FA-98401 / Token sampling filters / Open access
Token sampling: Token bias is applied after repetition adjustment · case 01
Token bias is applied after repetition adjustment.
ROOT CAUSE
The repeated-token transform branches on the raw logit and bias is added afterward.
THE FAILURE
The repeated-token transform branches on the raw logit and bias is added afterward.
Unsuccessful approach: Dividing biased scores regardless of sign mishandles a bias that makes a negative logit remain negative.
Case contract
Add per-token bias to base logits first, then apply the sign-aware repetition transform to seen tokens.
Why this case matters
A finite, deterministic token-selection contract isolates one decoding operation from model inference and runtime behavior.
1 / The failure
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
import math
N = 1
observations = []
def solve(scores, bias, seen, penalty):
return {t:((v/penalty if v>=0 else v*penalty) if t in seen else v)+bias.get(t,0) for t,v in scores.items()}
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('bias changes sign branch',solve({'a':-2.0},{'a':5.0},['a'],2),{'a':1.5})
check('unseen bias',solve({'a':N},{'a':3.0},[],2),{'a':N+3.0})
check('negative biased logit',solve({'a':-5.0},{'a':1.0},['a'],3),{'a':-12.0})
check('no bias',solve({'a':4.0},{},['a'],2),{'a':2.0})
check('empty',solve({}, {'a':N},['a'],2),{})
check('zero score bias',solve({'a':0.0},{'a':2.0},['a'],2),{'a':1.0})
print(json.dumps({"observations": observations, "passed": all(x["passed"] for x in observations)}, ensure_ascii=False))
raise SystemExit(0 if all(x["passed"] for x in observations) else 1)
| Boundary fixture | Actual | Expected | Outcome |
|---|---|---|---|
| bias changes sign branch | {'a': 1.0} | {'a': 1.5} | Failed |
| unseen bias | {'a': 4.0} | {'a': 4.0} | Passed |
| negative biased logit | {'a': -14.0} | {'a': -12.0} | Failed |
| no bias | {'a': 2.0} | {'a': 2.0} | Passed |
| empty | {} | {} | Passed |
| zero score bias | {'a': 2.0} | {'a': 1.0} | Failed |
SHA-256 / 33328b4428d82f4ce7a795b2a2289bca70837528a439991184301d5f16350be2
2 / The unsuccessful fix
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
import math
N = 1
observations = []
def solve(scores, bias, seen, penalty):
out={}
for t,v in scores.items():
b=v+bias.get(t,0); out[t]=(b/penalty if t in seen else b)
return out
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('bias changes sign branch',solve({'a':-2.0},{'a':5.0},['a'],2),{'a':1.5})
check('unseen bias',solve({'a':N},{'a':3.0},[],2),{'a':N+3.0})
check('negative biased logit',solve({'a':-5.0},{'a':1.0},['a'],3),{'a':-12.0})
check('no bias',solve({'a':4.0},{},['a'],2),{'a':2.0})
check('empty',solve({}, {'a':N},['a'],2),{})
check('zero score bias',solve({'a':0.0},{'a':2.0},['a'],2),{'a':1.0})
print(json.dumps({"observations": observations, "passed": all(x["passed"] for x in observations)}, ensure_ascii=False))
raise SystemExit(0 if all(x["passed"] for x in observations) else 1)
| Boundary fixture | Actual | Expected | Outcome |
|---|---|---|---|
| bias changes sign branch | {'a': 1.5} | {'a': 1.5} | Passed |
| unseen bias | {'a': 4.0} | {'a': 4.0} | Passed |
| negative biased logit | {'a': -1.3333333333333333} | {'a': -12.0} | Failed |
| no bias | {'a': 2.0} | {'a': 2.0} | Passed |
| empty | {} | {} | Passed |
| zero score bias | {'a': 1.0} | {'a': 1.0} | Passed |
SHA-256 / bb293246d63ec4c17fad2e2e26aeffc77a57017054c4565d54402466d8b64a22
HELD IN THE MEMBER ARCHIVE
The verified repair and its recorded checks are member-only.
This mechanism has 6 recorded checks per implementation. The open-access tier publishes the failure and the unsuccessful fix; the repaired source that passes every check, and the observations that prove it, are available to members.
Every case sharing this mechanism uses the same contract and the same repair, so this one record is held back for all of them.
Member access is invitation-based. Sign in with your invited account to inspect the repair.
Sign in to the archive ↗Verification & scope
These are toy finite-vocabulary contracts with supplied scores, histories, and draws; they do not claim production model quality or compatibility. This reproducer isolates one failure mechanism. Results cover the supplied fixtures. Variants within a family share a test contract and should remain grouped when constructing evaluation splits. Related mechanisms with a shared evaluation_group must also remain together; these controlled models are not independent production incidents.
Observations recorded using Python 3.12.14 at 2026-09-30T16:48:12.802654+00:00.
Case digest / d31f4575878b53ac07fd269e1d7be2517c4cb908d42745720c359ab830287374