FA-99031 / Optimizer update state / Open access

Bias parameter receives unintended weight decay · case 01

Bias parameter receives unintended weight decay.

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

ROOT CAUSE

All parameter roles share the same decay path.

THE FAILURE

All parameter roles share the same decay path.

Unsuccessful approach: Removing decay from all roles repairs biases but also exempts regularized weights.

Case contract

Records [role,p,g], lr=1, wd=.5; p_next=p-g-wd*p only for role weight.

Why this case matters

Optimizer implementations must associate each parameter with its own mathematically specified update state.

1 / The failure

Exit 1

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

N = 1
observations = []
def solve(records):
    return [[role,p-g-0.5*p] for role,p,g in records]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('weight and bias',solve([['weight',4,1],['bias',4,1]]),[['weight',1],['bias',3]])
check('bias only',solve([['bias',N,2]]),[['bias',N-2]])
check('weight only',solve([['weight',N,0]]),[['weight',N/2]])
check('empty params',solve([]),[])
check('negative weight',solve([['weight',-4,0]]),[['weight',-2]])
check('negative bias',solve([['bias',-4,1]]),[['bias',-5]])
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
weight and bias	`[['weight', 1.0], ['bias', 1.0]]`	`[['weight', 1], ['bias', 3]]`	Failed
bias only	`[['bias', -1.5]]`	`[['bias', -1]]`	Failed
weight only	`[['weight', 0.5]]`	`[['weight', 0.5]]`	Passed
empty params	`[]`	`[]`	Passed
negative weight	`[['weight', -2.0]]`	`[['weight', -2]]`	Passed
negative bias	`[['bias', -3.0]]`	`[['bias', -5]]`	Failed

SHA-256 / 382a4ba9495639e17081021b4fe224c9792d13f3e099301b45613200c7ef8652

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(records):
    return [[role,p-g] for role,p,g in records]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('weight and bias',solve([['weight',4,1],['bias',4,1]]),[['weight',1],['bias',3]])
check('bias only',solve([['bias',N,2]]),[['bias',N-2]])
check('weight only',solve([['weight',N,0]]),[['weight',N/2]])
check('empty params',solve([]),[])
check('negative weight',solve([['weight',-4,0]]),[['weight',-2]])
check('negative bias',solve([['bias',-4,1]]),[['bias',-5]])
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
weight and bias	`[['weight', 3], ['bias', 3]]`	`[['weight', 1], ['bias', 3]]`	Failed
bias only	`[['bias', -1]]`	`[['bias', -1]]`	Passed
weight only	`[['weight', 1]]`	`[['weight', 0.5]]`	Failed
empty params	`[]`	`[]`	Passed
negative weight	`[['weight', -4]]`	`[['weight', -2]]`	Failed
negative bias	`[['bias', -5]]`	`[['bias', -5]]`	Passed

SHA-256 / ee71557c8133519d5d54dacd727d40666c2c4b9a61c3fe31bf4b072c883a58eb

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

These are deterministic scalar or short-vector toy contracts, not claims about production optimizer APIs or training 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 2026-09-30T16:48:19.456541+00:00.

Case digest / 3af8c4ed5432006f4603a29bd483f8253cf6a87a8b05703a7bbdaeb1ed275f2d