FA-100816 / Model output stopping rules / Open access

Output stopping: An ordinary end marker is reported as a length finish · case 01

An ordinary end marker is reported as a length finish.

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

ROOT CAUSE

The end marker is checked after the length cap, allowing a boundary EOS to be mislabeled as length.

THE FAILURE

The end marker is checked after the length cap, allowing a boundary EOS to be mislabeled as length.

Unsuccessful approach: Renaming the EOS reason does not use the required canonical reason.

Case contract

Given token stream and end marker, report eos if marker appears, length if token count reaches cap, otherwise null; eos has precedence.

Why this case matters

A deterministic output-control model isolates one finish or streaming boundary from any particular runtime.

1 / The failure

Exit 1

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

N = 1
observations = []
def solve(tokens, end_id, cap):
    if len(tokens)>=cap:return 'length'
    if end_id in tokens:return 'eos'
    return None
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('eos before cap',solve([1,9],9,4),'eos')
check('eos at cap',solve([1,9],9,2),'eos')
check('cap reached',solve([1,2],9,2),'length')
check('under cap',solve([1],9,3),None)
check('empty zero cap',solve([],9,0),'length')
check('end zero',solve([0],0,1),'eos')
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
eos before cap	`eos`	`eos`	Passed
eos at cap	`length`	`eos`	Failed
cap reached	`length`	`length`	Passed
under cap	`None`	`None`	Passed
empty zero cap	`length`	`length`	Passed
end zero	`length`	`eos`	Failed

SHA-256 / 8fdebcce7c61d935e8ee38a90be58199839255226132c8189f74e1630ebc0af3

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(tokens, end_id, cap):
    if end_id in tokens:return 'end'
    if len(tokens)>=cap:return 'length'
    return None
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('eos before cap',solve([1,9],9,4),'eos')
check('eos at cap',solve([1,9],9,2),'eos')
check('cap reached',solve([1,2],9,2),'length')
check('under cap',solve([1],9,3),None)
check('empty zero cap',solve([],9,0),'length')
check('end zero',solve([0],0,1),'eos')
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
eos before cap	`end`	`eos`	Failed
eos at cap	`end`	`eos`	Failed
cap reached	`length`	`length`	Passed
under cap	`None`	`None`	Passed
empty zero cap	`length`	`length`	Passed
end zero	`end`	`eos`	Failed

SHA-256 / 94dbaa4c6fad97285b915b9faaf1621b1612bed3be656b7ac3ef10b46a1f9d91

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

This toy contract covers the supplied token records only; it does not claim compatibility with a particular model runtime or tokenizer. 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 2026-09-30T16:48:40.054596+00:00.

Case digest / 0fb4c41ebb4938be56dd88bd916f1c0a9c677b2b536e94cf358cbb68fe3e1e93