FA-98251 / Autoregressive target shifting / Open access

Start marker context is copied as its own target · case 01

Start marker context is copied as its own target.

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

ROOT CAUSE

The target at each position is copied from the same augmented-sequence index instead of the successor index.

THE FAILURE

The target at each position is copied from the same augmented-sequence index instead of the successor index.

Unsuccessful approach: Dropping the marker entirely misaligns later positions when it should provide context.

Case contract

For observed tokens and bos, return [context,target] pairs beginning [bos, first token]; empty observations produce no pairs.

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, bos):
    s=[bos]+tokens
    return [[s[i],s[i]] for i in range(len(tokens))]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('first target',solve([N,N+1],0),[[0,N],[N,N+1]])
check('one observed',solve([8],3),[[3,8]])
check('empty',solve([],2),[])
check('marker differs',solve([4,5],9),[[9,4],[4,5]])
check('repeated token',solve([2,2],1),[[1,2],[2,2]])
check('long stream',solve([1,2,3,4],7),[[7,1],[1,2],[2,3],[3,4]])
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
first target	`[[0, 0], [1, 1]]`	`[[0, 1], [1, 2]]`	Failed
one observed	`[[3, 3]]`	`[[3, 8]]`	Failed
empty	`[]`	`[]`	Passed
marker differs	`[[9, 9], [4, 4]]`	`[[9, 4], [4, 5]]`	Failed
repeated token	`[[1, 1], [2, 2]]`	`[[1, 2], [2, 2]]`	Failed
long stream	`[[7, 7], [1, 1], [2, 2], [3, 3]]`	`[[7, 1], [1, 2], [2, 3], [3, 4]]`	Failed

SHA-256 / eb6c56ede1f38be1d1dbcb23e994bbb1b04c82b4e653926774dfa7099d7d0792

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(tokens, bos):
    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('first target',solve([N,N+1],0),[[0,N],[N,N+1]])
check('one observed',solve([8],3),[[3,8]])
check('empty',solve([],2),[])
check('marker differs',solve([4,5],9),[[9,4],[4,5]])
check('repeated token',solve([2,2],1),[[1,2],[2,2]])
check('long stream',solve([1,2,3,4],7),[[7,1],[1,2],[2,3],[3,4]])
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
first target	`[[1, 2]]`	`[[0, 1], [1, 2]]`	Failed
one observed	`[]`	`[[3, 8]]`	Failed
empty	`[]`	`[]`	Passed
marker differs	`[[4, 5]]`	`[[9, 4], [4, 5]]`	Failed
repeated token	`[[2, 2]]`	`[[1, 2], [2, 2]]`	Failed
long stream	`[[1, 2], [2, 3], [3, 4]]`	`[[7, 1], [1, 2], [2, 3], [3, 4]]`	Failed

SHA-256 / 94f8a746603a65de983da0ea286f27b71a181261a71f8c3061a5f68ade7200ad

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 2026-09-30T16:48:11.053892+00:00.

Case digest / 5e1aa9cc5084a8381c279e0b5dca2db5dfb8fa4e26b3b1a3958bbf28716bf4f6