Co-Attested Claims: A Protocol Gate Against Chimeric Answers in Multi-Hop Biology QA

When a language-model agent answers a question that needs several pieces of evidence chained together, it tends to fail in a particular way. It finds one fact in one paper and a related fact in another, then fuses them into a single confident statement that no source supports. In molecular biology the stakes of this "stitching" are concrete. A long non-coding RNA (lncRNA, an RNA molecule that makes no protein but can act as a physical scaffold) often binds several protein complexes at once and tethers them to specific spots on DNA. A question such as "which repressive complex does the lncRNA that silences CDKN1A in gastric cancer bring to that promoter?" has an answer that is a multi-part assembly: one RNA, one or more proteins, one genomic locus, one cell context. If an agent joins the RNA from paper A to the enzyme from paper B, it produces a chimera. The result sounds like a real mechanism, uses correct gene names, and passes a casual citation check.

We propose to stop these chimeras at the protocol layer. The agent passes intermediate findings only as structured claim messages. Each message names every component of an assembly together with a verbatim span from a document the agent actually opened. A deterministic gate in the environment accepts a claim only if one span co-attests all of its components, meaning they all appear together in that single passage. Claims that fail the gate never enter the agent's working context or reach other agents, so a bad join cannot cascade into later hops. The proposal combines four directions. The lncRNA scaffolding literature is the application regime and the source of the higher-order relation structure. Multi-hop retrieval is the task. Standardized agent messages are the mechanism. Hallucination is the outcome we measure.

Core Claim

Existing runtime fixes for multi-hop agents check whether the agent looked at evidence. They do not check whether a claim's parts were ever seen together. Before Reasoning Can Fail introduces READ-GATE, which refuses to let an agent finalize while it has issued zero read actions. On HotpotQA, 2WikiMultiHopQA and MuSiQue this lifts accuracy by 14.9 to 19.9 points on trajectories that skipped reading, and by 3.2 to 9.4 points overall. Accelerating Scientific Research with Gemini in the Real-World checks every quantitative manuscript claim against raw execution logs and cuts severe result hallucinations from 90% to 4%. That check only works because the logs give it an exact-match target.

Our claim is that scaffolding-style questions open a third failure mode that neither gate addresses. The agent reads real documents, and every individual fact it states is attested somewhere, yet the n-ary assembly it asserts is attested nowhere. We hypothesize that a span-level co-attestation gate does three things. It sharply reduces this "chimeric assembly" error. It does so more cheaply than a learned verifier. Because it acts inside the loop, it beats the same check applied after the fact. The third part is the risky one, and the experiments are designed to test it.

Why Scaffolding Is the Right Stress Test

The biology literature motivates the task directly. Sequences to systems argues that gene regulation is a three-dimensional wiring diagram rather than a flat parts list. Individual papers report exactly the multi-part assemblies that pairwise knowledge graphs flatten. LINC01094 binds RBMS2 and HDAC1 to target CDKN1A through two routes. LINC00930 co-recruits methyltransferases and acetyltransferases to a single promoter. DANCR guides PRC2, the Polycomb repressive complex, to tumor-suppressor promoters. USP30-AS1 acts through different partners in the nucleus and in the cytoplasm. The USP30-AS1 case is a ready-made trap for compartment-swapping errors.

The corpus is dense with near-duplicate names (LINC009xx, many lncRNAs pairing with EZH2, HDAC1 and p21) and the relations are many-way. This is the setting where the entity- and role-swapping errors reported for tabular clinical generation (Milecki) and the overconfident conversion of tentative findings (Tang et al.) should be most frequent. The direction profile itself names the "computational representation gap": interactions get stored as pairwise edges when they are really multi-component assemblies. A co-attestation requirement is the simplest operational answer to that gap.

Research Program

We will build three things.

The first is ScaffoldHop, a multi-hop question set over a corpus of lncRNA mechanism abstracts and full texts. Following the bottom-up composition of MuSiQue, we start from curated single-paper assemblies of the form (lncRNA, protein partner(s), locus or target gene, cell or compartment context, effect) and chain them into two- to four-hop questions. A typical chain runs from phenotype to gene to lncRNA to the recruited complex. Also following MuSiQue, we add unanswerable contrast questions. These are built by pairing an lncRNA with a partner that a different paper reports for a different lncRNA, so the "answer" an agent would stitch together is precisely a chimera. Two annotators with molecular biology training will validate each assembly and each unanswerable item.

The second is the claim protocol. Agent messages follow the intent / action / observation decomposition argued for in LACP. Formal commitments are wrapped in XML tags so they can be parsed apart from free dialogue, the device NegotiationArena uses for binding offers. A claim message lists its typed components, a document ID and a quoted span. The environment checks three conditions, all deterministic:

  1. The document was read in this trajectory.
  2. The span occurs verbatim in that document, after normalization.
  3. Every listed component, mapped through a gene and protein synonym table, occurs within that span or its enclosing paragraph.

Failed claims return an error hint, as READ-GATE does. Accepted claims are written to an external ledger, in the spirit of the experiment ledger in SENPAI. Only ledger entries, not raw retrieved paragraphs, are carried into later hops and passed between a planner agent and a reader agent. This also attacks the context-accumulation problem that IRCoT leaves open.

Some correct answers really do require cross-paper synthesis. For these, the protocol allows an explicitly typed "composed" claim. A composed claim must reference two accepted claims that share a component, and it is flagged in the final answer rather than asserted as a single attested mechanism.

The third is the analysis. We will build an error taxonomy separating four cases: no-read guessing (what READ-GATE targets), fabricated spans, chimeric assemblies, and context or compartment swaps. For each condition we measure how often each occurs.

Baselines

The naive reference point is a ReAct-style agent with search and read tools and no gate, run on the same models, corpus index and step budget. It shows how common chimeras are when nothing stops them.

The first published method is IRCoT (Trivedi et al.), which interleaves each reasoning sentence with a retrieval call. It reports retrieval-recall gains of 11–21 points, F1 gains of up to 15 points, and up to 50% fewer factual errors in reasoning chains on Wikipedia multi-hop sets. We will rerun it on ScaffoldHop with the same retriever and model.

The strongest existing competitor is READ-GATE (Roh & Han). It reports net gains of 3.2 to 9.4 accuracy points across HotpotQA, 2WikiMultiHopQA and MuSiQue in minimal-reasoning settings. It is the strongest competitor because it is also a zero-training runtime invariant, so any gain we show must come from co-attestation specifically and not from "forcing reading." We will reproduce it on MuSiQue first to confirm our harness lands within its reported range. Then we run it on ScaffoldHop with identical tools, models (GPT-4o-mini-class and one open model) and step limits. We found no reported result on Lacuna for any method on lncRNA or scaffolding multi-hop QA, so on ScaffoldHop all baselines are measured fresh.

We also include a learned verifier: an off-the-shelf entailment or attribution checker of the kind surveyed in Attribution Techniques for Mitigating Hallucinated Information in RAG Systems. It scores each intermediate claim against its cited passage and is given matched token budgets.

Three ablations isolate the key ingredient:

  1. Span-existence only. The gate keeps conditions 1 and 2 and drops condition 3.
  2. Post-hoc co-attestation. The full check is applied only to the final answer, which is either rejected or turned into an abstention.
  3. Prompt-only. The co-attestation rule is stated in the system prompt but not enforced.

Beating the baselines means two things on ScaffoldHop. First, a statistically significant drop in the chimeric-assembly rate relative to READ-GATE and to the post-hoc variant. Second, answer F1 on answerable questions within two points of the best baseline or better, with abstention F1 on the contrast questions also improving. On the three Wikipedia benchmarks, the gate must not fall below READ-GATE's accuracy, since those benchmarks contain few n-ary claims.

Why not just…?

Why not just use READ-GATE on biology papers? READ-GATE removes guessing from preview snippets, but an agent that reads paper A and paper B can still join them. Our contrast questions are built so that both component facts are retrievable and readable. If READ-GATE's chimera rate on these questions matches ours, our hypothesis is wrong. The experiment measures this directly, and we treat it as a real possibility.

Why not just check the final answer against sources afterwards? This is the most serious threat. The deterministic check is identical; only its placement differs. Our argument is that in-loop rejection lets the agent recover a correct path, while a post-hoc check can only abstain. Cascading divergence after one bad intermediate step is a documented failure in multi-hop chains. The post-hoc ablation settles the question. If it matches the in-loop gate on F1 and on hallucination rate, the contribution reduces to "an n-ary attribution check," and we will report it that way.

Why not just use a stronger model or a learned verifier? A stronger model may stitch less often, and entailment verifiers catch some unsupported claims. However, the four-axis survey (Retrieval–Reasoning Processes for Multi-hop QA) documents that step-level verifiers buy faithfulness at large token and latency cost. Entailment models also judge one claim against one passage, which says nothing about whether the components co-occur. We will report chimera rate against tokens for each condition. A frontier model with no gate is included, so that we know whether the problem simply disappears with scale.

Why not just build a pairwise knowledge graph of lncRNA–protein edges? Pairwise edges are what make chimeras expressible in the first place. RNA–HDAC1 and HDAC1–CDKN1A can both be true without a single assembly linking them. We will include a graph-retrieval variant built from pairwise edges to show this, and report its chimera rate.

Proposed Experiments and Decisive Checks

The main experiment runs every condition above on ScaffoldHop, about 2,000 answerable and 600 contrast questions, and on the MuSiQue and HotpotQA dev splits. We use three seeds and report bootstrap confidence intervals. The measures are:

  • Answer F1, compared against a MuSiQue human reference of 78.0 F1 as a ceiling indicator for the Wikipedia setting.
  • Chimeric-assembly rate, judged by annotators blind to condition, with a deterministic co-occurrence script as a second check.
  • Abstention F1 on the contrast questions.
  • Gate rejection and recovery counts.
  • Tokens per question.

Three checks are decisive. The first is the in-loop versus post-hoc comparison described above. The second is a precision audit of the gate: on gold-annotated answerable items, how often does it wrongly reject a correct claim because of synonym gaps or paragraph boundaries? If false rejections exceed 10% of correct claims, the deterministic design is too brittle and needs a learned alias resolver. The third measures the share of correct ScaffoldHop answers that need composed claims. If that share is large, co-attestation is the wrong default for this literature. Across all of these, the result reported for each condition is its full error taxonomy, so a reader can see which failure mode each method actually moves.