FA-98341 / Token sampling filters / Open access

Token sampling: Top-k drops the final retained candidate · case 01

Top-k drops the final retained candidate.

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

ROOT CAUSE

The cutoff subtracts one from a count that is already a slice length.

THE FAILURE

The cutoff subtracts one from a count that is already a slice length.

Unsuccessful approach: Clamping the erroneous k-1 still excludes one valid candidate.

Case contract

Return the k highest-scoring token IDs, ties by ascending ID; k<=0 returns empty and k beyond vocabulary returns all.

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, k):
    ranked=sorted(scores,key=lambda t:(-scores[t],t))
    return ranked[:max(0,k-1)]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('top two',solve({'a':N+2,'b':N+1,'c':N},2),['a','b'])
check('singleton',solve({'a':N},1),['a'])
check('zero',solve({'a':N},0),[])
check('oversized',solve({'b':N,'a':N},9),['a','b'])
check('tie order',solve({'z':N,'a':N},1),['a'])
check('empty',solve({},2),[])
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
top two	`['a']`	`['a', 'b']`	Failed
singleton	`[]`	`['a']`	Failed
zero	`[]`	`[]`	Passed
oversized	`['a', 'b']`	`['a', 'b']`	Passed
tie order	`[]`	`['a']`	Failed
empty	`[]`	`[]`	Passed

SHA-256 / ead7a2b617ec620bd241b212de4085781dfe758cfe051c1d07f1eb960107de6c

2 / The unsuccessful fix

Exit 1

"""Failure Map reference implementation. Python standard library only."""
import json
import math
N = 1
observations = []
def solve(scores, k):
    ranked=sorted(scores,key=lambda t:(-scores[t],t))
    return ranked[:max(0,min(k-1,len(ranked)))]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('top two',solve({'a':N+2,'b':N+1,'c':N},2),['a','b'])
check('singleton',solve({'a':N},1),['a'])
check('zero',solve({'a':N},0),[])
check('oversized',solve({'b':N,'a':N},9),['a','b'])
check('tie order',solve({'z':N,'a':N},1),['a'])
check('empty',solve({},2),[])
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
top two	`['a']`	`['a', 'b']`	Failed
singleton	`[]`	`['a']`	Failed
zero	`[]`	`[]`	Passed
oversized	`['a', 'b']`	`['a', 'b']`	Passed
tie order	`[]`	`['a']`	Failed
empty	`[]`	`[]`	Passed

SHA-256 / d33875b1ef102eb9318575b3988f312647795b8ad60577e049cd770b9627c5a9

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 2026-09-30T16:48:11.933319+00:00.

Case digest / 588b2c7a5a3a62f28b6c6f8abfcf485450abe6157443443fb7b8be2bfdeef3dd