FA-100661 / Speculative decoding bookkeeping / Open access

Decoding: Rollback removes only the final draft token · case 01

Rollback removes only the final draft token.

Verified by executionVariant 1 · 7 checks per implementationDownload source bundle ↓JSON ↗

ROOT CAUSE

Rollback removes only the single rejected token.

THE FAILURE

Rollback removes only the single rejected token.

Unsuccessful approach: Removing the last token loses a valid suffix whenever rejection is earlier.

Case contract

Keep the first k draft tokens and discard the rejected token plus all later draft tokens.

Why this case matters

A self-contained bookkeeping model makes prefix and token-count semantics observable without a model runtime.

1 / The failure

Exit 1

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

N = 1
observations = []
def solve(draft, k):
    return draft[:-1]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('middle rollback', solve([1,2,3,4],2),[1,2])
check('reject first', solve([1,2],0),[])
check('accept all', solve([1,2],2),[1,2])
check('empty draft', solve([],0),[])
check('singleton rejected', solve([N],0),[])
check('keep one', solve([1,2,3],1),[1])
check('drop long suffix', solve([0,1,2,3,4],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 rollback	`[1, 2, 3]`	`[1, 2]`	Failed
reject first	`[1]`	`[]`	Failed
accept all	`[1]`	`[1, 2]`	Failed
empty draft	`[]`	`[]`	Passed
singleton rejected	`[]`	`[]`	Passed
keep one	`[1, 2]`	`[1]`	Failed
drop long suffix	`[0, 1, 2, 3]`	`[0]`	Failed

SHA-256 / 00c49cd26c5f7eee165f7e68883fa6e6b6a9c9fee102c9946f9755aa53250c4a

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(draft, k):
    return draft[:k+1]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('middle rollback', solve([1,2,3,4],2),[1,2])
check('reject first', solve([1,2],0),[])
check('accept all', solve([1,2],2),[1,2])
check('empty draft', solve([],0),[])
check('singleton rejected', solve([N],0),[])
check('keep one', solve([1,2,3],1),[1])
check('drop long suffix', solve([0,1,2,3,4],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 rollback	`[1, 2, 3]`	`[1, 2]`	Failed
reject first	`[1]`	`[]`	Failed
accept all	`[1, 2]`	`[1, 2]`	Passed
empty draft	`[]`	`[]`	Passed
singleton rejected	`[1]`	`[]`	Failed
keep one	`[1, 2]`	`[1]`	Failed
drop long suffix	`[0, 1]`	`[0]`	Failed

SHA-256 / 7372ece63bd513fa54f4592c52e837406613655367dbddf729fb5d6dff3fc19f

HELD IN THE MEMBER ARCHIVE

The verified repair and its recorded checks are member-only.

This mechanism has 7 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

These are finite toy sequences with explicit contracts; they make no claim about production decoder behavior or model quality. 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 2026-09-30T16:48:38.098905+00:00.

Case digest / cbd606ba3f38f736d8208bb7af23a6a0f6ab4a29e81a0268b07cfd24d0f74710