FA-98126 / Sequence batch packing / Open access
Prompt tokens leak into response-only truncation · case 01
Prompt tokens leak into response-only truncation.
ROOT CAUSE
Truncation is applied to the concatenated prompt and response, so prompt tokens can leak into the response-only result.
THE FAILURE
Truncation is applied to the concatenated prompt and response, so prompt tokens can leak into the response-only result.
Unsuccessful approach: Taking a response prefix removes the latest response supervision even though it no longer leaks prompt tokens.
Case contract
Return response tokens only, preserving the full response when it fits and otherwise retaining its final limit tokens; output (tokens, response_start), where response_start is zero.
Why this case matters
This deterministic toy packing rule models a common boundary in preparing bounded training sequences.
1 / The failure
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
N = 1
observations = []
def solve(prompt, response, limit):
joined=prompt+response; kept=joined[-limit:] if limit else []; return kept,max(0,len(prompt)-max(0,len(joined)-limit))
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
p=list(range(N+2)); r=[8,9]
check('response exceeds limit', solve(p,r,1), ([9],0))
check('response fits', solve(p,r,N+4), (r,0))
check('zero limit', solve(p,r,0), ([],0))
check('empty response', solve(p,[],N), ([],0))
check('single final token', solve([1],[7],1), ([7],0))
check('several response tokens', solve([1],[2,3,4],2), ([3,4],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 |
|---|---|---|---|
| response exceeds limit | [[9], 0] | [[9], 0] | Passed |
| response fits | [[0, 1, 2, 8, 9], 3] | [[8, 9], 0] | Failed |
| zero limit | [[], 0] | [[], 0] | Passed |
| empty response | [[2], 1] | [[], 0] | Failed |
| single final token | [[7], 0] | [[7], 0] | Passed |
| several response tokens | [[3, 4], 0] | [[3, 4], 0] | Passed |
SHA-256 / f415ae06f8e9d683aac3f268d7c875a22b3da284a9783ca58d6958c8624edb88
2 / The unsuccessful fix
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
N = 1
observations = []
def solve(prompt, response, limit):
kept=response[:limit] if limit else []; return kept,0
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
p=list(range(N+2)); r=[8,9]
check('response exceeds limit', solve(p,r,1), ([9],0))
check('response fits', solve(p,r,N+4), (r,0))
check('zero limit', solve(p,r,0), ([],0))
check('empty response', solve(p,[],N), ([],0))
check('single final token', solve([1],[7],1), ([7],0))
check('several response tokens', solve([1],[2,3,4],2), ([3,4],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 |
|---|---|---|---|
| response exceeds limit | [[8], 0] | [[9], 0] | Failed |
| response fits | [[8, 9], 0] | [[8, 9], 0] | Passed |
| zero limit | [[], 0] | [[], 0] | Passed |
| empty response | [[], 0] | [[], 0] | Passed |
| single final token | [[7], 0] | [[7], 0] | Passed |
| several response tokens | [[2, 3], 0] | [[3, 4], 0] | Failed |
SHA-256 / 3bb7bda05d446391e0aca4e3d39960c8c34f53ed5d2174cd89a18c1f20f77e95
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.
Sign in to the archive ↗Verification & scope
Inputs are short integer-token lists under the stated toy contract; results do not claim compatibility with a specific training framework. 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 2026-09-30T16:48:09.757529+00:00.
Case digest / 1c4fccee395013c7cc13ff549dece1332ee6cd95073f929dd2224044cf00da4e