
5h 5d
Board
Td 9h Kh
Pot
5.72 BB
Stack
19.25 BB
Legal actions
FOLDCHECKBET 1 BB-19.25 BBRAISE to 1 BB-19.25 BB
Action context
UTG FOLDUTG+1 FOLDMP1 FOLDMP2 FOLDCO RAISE to 2.31 BBBTN FOLDSB FOLDBB CALL 1.21 BBBB CHECK
Across stages: machinery changes.
Different machinery
Most common action rule
No weights are learned for this hand-level decision; the rule selects the training majority action.
Training action counts point to the most common action.
Different answer
Baseline
DifferentPredicts one direct action from training action frequency.
Predicted action
FOLD
Rule
training majority action
What changed?
This is the starting point: the hand first passes through a non-neural frequency rule.
The model can now combine poker features instead of always choosing the most common action; for this hand it changes to BET.
Within-stage training
Within a stage: training progresses. The primary view shows initial, middle and final evidence for this exact hand.
This non-neural baseline has no epoch replay.
Attention
This stage did not declare attention capture.
The inherited-hero world
Historical v6 diagnostic evaluation
The original final generation sample used a 10-token cap, which can cut off valid sized targets before EOS. Its weights match the current reconstruction, but its decode budget and sample differ. The primary journey reports full-test evaluation with the corrected declared policy. These historical results are preserved, not overwritten.
ActionLearned strengthTeachTeacherGenGeneratedStatus
CHECK
83% recall
strong
CALL
14% recall
faint signal
BET
33% recall
emerging
RAISE
19% recall
faint signal
Epoch 12 validation sample
Action accuracy
61%
Macro F1
0.481
Valid generation
81%
Validation loss
0.264
These are aggregate validation-sample metrics recorded at epoch 12; they are not one prompt being replayed over time.
Teacher vs generated counts
Secondary view of the same selected validation sample.
Teacher countGenerated count
Action accuracy and validation loss by epoch
A compact trace of the exact measures shown in the epoch summary; accuracy uses the left axis as percent, loss uses the right axis.
Action accuracyValidation loss
Original aggregate evidence
{
"bestEpoch": 12,
"finalEpoch": 12,
"validationPerplexity": 1.301608,
"testPerplexity": 1.309724,
"actionAccuracy": 0.648438,
"macroF1": 0.452918,
"validGenerationRate": 0.761719,
"exactSizingAccuracyOnSized": 0,
"sizingMaeCentibbOnDecodable": 620.2,
"actionCollapsePersists": false,
"dominantPredictedAction": "FOLD",
"dominantPredictedActionShare": 0.554688,
"evaluationSampleSize": 256,
"perActionRecall": {
"FOLD": 0.984848,
"CHECK": 0.833333,
"CALL": 0.208333,
"BET": 0.1,
"RAISE": 0.160714
},
"representativeOutputs": [
{
"exampleId": "pokerese_0020260804_T226_0529230991:H00055:D002:teacher_concise",
"street": "flop",
"teacherPokerese": "ACTION CHECK <EOS>",
"generatedPokerese": "ACTION CHECK <EOS>",
"teacherAction": "CHECK",
"generatedAction": "CHECK",
"teacherSizingCentibb": null,
"generatedSizingCentibb": null,
"grammarError": null,
"interpretation": "Valid Pokerese and the action matches the teacher.",
"selectionCategory": "valid_correct_action"
},
{
"exampleId": "pokerese_0020260804_T248_0668491052:H00046:D002:teacher_concise",
"street": "flop",
"teacherPokerese": "ACTION BET SIZE_KIND AMOUNT SIZE_BB <NUM> D9 D0 D3 </NUM> <EOS>",
"generatedPokerese": "ACTION CHECK <EOS>",
"teacherAction": "BET",
"generatedAction": "CHECK",
"teacherSizingCentibb": 903,
"generatedSizingCentibb": null,
"grammarError": null,
"interpretation": "Valid Pokerese, but the model chose CHECK instead of BET.",
"selectionCategory": "valid_wrong_action"
},
{
"exampleId": "pokerese_0020260804_T226_0529230991:H00017:D002:teacher_concise",
"street": "flop",
"teacherPokerese": "ACTION BET SIZE_KIND AMOUNT SIZE_BB <NUM> D1 D8 D9 </NUM> <EOS>",
"generatedPokerese": "ACTION BET SIZE_KIND AMOUNT SIZE_BB <NUM> D1 D8 D9 </NUM>",
"teacherAction": "BET",
"generatedAction": "UNK",
"teacherSizingCentibb": 189,
"generatedSizingCentibb": null,
"grammarError": "<EOS> must occur exactly once at the end",
"interpretation": "The model produced BET tokens, but the sentence was malformed, so it could not be decoded reliably.",
"selectionCategory": "malformed_grammar"
},
{
"exampleId": "pokerese_0020260804_T227_2778673778:H00029:D002:teacher_concise",
"street": "flop",
"teacherPokerese": "ACTION CHECK <EOS>",
"generatedPokerese": "ACTION BET SIZE_KIND AMOUNT SIZE_BB <NUM> D1 D8 </NUM> <EOS>",
"teacherAction": "CHECK",
"generatedAction": "BET",
"teacherSizingCentibb": null,
"generatedSizingCentibb": 18,
"grammarError": null,
"interpretation": "Valid Pokerese, but the model chose BET instead of CHECK.",
"selectionCategory": "valid_active_generation"
}
],
"missingRepresentativeCategories": [
"correct action but wrong or missing size"
]
}