FA-99216 / Checkpoint tensor mapping / Open access

Checkpoint mapping: unlisted numeric conversions are accepted · case 01

unlisted numeric conversions are accepted.

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

ROOT CAUSE

General numeric membership ignores the explicit conversion policy.

THE FAILURE

General numeric membership ignores the explicit conversion policy.

Unsuccessful approach: Treating every float-to-float cast as safe admits an unlisted conversion.

Case contract

This toy policy accepts identical dtype labels and the directed pairs f16-to-f32 and i8-to-f16; all other conversions are rejected.

Why this case matters

A bounded offline model of checkpoint tensor bookkeeping in a training workflow.

1 / The failure

Exit 1

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

N = 1
observations = []
def solve(source_dtype, target_dtype):
    return source_dtype in ("i8","f32","f16") and target_dtype in ("i8","f32","f16")
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('allowed float', solve(*('f16', 'f32')), True)
check('allowed integer route', solve(*('i8', 'f16')), True)
check('disallowed float', solve(*('f32', 'f16')), False)
check('same', solve(*('i8', 'i8')), True)
check('unknown', solve(*('x', 'x')), True)
check('reverse', solve(*('f16', 'i8')), False)
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
allowed float	`True`	`True`	Passed
allowed integer route	`True`	`True`	Passed
disallowed float	`True`	`False`	Failed
same	`True`	`True`	Passed
unknown	`False`	`True`	Failed
reverse	`True`	`False`	Failed

SHA-256 / e1e1e0a25a870f0ed2e2cd7bee77639ca9152f7693f37e260f99e584d3c372b0

2 / The unsuccessful fix

Exit 1

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

N = 1
observations = []
def solve(source_dtype, target_dtype):
    return source_dtype==target_dtype or source_dtype.startswith("f") and target_dtype.startswith("f")
def check(label, actual, expected):
    observations.append({"check": label, "actual": actual, "expected": expected, "passed": actual == expected})
check('allowed float', solve(*('f16', 'f32')), True)
check('allowed integer route', solve(*('i8', 'f16')), True)
check('disallowed float', solve(*('f32', 'f16')), False)
check('same', solve(*('i8', 'i8')), True)
check('unknown', solve(*('x', 'x')), True)
check('reverse', solve(*('f16', 'i8')), False)
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
allowed float	`True`	`True`	Passed
allowed integer route	`False`	`True`	Failed
disallowed float	`True`	`False`	Failed
same	`True`	`True`	Passed
unknown	`True`	`True`	Passed
reverse	`False`	`False`	Passed

SHA-256 / 34ce177ead5c52523462624bd4245c690042cddfb3bc6fa27c942efeaba1daeb

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 already decoded Python values; this toy contract makes no claim about a framework, file format, or production checkpoint. 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 2026-09-30T16:48:21.699890+00:00.

Case digest / 6aa3dca3d0c3c32f7cc4c474da6b0c81511c22f9a9a33ce91c3aadcfe1b21b1a