FA-100561 / Mixture of experts dispatch / Open access

Selected gate weights retain probability mass from discarded experts · case 01

Selected gate weights retain probability mass from discarded experts.

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

ROOT CAUSE

The full gate vector is normalized and then truncated, so retained weights can sum below one.

THE FAILURE

The full gate vector is normalized and then truncated, so retained weights can sum below one.

Unsuccessful approach: Selecting first but returning raw selected weights still fails the per-token unit-mass contract.

Case contract

Select top k (expert,weight) pairs by descending weight then name; normalize selected weights to sum one, with zero total producing an empty list.

Why this case matters

A deterministic toy router makes this dispatch decision independently inspectable.

1 / The failure

Exit 1

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

N = 1
observations = []
def solve(routes,k):
    total=sum(v for _,v in routes)
    return [(e,w/total) for e,w in sorted(routes,key=lambda x:(-x[1],x[0]))[:k]] if total else []
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
rows=[('a',N+1),('b',N),('c',1)]
check('top two mass',solve(rows,2),[('a',(N+1)/(2*N+1)),('b',N/(2*N+1))])
check('all mass',solve(rows,8),[('a',(N+1)/(2*N+2)),('b',N/(2*N+2)),('c',1/(2*N+2))])
check('one route',solve(rows,1),[('a',1.0)])
check('zero k',solve(rows,0),[])
check('empty',solve([],N),[])
check('zero total',solve([('a',0),('b',0)],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
top two mass	`[['a', 0.5], ['b', 0.25]]`	`[['a', 0.6666666666666666], ['b', 0.3333333333333333]]`	Failed
all mass	`[['a', 0.5], ['b', 0.25], ['c', 0.25]]`	`[['a', 0.5], ['b', 0.25], ['c', 0.25]]`	Passed
one route	`[['a', 0.5]]`	`[['a', 1.0]]`	Failed
zero k	`[]`	`[]`	Passed
empty	`[]`	`[]`	Passed
zero total	`[]`	`[]`	Passed

SHA-256 / 1c9408032c54eff36467f3e6a8235131bcedb090d55a74f4163c2c4c61042235

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(routes,k):
    chosen=sorted(routes,key=lambda x:(-x[1],x[0]))[:k]
    return chosen
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
rows=[('a',N+1),('b',N),('c',1)]
check('top two mass',solve(rows,2),[('a',(N+1)/(2*N+1)),('b',N/(2*N+1))])
check('all mass',solve(rows,8),[('a',(N+1)/(2*N+2)),('b',N/(2*N+2)),('c',1/(2*N+2))])
check('one route',solve(rows,1),[('a',1.0)])
check('zero k',solve(rows,0),[])
check('empty',solve([],N),[])
check('zero total',solve([('a',0),('b',0)],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
top two mass	`[['a', 2], ['b', 1]]`	`[['a', 0.6666666666666666], ['b', 0.3333333333333333]]`	Failed
all mass	`[['a', 2], ['b', 1], ['c', 1]]`	`[['a', 0.5], ['b', 0.25], ['c', 0.25]]`	Failed
one route	`[['a', 2]]`	`[['a', 1.0]]`	Failed
zero k	`[]`	`[]`	Passed
empty	`[]`	`[]`	Passed
zero total	`[['a', 0], ['b', 0]]`	`[]`	Failed

SHA-256 / e86e2065448291717b990a7706ce8f075420ceeafd28fdf037e926da4d11d4db

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

Inputs are finite toy scores and routes; the contract does not model a production model, accelerator, or learned routing 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 2026-09-30T16:48:37.161019+00:00.

Case digest / a30265a94da688df76cce4a0dda42cb9e4da57163d1efa9f6df15bc28c76ea2d