FA-98306 / Autoregressive target shifting / Open access

Target gather uses the context position instead of its successor · case 01

Target gather uses the context position instead of its successor.

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

ROOT CAUSE

A lookup reads the target stored at the context position rather than the next absolute position.

THE FAILURE

A lookup reads the target stored at the context position rather than the next absolute position.

Unsuccessful approach: Using the enumeration offset as a target key fails when the context stream starts away from zero.

Case contract

Contexts are [absolute_position,input_token] rows and targets maps absolute positions to tokens. Return [input,target_at_position_plus_one] for each context row.

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(contexts, targets):
    return [[token,targets[position]] for position,token in contexts]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('shifted lookup',solve([[4,10],[5,11]],{4:90,5:12,6:13}),[[10,12],[11,13]])
check('nonzero origin',solve([[N,7]],{N:7,N+1:8}),[[7,8]])
check('negative positions',solve([[-2,1],[-1,2]],{-2:1,-1:2,0:3}),[[1,2],[2,3]])
check('empty contexts',solve([],{1:8}),[])
check('single context',solve([[20,4]],{20:4,21:9}),[[4,9]])
check('target differs from input',solve([[0,2]],{0:2,1:8}),[[2,8]])
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
shifted lookup	`[[10, 90], [11, 12]]`	`[[10, 12], [11, 13]]`	Failed
nonzero origin	`[[7, 7]]`	`[[7, 8]]`	Failed
negative positions	`[[1, 1], [2, 2]]`	`[[1, 2], [2, 3]]`	Failed
empty contexts	`[]`	`[]`	Passed
single context	`[[4, 4]]`	`[[4, 9]]`	Failed
target differs from input	`[[2, 2]]`	`[[2, 8]]`	Failed

SHA-256 / 630494ba666db7af7ad3592a6db7b36a65b3a5bf497f248eb2394bcabf1e1c0d

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(contexts, targets):
    return [[token,targets.get(i+1,targets.get(position,-1))] for i,(position,token) in enumerate(contexts)]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('shifted lookup',solve([[4,10],[5,11]],{4:90,5:12,6:13}),[[10,12],[11,13]])
check('nonzero origin',solve([[N,7]],{N:7,N+1:8}),[[7,8]])
check('negative positions',solve([[-2,1],[-1,2]],{-2:1,-1:2,0:3}),[[1,2],[2,3]])
check('empty contexts',solve([],{1:8}),[])
check('single context',solve([[20,4]],{20:4,21:9}),[[4,9]])
check('target differs from input',solve([[0,2]],{0:2,1:8}),[[2,8]])
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
shifted lookup	`[[10, 90], [11, 12]]`	`[[10, 12], [11, 13]]`	Failed
nonzero origin	`[[7, 7]]`	`[[7, 8]]`	Failed
negative positions	`[[1, 1], [2, 2]]`	`[[1, 2], [2, 3]]`	Failed
empty contexts	`[]`	`[]`	Passed
single context	`[[4, 4]]`	`[[4, 9]]`	Failed
target differs from input	`[[2, 8]]`	`[[2, 8]]`	Passed

SHA-256 / bc75986eccd428f68867ea7d7a351ceb986a2aadbd2b84e2fc55d4f0cc44863f

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 2026-09-30T16:48:11.500424+00:00.

Case digest / 009f4b74d6d757080f91a4396aabc978e4b8ca610201b0ed92e7c4a8aae004a0