FA-98361 / Token sampling filters / Open access

Token sampling: Epsilon cutoff ignores the candidate count · case 01

Epsilon cutoff ignores the candidate count.

Verified by executionVariant 1 · 6 checks per implementationDownload source bundle ↓JSON ↗

ROOT CAUSE

Epsilon is treated as an absolute cutoff instead of a vocabulary-adjusted cutoff.

THE FAILURE

Epsilon is treated as an absolute cutoff instead of a vocabulary-adjusted cutoff.

Unsuccessful approach: Strict comparison still removes a token whose probability is exactly the adjusted floor.

Case contract

Keep tokens whose probability is at least epsilon divided by vocabulary size; empty input returns empty.

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(probabilities, epsilon):
    return sorted(t for t,p in probabilities.items() if p>=epsilon)
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('vocabulary-scaled floor',solve({'a':.4,'b':.2,'c':.1},.3),['a','b','c'])
check('exact floor',solve({'a':.5,'b':.25,'c':.25},.75),['a','b','c'])
check('empty',solve({},.5),[])
check('singleton',solve({'x':.2},.3),[])
check('zero epsilon',solve({'a':.1,'b':0},0),['a','b'])
check('larger vocab',solve({'a':.5,'b':.2,'c':.1,'d':.05},.2),['a','b','c','d'])
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
vocabulary-scaled floor	`['a']`	`['a', 'b', 'c']`	Failed
exact floor	`[]`	`['a', 'b', 'c']`	Failed
empty	`[]`	`[]`	Passed
singleton	`[]`	`[]`	Passed
zero epsilon	`['a', 'b']`	`['a', 'b']`	Passed
larger vocab	`['a', 'b']`	`['a', 'b', 'c', 'd']`	Failed

SHA-256 / 4b289d6daca67866f176920b5eea8f6a4ef41c6f28e2d967f62ddf72f2abace1

2 / The unsuccessful fix

Exit 1

"""Failure Map reference implementation. Python standard library only."""
import json
import math
N = 1
observations = []
def solve(probabilities, epsilon):
    floor=epsilon/max(1,len(probabilities))
    return sorted(t for t,p in probabilities.items() if p>floor)
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('vocabulary-scaled floor',solve({'a':.4,'b':.2,'c':.1},.3),['a','b','c'])
check('exact floor',solve({'a':.5,'b':.25,'c':.25},.75),['a','b','c'])
check('empty',solve({},.5),[])
check('singleton',solve({'x':.2},.3),[])
check('zero epsilon',solve({'a':.1,'b':0},0),['a','b'])
check('larger vocab',solve({'a':.5,'b':.2,'c':.1,'d':.05},.2),['a','b','c','d'])
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
vocabulary-scaled floor	`['a', 'b', 'c']`	`['a', 'b', 'c']`	Passed
exact floor	`['a']`	`['a', 'b', 'c']`	Failed
empty	`[]`	`[]`	Passed
singleton	`[]`	`[]`	Passed
zero epsilon	`['a']`	`['a', 'b']`	Failed
larger vocab	`['a', 'b', 'c']`	`['a', 'b', 'c', 'd']`	Failed

SHA-256 / a21498749c37763ffb155f086b666a31f9c594672790ebfc01f9b313ee4704bc

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.

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 2026-09-30T16:48:12.371166+00:00.

Case digest / a14325128e9d4e8462aa86dd50907ceded9b28a84326243e4a8a199e8c99aaca