File size: 7,649 Bytes
0c1817e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
{
  "schema": "AutonomaFableRouterCurriculum.v2",
  "status": "data_contract_frozen_training_not_authorized_nonrouting",
  "nonRouting": true,
  "trainingAuthorized": false,
  "goalOrder": [
    "capability",
    "succinctness_and_termination",
    "expert_efficiency",
    "throughput"
  ],
  "immutableSources": {
    "host": {
      "repo": "ermiaazarkhalili/LFM2.5-2.6B-SFT-Fable5",
      "revision": "72d68fc"
    },
    "reasoning": {
      "repo": "HelioAI/Claude-Fable-5-5500x",
      "revision": "440267fbdb1b00a40216e7233dbce25530a0ed09"
    },
    "agentCoding": {
      "repo": "greghavens/fable-5-coding-and-debugging-traces",
      "revision": "c63e82adec30798edcbd6e1dcb0014d2b15de236"
    },
    "selectedBanks": {
      "path": "config/benching/fable-donor-selected-banks.v1.json",
      "sha256": "158689cca0679423925cfe3c2eefbd6a248045ed01d2f7dad4d74e05e3ad0d81"
    }
  },
  "dataPolicy": {
    "splitIdentity": {
      "reasoning": "sha256(trimmed prompt)",
      "agentCoding": "source_trajectory_sha256"
    },
    "splitBuckets": {
      "train": "0..899",
      "validation": "900..949",
      "test": "950..999"
    },
    "reasoningDeduplication": "retain the shortest admitted non-empty completion per normalized prompt",
    "historicalTrajectorySampling": "at most the first tool-producing prefix plus the final prefix",
    "attestedTrajectorySampling": "retain all verifier-attested assistant decisions",
    "maximumTokensBeforeTemplate": 1900,
    "finalTemplateMaximumTokens": 2048,
    "finalTrainerMustRetokenizeAndRejectOverflow": true,
    "frozenEvaluationIsolation": "never read sealed/frozen evaluation prompts or answers while building or training"
  },
  "lanes": {
    "host_preservation": {
      "routerEligibility": "preserve",
      "sftWeight": 1.0,
      "baseFableKlWeight": 1.0,
      "expertActivityTarget": "zero unless routed loss beats frozen host by the benefit margin"
    },
    "verified_expert": {
      "routerEligibility": "expert",
      "sftWeight": 1.0,
      "baseFableKlWeight": 0.25,
      "requires": "model_attested=true and non-empty verifier"
    },
    "interaction_pattern": {
      "routerEligibility": "conditional",
      "sftWeight": 0.1,
      "baseFableKlWeight": 0.5,
      "purpose": "tool topology and recovery only; not correctness supervision"
    },
    "loop_negative": {
      "routerEligibility": "inherit_positive",
      "preferenceWeight": 0.5,
      "purpose": "reject repeated completions and duplicate identical tool calls"
    }
  },
  "routingObjective": {
    "counterfactual": "compare frozen-host token NLL with independently explored routed token NLL; never rely only on the router's current choice",
    "offClass": "host-only is an explicit route target when no sampled expert helps",
    "exploration": {
      "expertScaleDuringDiscovery": "initialize at 0.005 and cap at 0.025 so incorrect trials cannot materially override the host",
      "oracleProbeScales": [
        0.025,
        0.05,
        0.1
      ],
      "probeIsolation": "counterfactual probe branches are detached from the served/main output; only the low-scale chosen branch can affect host logits during discovery",
      "candidateExpertsPerEligibleToken": 4,
      "sampling": "stratified without replacement across steps and layers, independent of router logits; half uniform coverage and half frozen-profile-prior sampling",
      "coverageBeforePruning": "every selected expert must receive counterfactual trials on validation-eligible token classes",
      "forcedExplorationProbability": {
        "initial": 0.25,
        "final": 0.05,
        "annealSteps": 300
      },
      "eligibleTokenEntropyBonus": {
        "initial": 0.01,
        "final": 0.0,
        "annealSteps": 300
      },
      "minimumEligibleRouteMass": {
        "initial": 0.15,
        "final": 0.0,
        "annealSteps": 300,
        "scope": "expert-eligible tokens only; never host-preservation tokens"
      },
      "positiveDiscoveryReplay": "retain positive-benefit token/expert pairs in a bounded balanced replay buffer"
    },
    "routerSupervision": {
      "target": "best sampled positive-benefit expert, with host-only off as a competing class",
      "loss": "benefit-weighted ranking/classification plus end-to-end token NLL",
      "negativeExperts": "sampled wrong experts teach relative ranking without requiring their high-impact activation",
      "allOffProtection": "ranking targets and the temporary eligible-token route floor provide gradients even when current router logits prefer off"
    },
    "benefitMarginNats": 0.02,
    "benefitMarginSchedule": "0.0 during discovery, anneal to 0.02 nats over steps 100..300",
    "routeActivation": "after discovery, reward only when routed NLL improves beyond the annealed margin",
    "routeCost": {
      "expertScaleL1": 0.015,
      "activeExpertPenalty": 0.005,
      "targetActiveExpertsPerEligibleToken": "1..2",
      "schedule": "zero for the first 100 discovery steps, then linearly anneal to full weight by step 300"
    },
    "conditionalLoadBalance": {
      "weight": 0.01,
      "applyOnlyTo": "expert-eligible tokens that pass the benefit gate",
      "neverApplyTo": [
        "host_preservation tokens",
        "tokens where the host-only path is equal or better"
      ]
    },
    "antiDeadExpert": {
      "measurement": "usage and marginal NLL improvement per layer, expert, lane, and donor bank",
      "minimumUse": "no global quota; require coverage only among positive-benefit eligible tokens",
      "pruneRule": "remove experts only after completed stratified exploration finds no validation-set marginal benefit; router non-use alone is not evidence",
      "collapseGate": "reject checkpoints that choose off for all eligible tokens while the oracle replay buffer contains positive-benefit routes"
    }
  },
  "behaviorObjective": {
    "stopBoundaryLoss": 0.25,
    "loopUnlikelihood": 0.5,
    "duplicateToolCallPenalty": 0.75,
    "maximumRepeatedNgramRateRegression": 0.0,
    "maximumMedianAnswerLengthRegression": 0.05,
    "maximumToolCallsAcrossFrozenSuite": 24,
    "reject": [
      "verbatim answer repetition",
      "duplicate identical tool calls without an intervening result",
      "continued reasoning after a valid terminal answer",
      "expert activation without measured token-level benefit"
    ]
  },
  "curriculum": [
    {
      "stage": "A-host-anchor",
      "steps": 200,
      "mixture": {
        "host_preservation": 0.75,
        "verified_expert": 0.2,
        "interaction_pattern": 0.05
      },
      "routerScaleMaximum": 0.025
    },
    {
      "stage": "B-benefit-gated-routing",
      "steps": 600,
      "mixture": {
        "host_preservation": 0.3,
        "verified_expert": 0.6,
        "interaction_pattern": 0.1
      },
      "routerScaleMaximum": 0.1
    },
    {
      "stage": "C-succinct-calibration",
      "steps": 200,
      "mixture": {
        "host_preservation": 0.55,
        "verified_expert": 0.35,
        "interaction_pattern": 0.1
      },
      "loopNegativeRatio": 0.25,
      "routerScaleMaximum": 0.1
    }
  ],
  "checkpointSelection": {
    "order": [
      "verified coding and tool correctness",
      "host capability non-regression",
      "succinctness and clean termination",
      "expert-use efficiency",
      "decode throughput"
    ],
    "rejectOn": [
      "any sealed correctness regression beyond the predeclared tolerance",
      "loop or duplicate-tool regression",
      "median verbosity regression above 5%",
      "expert traffic without positive marginal validation benefit"
    ]
  }
}