FA-98951 / Optimizer update state / Open access

Momentum dampening scales the carried buffer · case 01

Momentum dampening scales the carried buffer.

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

ROOT CAUSE

Dampening is applied to the whole momentum recurrence.

THE FAILURE

Dampening is applied to the whole momentum recurrence.

Unsuccessful approach: Removing dampening from the entire recurrence preserves the carried buffer but leaves the new gradient undampened.

Case contract

v_next=mu*v+(1-dampening)*g; return [p-lr*v_next,v_next].

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(p, g, v, lr, mu, damp):
    nv=(1-damp)*(mu*v+g)
    return [p-lr*nv,nv]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('fresh buffer',solve(8,4,0,1,0.5,0.5),[6,2])
check('retained velocity',solve(8,4,6,1,0.5,0.5),[3,5])
check('full dampening',solve(3,7,2,1,0.8,1),[1.4,1.6])
check('zero momentum',solve(4,-2,9,2,0,0.25),[7,-1.5])
check('zero gradient',solve(5,0,4,1,0.5,0.25),[3,2])
check('zero learning rate',solve(2,3,4,0,0.5,0.5),[2,3.5])
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
fresh buffer	`[6.0, 2.0]`	`[6, 2]`	Passed
retained velocity	`[4.5, 3.5]`	`[3, 5]`	Failed
full dampening	`[3.0, 0.0]`	`[1.4, 1.6]`	Failed
zero momentum	`[7.0, -1.5]`	`[7, -1.5]`	Passed
zero gradient	`[3.5, 1.5]`	`[3, 2]`	Failed
zero learning rate	`[2.0, 2.5]`	`[2, 3.5]`	Failed

SHA-256 / 79670437008ff5fae2fb5827cc4acffd3b1cc041a009ba1bc00810b63d42f753

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(p, g, v, lr, mu, damp):
    nv=mu*v+g
    return [p-lr*nv,nv]
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('fresh buffer',solve(8,4,0,1,0.5,0.5),[6,2])
check('retained velocity',solve(8,4,6,1,0.5,0.5),[3,5])
check('full dampening',solve(3,7,2,1,0.8,1),[1.4,1.6])
check('zero momentum',solve(4,-2,9,2,0,0.25),[7,-1.5])
check('zero gradient',solve(5,0,4,1,0.5,0.25),[3,2])
check('zero learning rate',solve(2,3,4,0,0.5,0.5),[2,3.5])
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
fresh buffer	`[4.0, 4.0]`	`[6, 2]`	Failed
retained velocity	`[1.0, 7.0]`	`[3, 5]`	Failed
full dampening	`[-5.6, 8.6]`	`[1.4, 1.6]`	Failed
zero momentum	`[8, -2]`	`[7, -1.5]`	Failed
zero gradient	`[3.0, 2.0]`	`[3, 2]`	Passed
zero learning rate	`[2.0, 5.0]`	`[2, 3.5]`	Failed

SHA-256 / 32280c960ca88e33e95cca727c80c36c124741ba6663aca984118e4b051c3bef

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 2026-09-30T16:48:18.534176+00:00.

Case digest / d9f3100b4f6b7b67ea98cc79813c55b27f36f8dc15a948aa5fbe51e2f7b8e036