FA-98911 / Gradient accumulation weighting / Open access

Flush emits a normalized result without clearing its accumulator · case 01

Flush emits a normalized result without clearing its accumulator.

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

ROOT CAUSE

The broken flush returns the normalized value but leaves all accumulated state available for accidental reuse.

THE FAILURE

The broken flush returns the normalized value but leaves all accumulated state available for accidental reuse.

Unsuccessful approach: Clearing only the count leaves a stale numerator in the next window.

Case contract

Given [sum,count], flush returns [mean, [0,0]]. Empty state has mean zero; all state is cleared for the next window.

Why this case matters

A deterministic toy model isolates one local accumulation contract.

1 / The failure

Exit 1

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

N = 1
observations = []
def solve(state):
    return [state[0] / state[1] if state[1] else 0,state]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('flush populated',solve([12*N,3*N]),[4, [0,0]])
check('flush singleton',solve([5*N,N]),[5,[0,0]])
check('flush empty',solve([0,0]),[0,[0,0]])
check('flush zero numerator',solve([0,3*N]),[0,[0,0]])
check('flush unequal',solve([14*N,4*N]),[3.5,[0,0]])
check('flush unit',solve([9*N,3*N]),[3,[0,0]])
check('flush larger',solve([20*N,4*N]),[5,[0,0]])
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
flush populated	`[4.0, [12, 3]]`	`[4, [0, 0]]`	Failed
flush singleton	`[5.0, [5, 1]]`	`[5, [0, 0]]`	Failed
flush empty	`[0, [0, 0]]`	`[0, [0, 0]]`	Passed
flush zero numerator	`[0.0, [0, 3]]`	`[0, [0, 0]]`	Failed
flush unequal	`[3.5, [14, 4]]`	`[3.5, [0, 0]]`	Failed
flush unit	`[3.0, [9, 3]]`	`[3, [0, 0]]`	Failed
flush larger	`[5.0, [20, 4]]`	`[5, [0, 0]]`	Failed

SHA-256 / 206854cb1254f931634f1e8dbe241f04e43b1590b1ce817f175564f2e6df9e04

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(state):
    return [state[0] / state[1] if state[1] else 0,[state[0],0]]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('flush populated',solve([12*N,3*N]),[4, [0,0]])
check('flush singleton',solve([5*N,N]),[5,[0,0]])
check('flush empty',solve([0,0]),[0,[0,0]])
check('flush zero numerator',solve([0,3*N]),[0,[0,0]])
check('flush unequal',solve([14*N,4*N]),[3.5,[0,0]])
check('flush unit',solve([9*N,3*N]),[3,[0,0]])
check('flush larger',solve([20*N,4*N]),[5,[0,0]])
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
flush populated	`[4.0, [12, 0]]`	`[4, [0, 0]]`	Failed
flush singleton	`[5.0, [5, 0]]`	`[5, [0, 0]]`	Failed
flush empty	`[0, [0, 0]]`	`[0, [0, 0]]`	Passed
flush zero numerator	`[0.0, [0, 0]]`	`[0, [0, 0]]`	Passed
flush unequal	`[3.5, [14, 0]]`	`[3.5, [0, 0]]`	Failed
flush unit	`[3.0, [9, 0]]`	`[3, [0, 0]]`	Failed
flush larger	`[5.0, [20, 0]]`	`[5, [0, 0]]`	Failed

SHA-256 / 06e32d6575c3134824c5813e92cf5f661d485837b9573089d1b6d45e5811715c

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

Inputs are scalar or short-list summaries and do not model autodiff, devices, or an optimizer. 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 2026-09-30T16:48:18.065606+00:00.

Case digest / ef84bb13f12366c2f3388b0d6fe0e740a606b21a1830f524a13810a27dc77318