FA-98206 / Attention mask construction / Open access
Block-causal mask lets keys cross supplied block boundaries · case 01
Block-causal mask lets keys cross supplied block boundaries.
ROOT CAUSE
The mask compares flattened positions and never derives the owner of each position.
THE FAILURE
The mask compares flattened positions and never derives the owner of each position.
Unsuccessful approach: Restricting a block to adjacent token positions removes valid long-range causal keys.
Case contract
Given nonnegative block lengths, expand owner IDs; visibility is same owner and causal. Zero-length blocks are allowed.
Why this case matters
A small explicit visibility matrix makes sequence-attention boundary behavior reviewable.
1 / The failure
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
N = 1
observations = []
def solve(lengths):
return [[j<=i for j in range(sum(lengths))] for i in range(sum(lengths))]
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
l=[N,0,2]
o=[]
for b,z in enumerate(l): o.extend([b]*z)
check('zero block boundary',solve(l),[[o[i]==o[j] and j<=i for j in range(len(o))] for i in range(len(o))])
check('two singleton blocks',solve([1,1]),[[True,False],[False,True]])
check('no blocks',solve([]),[])
check('all empty blocks',solve([0,0]),[])
check('one block',solve([3]),[[True,False,False],[True,True,False],[True,True,True]])
check('trailing empty block',solve([2,0]),[[True,False],[True,True]])
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 |
|---|---|---|---|
| zero block boundary | [[True, False, False], [True, True, False], [True, True, True]] | [[True, False, False], [False, True, False], [False, True, True]] | Failed |
| two singleton blocks | [[True, False], [True, True]] | [[True, False], [False, True]] | Failed |
| no blocks | [] | [] | Passed |
| all empty blocks | [] | [] | Passed |
| one block | [[True, False, False], [True, True, False], [True, True, True]] | [[True, False, False], [True, True, False], [True, True, True]] | Passed |
| trailing empty block | [[True, False], [True, True]] | [[True, False], [True, True]] | Passed |
SHA-256 / f0ae500e23034cc6fff01e7490ff94b2f8487566f7bddd0f4fe01b241ca5260d
2 / The unsuccessful fix
Exit 1"""Failure Map reference implementation. Python standard library only."""
import json
N = 1
observations = []
def solve(lengths):
owners=[]
for b,size in enumerate(lengths): owners.extend([b]*size)
return [[owners[i]==owners[j] and j<=i and i-j<=1 for j in range(len(owners))] for i in range(len(owners))]
def check(label, actual, expected):
observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
l=[N,0,2]
o=[]
for b,z in enumerate(l): o.extend([b]*z)
check('zero block boundary',solve(l),[[o[i]==o[j] and j<=i for j in range(len(o))] for i in range(len(o))])
check('two singleton blocks',solve([1,1]),[[True,False],[False,True]])
check('no blocks',solve([]),[])
check('all empty blocks',solve([0,0]),[])
check('one block',solve([3]),[[True,False,False],[True,True,False],[True,True,True]])
check('trailing empty block',solve([2,0]),[[True,False],[True,True]])
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 |
|---|---|---|---|
| zero block boundary | [[True, False, False], [False, True, False], [False, True, True]] | [[True, False, False], [False, True, False], [False, True, True]] | Passed |
| two singleton blocks | [[True, False], [False, True]] | [[True, False], [False, True]] | Passed |
| no blocks | [] | [] | Passed |
| all empty blocks | [] | [] | Passed |
| one block | [[True, False, False], [True, True, False], [False, True, True]] | [[True, False, False], [True, True, False], [True, True, True]] | Failed |
| trailing empty block | [[True, False], [True, True]] | [[True, False], [True, True]] | Passed |
SHA-256 / f663cfe5b4c8cc9d7e7421a332e0e88bce8c2a41f7e8d0f3fa831af362f76a7f
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.
Sign in to the archive ↗Verification & scope
These are toy Boolean matrix contracts; they do not define a production model architecture or tensor library behavior. 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 2026-09-30T16:48:10.630126+00:00.
Case digest / a95cf34d474e1728333e2a2a2257a6c96f4249c1b6cc81938b535079db74bc14