FA-98281 / Autoregressive target shifting / Open access

A context window predicts each token from itself · case 01

A context window predicts each token from itself.

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

ROOT CAUSE

The target index is copied from the input index instead of applying the one-token causal offset.

THE FAILURE

The target index is copied from the input index instead of applying the one-token causal offset.

Unsuccessful approach: Adding a lookahead to the target slice but retaining the old input start skips a context token.

Case contract

For tokens and start, return adjacent pairs whose input indexes are start through len(tokens)-2; negative start clamps to zero.

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, start):
    return [[tokens[i],tokens[i]] for i in range(max(0,start),len(tokens)-1)]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('middle window',solve([0,1,2,3,4],2),[[2,3],[3,4]])
check('start zero',solve([N,N+1,N+2],0),[[N,N+1],[N+1,N+2]])
check('negative clamps',solve([1,2,3],-4),[[1,2],[2,3]])
check('last pair',solve([1,2,3,4],2),[[3,4]])
check('past end',solve([1,2,3],8),[])
check('short',solve([1],0),[])
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
middle window	`[[2, 2], [3, 3]]`	`[[2, 3], [3, 4]]`	Failed
start zero	`[[1, 1], [2, 2]]`	`[[1, 2], [2, 3]]`	Failed
negative clamps	`[[1, 1], [2, 2]]`	`[[1, 2], [2, 3]]`	Failed
last pair	`[[3, 3]]`	`[[3, 4]]`	Failed
past end	`[]`	`[]`	Passed
short	`[]`	`[]`	Passed

SHA-256 / 942b99cb10e095742a18c2701d6da93c52a5e84c7d5a97b8c5d17cb425b92514

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(tokens, start):
    return [[tokens[i+1],tokens[i+2]] for i in range(max(0,start),len(tokens)-2)]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('middle window',solve([0,1,2,3,4],2),[[2,3],[3,4]])
check('start zero',solve([N,N+1,N+2],0),[[N,N+1],[N+1,N+2]])
check('negative clamps',solve([1,2,3],-4),[[1,2],[2,3]])
check('last pair',solve([1,2,3,4],2),[[3,4]])
check('past end',solve([1,2,3],8),[])
check('short',solve([1],0),[])
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
middle window	`[[3, 4]]`	`[[2, 3], [3, 4]]`	Failed
start zero	`[[2, 3]]`	`[[1, 2], [2, 3]]`	Failed
negative clamps	`[[2, 3]]`	`[[1, 2], [2, 3]]`	Failed
last pair	`[]`	`[[3, 4]]`	Failed
past end	`[]`	`[]`	Passed
short	`[]`	`[]`	Passed

SHA-256 / e5e5f1ba5a26b87d498ebfd8606a95d578f06846121421c6c5e54545fce29d9f

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 2026-09-30T16:48:11.500501+00:00.

Case digest / 0998609399371d49722d6e0cc7f73861b99b54f29931722b3a0ffd85cde7db2b