FA-98331 / Autoregressive target shifting / Open access

A special separator token is predicted as ordinary content · case 01

A special separator token is predicted as ordinary content.

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

ROOT CAUSE

A separator is included as a target despite the contract assigning it only as context boundary.

THE FAILURE

A separator is included as a target despite the contract assigning it only as context boundary.

Unsuccessful approach: Filtering pairs by input token drops predictions after the separator as well.

Case contract

For tokens and sep, return adjacent pairs except any pair whose target equals sep.

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, sep):
    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('separator target skipped',solve([1,9,2],9),[[9,2]])
check('ordinary pairs',solve([1,2,3],9),[[1,2],[2,3]])
check('separator first',solve([9,2,3],9),[[9,2],[2,3]])
check('terminal separator',solve([1,2,9],9),[[1,2]])
check('singleton',solve([9],9),[])
check('empty',solve([],9),[])
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
separator target skipped	`[[1, 9], [9, 2]]`	`[[9, 2]]`	Failed
ordinary pairs	`[[1, 2], [2, 3]]`	`[[1, 2], [2, 3]]`	Passed
separator first	`[[9, 2], [2, 3]]`	`[[9, 2], [2, 3]]`	Passed
terminal separator	`[[1, 2], [2, 9]]`	`[[1, 2]]`	Failed
singleton	`[]`	`[]`	Passed
empty	`[]`	`[]`	Passed

SHA-256 / 784af3fb9b212df0a9517fc765e1d9a9e682e6d3f884200c0b1d0f3a0fdf0f7f

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(tokens, sep):
    return [[tokens[i],tokens[i+1]] for i in range(len(tokens)-1) if tokens[i]!=sep]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('separator target skipped',solve([1,9,2],9),[[9,2]])
check('ordinary pairs',solve([1,2,3],9),[[1,2],[2,3]])
check('separator first',solve([9,2,3],9),[[9,2],[2,3]])
check('terminal separator',solve([1,2,9],9),[[1,2]])
check('singleton',solve([9],9),[])
check('empty',solve([],9),[])
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
separator target skipped	`[[1, 9]]`	`[[9, 2]]`	Failed
ordinary pairs	`[[1, 2], [2, 3]]`	`[[1, 2], [2, 3]]`	Passed
separator first	`[[2, 3]]`	`[[9, 2], [2, 3]]`	Failed
terminal separator	`[[1, 2], [2, 9]]`	`[[1, 2]]`	Failed
singleton	`[]`	`[]`	Passed
empty	`[]`	`[]`	Passed

SHA-256 / 1c171e0b8d01e2638b866fce8eff6812d6201112be533e7367799016368c43a2

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 2026-09-30T16:48:11.933363+00:00.

Case digest / 61028dbbf2a6b691529a4b5a913c6d170d5f7becf18d1f7d8d39e4a21d429bd8