FA-100636 / Mixture of experts dispatch / Open access

Extreme router logits bypass the declared score bound · case 01

Extreme router logits bypass the declared score bound.

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

ROOT CAUSE

Gate scores are ranked without applying the toy router clipping interval.

THE FAILURE

Gate scores are ranked without applying the toy router clipping interval.

Unsuccessful approach: Clamping only the upper tail still lets large negative logits dominate lower bounded scores.

Case contract

For [expert,score] logits clamp score into [-1,1], then return pairs descending by clamped score and expert ID.

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(scores):
    return sorted(scores,key=lambda x:(-x[1],x[0]))
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
rows=[('a',N+2),('b',1),('c',-N-2)]
check('upper clamp tie',solve(rows),[('a',1),('b',1),('c',-1)])
check('lower clamp',solve([('a',-N-1),('b',-1)]),[('a',-1),('b',-1)])
check('inside unchanged',solve([('x',0.5),('y',-0.5)]),[('x',0.5),('y',-0.5)])
check('empty',solve([]),[])
check('zero logits',solve([('b',0),('a',0)]),[('a',0),('b',0)])
check('singleton clipped',solve([('q',N+1)]),[('q',1)])
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
upper clamp tie	`[['a', 3], ['b', 1], ['c', -3]]`	`[['a', 1], ['b', 1], ['c', -1]]`	Failed
lower clamp	`[['b', -1], ['a', -2]]`	`[['a', -1], ['b', -1]]`	Failed
inside unchanged	`[['x', 0.5], ['y', -0.5]]`	`[['x', 0.5], ['y', -0.5]]`	Passed
empty	`[]`	`[]`	Passed
zero logits	`[['a', 0], ['b', 0]]`	`[['a', 0], ['b', 0]]`	Passed
singleton clipped	`[['q', 2]]`	`[['q', 1]]`	Failed

SHA-256 / 4bb6224d1bbdc0236dd94a30805f9fdb4c701ba0384ccb6a51f3b468b649b1ae

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(scores):
    return sorted([(e,min(1,s)) for e,s in scores],key=lambda x:(-x[1],x[0]))
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
rows=[('a',N+2),('b',1),('c',-N-2)]
check('upper clamp tie',solve(rows),[('a',1),('b',1),('c',-1)])
check('lower clamp',solve([('a',-N-1),('b',-1)]),[('a',-1),('b',-1)])
check('inside unchanged',solve([('x',0.5),('y',-0.5)]),[('x',0.5),('y',-0.5)])
check('empty',solve([]),[])
check('zero logits',solve([('b',0),('a',0)]),[('a',0),('b',0)])
check('singleton clipped',solve([('q',N+1)]),[('q',1)])
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
upper clamp tie	`[['a', 1], ['b', 1], ['c', -3]]`	`[['a', 1], ['b', 1], ['c', -1]]`	Failed
lower clamp	`[['b', -1], ['a', -2]]`	`[['a', -1], ['b', -1]]`	Failed
inside unchanged	`[['x', 0.5], ['y', -0.5]]`	`[['x', 0.5], ['y', -0.5]]`	Passed
empty	`[]`	`[]`	Passed
zero logits	`[['a', 0], ['b', 0]]`	`[['a', 0], ['b', 0]]`	Passed
singleton clipped	`[['q', 1]]`	`[['q', 1]]`	Passed

SHA-256 / d16cbd01929e40c55d831ddf41db1612ca442ad25c288ed6dfc900b5b37a2d90

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 2026-09-30T16:48:37.646491+00:00.

Case digest / 0686a1cac2fda84977fca7558f088b3c00a5b2e4e015aa91c54d26655bd06724