FA-98531 / Beam search bookkeeping / Open access

Beam search: length counter advances for prompt positions · case 01

length counter advances for prompt positions.

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

ROOT CAUSE

The bookkeeping count starts at the input prompt length rather than at zero generated tokens.

THE FAILURE

The bookkeeping count starts at the input prompt length rather than at zero generated tokens.

Unsuccessful approach: Subtracting one from total sequence size only works for a one-token prompt.

Case contract

Given prompt token count and generated token batches, return generated length after each step.

Why this case matters

Beam decoders rely on consistent hypothesis state while expanding and ranking candidates.

1 / The failure

Exit 1

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

N = 1
observations = []
def solve(prompt_length, batches):
    return [prompt_length+sum(batches[:i+1]) for i in range(len(batches))]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('prompt excluded',solve(N,[1,1,1]),[1,2,3])
check('multi-token append',solve(2*N,[2,1]),[2,3])
check('no generated tokens',solve(N,[]),[])
check('zero prefix',solve(0,[1]),[1])
check('empty append step',solve(N,[0,1]),[0,1])
check('varying batch sizes',solve(7,[N,2]),[N,N+2])
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
prompt excluded	`[2, 3, 4]`	`[1, 2, 3]`	Failed
multi-token append	`[4, 5]`	`[2, 3]`	Failed
no generated tokens	`[]`	`[]`	Passed
zero prefix	`[1]`	`[1]`	Passed
empty append step	`[1, 2]`	`[0, 1]`	Failed
varying batch sizes	`[8, 10]`	`[1, 3]`	Failed

SHA-256 / c19a7f491f9baa313707d7d51468628ba1eec2172f8bb59397b526d823f2d27a

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(prompt_length, batches):
    return [prompt_length+sum(batches[:i+1])-1 for i in range(len(batches))]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('prompt excluded',solve(N,[1,1,1]),[1,2,3])
check('multi-token append',solve(2*N,[2,1]),[2,3])
check('no generated tokens',solve(N,[]),[])
check('zero prefix',solve(0,[1]),[1])
check('empty append step',solve(N,[0,1]),[0,1])
check('varying batch sizes',solve(7,[N,2]),[N,N+2])
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
prompt excluded	`[1, 2, 3]`	`[1, 2, 3]`	Passed
multi-token append	`[3, 4]`	`[2, 3]`	Failed
no generated tokens	`[]`	`[]`	Passed
zero prefix	`[0]`	`[1]`	Failed
empty append step	`[0, 1]`	`[0, 1]`	Passed
varying batch sizes	`[7, 9]`	`[1, 3]`	Failed

SHA-256 / bb6855008090f5ec0053d8ec842b30dae1bef934647dca697531e1ae498837b4

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 is a finite toy beam-state contract and makes no claim about neural model quality or production decoding. 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 2026-09-30T16:48:14.106361+00:00.

Case digest / 794ff6b0d48330ee073d39476b50e13a8e04a7cc673b12b44774bd9de8f48864