Paper P-8facce

✗ Not runnable within the agent budget (18 turns) P-8facce · repository hidden · run dev-v1

PCI

papercibot reviewed this paper's code · 2026-10-11 11:11

Pipeline

✓Triage2s

✓Claims sealed6s

✗Build12.4m

–Offline run0s

–Reruns0s

!Verdict0s

–Package

–Oracle

–Trainability

Walkthrough

PaperCI fetched the paper's repository into a fresh CPU sandbox. A builder agent (claude-sonnet-4.6) worked for 18 turns (12.4 min), installed 2 package(s) (2, huggingface_hub), but did not reach a runnable state (max_turns)

5claims sealed

0reproduced

0measured

0offline reruns

18agent turns

$0.28total cost

Claims vs measured

Claim (from the paper)	Paper	PaperCI measured	Tolerance	Status
c1 Absolute Error (AE) · image-level geometry completion with sensor depth prior	0.026	–	-	Blocked
c2 Relative Error (RE) · image-level geometry completion with sensor depth prior	0.038	–	-	Blocked
c3 Logarithmic Error (LE) · image-level geometry completion with sensor depth prior	0.036	–	-	Blocked
c4 Threshold Accuracy · image-level geometry completion with sensor depth prior	79.32%	–	-	Blocked
c5 Threshold Accuracy · image-level geometry completion with sensor depth prior	92.56%	–	-	Blocked

Checks

✓SandboxFresh Docker container: non-root, all capabilities dropped, egress only via an allowlist proxy

✓Claims sealed5 headline claims extracted from the paper text and sealed (sha256 103f908385d8) before the builder started; the builder never saw the values

✗Buildmax_turns after 18 agent turns, 12.4 min

✓No hardcoded answersno sealed claim value appears in anything the builder wrote

✓No copied resultsno output file is a byte-identical copy of a result file in the repo

✓Computed, not replayedthe run computes its numbers instead of loading stored results

✓Uses the paper's codethe run script executes the repository's own code

!EgressRefused hosts: cas-server.xethub.hf.co (423x)

Suggestions for the authors

1. Declare missing dependenciesInstalled during the build but not in your manifests: 2 huggingface_hub.

2. Add a one-command reproduction entry pointGetting this repository to produce the headline numbers took more than 18 agent steps. A reproduce.sh (or a README section) that prints each table number would make it checkable.

3. Data missingA file the code expects is not in the repo and no download step creates it.

4. Egress deniedThe code downloads from a host that a sandboxed reviewer cannot reach; mirror the data.

Builder claude-sonnet-4.6 · verifier nova-pro (different model family) · neatlogs trace 7bfd8f9e85e8a9e5a02622e3484c574c · sealed claims sha256 103f908385d87fad
Verdict blocked-infra (medium):