FA-100566 / Mixture of experts dispatch / Open access

Dropped routes make surviving token weights sum to less than one · case 01

Dropped routes make surviving token weights sum to less than one.

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

ROOT CAUSE

The router normalizes each token’s route weights before capacity drops remove selected alternatives.

THE FAILURE

The router normalizes each token’s route weights before capacity drops remove selected alternatives.

Unsuccessful approach: Filtering capacity correctly but retaining pre-drop normalized weights still leaves surviving token mass below one.

Case contract

Routes [token,expert,weight] arrive in order; accept up to capacity per expert, then normalize each token’s accepted weights.

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,capacity):
    tot={}
    for t,e,w in routes: tot[t]=tot.get(t,0)+w
    normalized=[[t,e,w/tot[t]] for t,e,w in routes if tot[t]]
    out=[]; c={}
    for t,e,w in normalized:
     if c.get(e,0)<capacity: out.append([t,e,w]); c[e]=c.get(e,0)+1
    return out
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
a='a'+str(N); b='b'+str(N)
c='c'+str(N); rows=[[0,a,2],[0,b,1],[1,a,3],[1,c,1]]
check('survivors normalized',solve(rows,1),[[0,a,2/3],[0,b,1/3],[1,c,1.0]])
check('two capacity',solve(rows,2),[[0,a,2/3],[0,b,1/3],[1,a,3/4],[1,c,1/4]])
check('zero capacity',solve(rows,0),[])
check('empty',solve([],N),[])
check('one route',solve([[0,a,N]],1),[[0,a,1.0]])
check('single token',solve([[0,a,1],[0,b,1]],N),[[0,a,0.5],[0,b,0.5]])
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
survivors normalized	`[[0, 'a1', 0.6666666666666666], [0, 'b1', 0.3333333333333333], [1, 'c1', 0.25]]`	`[[0, 'a1', 0.6666666666666666], [0, 'b1', 0.3333333333333333], [1, 'c1', 1.0]]`	Failed
two capacity	`[[0, 'a1', 0.6666666666666666], [0, 'b1', 0.3333333333333333], [1, 'a1', 0.75], [1, 'c1', 0.25]]`	`[[0, 'a1', 0.6666666666666666], [0, 'b1', 0.3333333333333333], [1, 'a1', 0.75], [1, 'c1', 0.25]]`	Passed
zero capacity	`[]`	`[]`	Passed
empty	`[]`	`[]`	Passed
one route	`[[0, 'a1', 1.0]]`	`[[0, 'a1', 1.0]]`	Passed
single token	`[[0, 'a1', 0.5], [0, 'b1', 0.5]]`	`[[0, 'a1', 0.5], [0, 'b1', 0.5]]`	Passed

SHA-256 / 20d5d05f83c7e187e3f92e279d552089b0503f47c7900d352c33e3faf06b4893

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(routes,capacity):
    out=[]; c={}
    for t,e,w in routes:
     if c.get(e,0)<capacity: out.append([t,e,w]); c[e]=c.get(e,0)+1
    return out
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
a='a'+str(N); b='b'+str(N)
c='c'+str(N); rows=[[0,a,2],[0,b,1],[1,a,3],[1,c,1]]
check('survivors normalized',solve(rows,1),[[0,a,2/3],[0,b,1/3],[1,c,1.0]])
check('two capacity',solve(rows,2),[[0,a,2/3],[0,b,1/3],[1,a,3/4],[1,c,1/4]])
check('zero capacity',solve(rows,0),[])
check('empty',solve([],N),[])
check('one route',solve([[0,a,N]],1),[[0,a,1.0]])
check('single token',solve([[0,a,1],[0,b,1]],N),[[0,a,0.5],[0,b,0.5]])
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
survivors normalized	`[[0, 'a1', 2], [0, 'b1', 1], [1, 'c1', 1]]`	`[[0, 'a1', 0.6666666666666666], [0, 'b1', 0.3333333333333333], [1, 'c1', 1.0]]`	Failed
two capacity	`[[0, 'a1', 2], [0, 'b1', 1], [1, 'a1', 3], [1, 'c1', 1]]`	`[[0, 'a1', 0.6666666666666666], [0, 'b1', 0.3333333333333333], [1, 'a1', 0.75], [1, 'c1', 0.25]]`	Failed
zero capacity	`[]`	`[]`	Passed
empty	`[]`	`[]`	Passed
one route	`[[0, 'a1', 1]]`	`[[0, 'a1', 1.0]]`	Passed
single token	`[[0, 'a1', 1], [0, 'b1', 1]]`	`[[0, 'a1', 0.5], [0, 'b1', 0.5]]`	Failed

SHA-256 / d53bebf76a098621dd0c2817fd1a07838e21174f032fb3ac13ce10acdbd23b62

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 2026-09-30T16:48:37.194334+00:00.

Case digest / fc0fa0d5438ad385d0104f607b6401fedb22c087333508f3e5cd5b04392004b2