FA-98261 / Autoregressive target shifting / Open access

End marker is omitted from next-token supervision · case 01

End marker is omitted from next-token supervision.

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

ROOT CAUSE

The target stream ends at the final observed token.

THE FAILURE

The target stream ends at the final observed token.

Unsuccessful approach: Appending EOS but starting at index one drops the first observed-token target pair.

Case contract

Return [input,target] pairs for each token followed by eos; empty tokens yield no pair.

Why this case matters

A small deterministic token-stream contract makes next-token alignment errors inspectable.

1 / The failure

Exit 1

"""Failure Map reference implementation. Python standard library only."""
import json

N = 1
observations = []
def solve(tokens, eos):
    return [[tokens[i],tokens[i+1]] for i in range(len(tokens)-1)]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('last target eos',solve([N,N+1],99),[[N,N+1],[N+1,99]])
check('one token',solve([4],7),[[4,7]])
check('empty',solve([],7),[])
check('eos equals token',solve([3,3],3),[[3,3],[3,3]])
check('unique tail',solve([1,2,8],5),[[1,2],[2,8],[8,5]])
check('negative token',solve([-1,0],4),[[-1,0],[0,4]])
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
last target eos	`[[1, 2]]`	`[[1, 2], [2, 99]]`	Failed
one token	`[]`	`[[4, 7]]`	Failed
empty	`[]`	`[]`	Passed
eos equals token	`[[3, 3]]`	`[[3, 3], [3, 3]]`	Failed
unique tail	`[[1, 2], [2, 8]]`	`[[1, 2], [2, 8], [8, 5]]`	Failed
negative token	`[[-1, 0]]`	`[[-1, 0], [0, 4]]`	Failed

SHA-256 / 2bd3de120a1fdefb0f4e79d1fa8d191ffa8df461a572f473d92b84c74c523435

2 / The unsuccessful fix

Exit 1

"""Failure Map reference implementation. Python standard library only."""
import json

N = 1
observations = []
def solve(tokens, eos):
    s=tokens+[eos]
    return [[s[i],s[i+1]] for i in range(1,len(s)-1)]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('last target eos',solve([N,N+1],99),[[N,N+1],[N+1,99]])
check('one token',solve([4],7),[[4,7]])
check('empty',solve([],7),[])
check('eos equals token',solve([3,3],3),[[3,3],[3,3]])
check('unique tail',solve([1,2,8],5),[[1,2],[2,8],[8,5]])
check('negative token',solve([-1,0],4),[[-1,0],[0,4]])
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
last target eos	`[[2, 99]]`	`[[1, 2], [2, 99]]`	Failed
one token	`[]`	`[[4, 7]]`	Failed
empty	`[]`	`[]`	Passed
eos equals token	`[[3, 3]]`	`[[3, 3], [3, 3]]`	Failed
unique tail	`[[2, 8], [8, 5]]`	`[[1, 2], [2, 8], [8, 5]]`	Failed
negative token	`[[0, 4]]`	`[[-1, 0], [0, 4]]`	Failed

SHA-256 / cd7d206ac643a503b3e6f7f7bc666dec01323fb13d69c7c01562f9e443bb1aa3

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

Inputs are toy integer token sequences; this contract says nothing about model quality, batching, sampling, or production training frameworks. 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 2026-09-30T16:48:11.053612+00:00.

Case digest / ec707ea717eec0c76f9bba60c72f816f6d2797f8e02fc4816d190fe8b8fa5bbc