Verify the whole after recombining parts

Learning objective

Add integration and verification steps that recombine subtask results and check for gaps, duplication, and contradictions.

Core idea

Local acceptance proves only that each returned part met its assigned contract. The combined result needs a separate test against the original outcome. A plan can omit a requirement, two workers can count the same evidence, and individually supported statements can become contradictory when a summary removes their qualifiers.

For the architect, integration is therefore an explicit stage with its own inputs, checks, and release decision. Anthropic’s prompt-chaining pattern places programmatic gates between calls; its evaluator-optimizer pattern adds feedback against clear criteria. Use exact checks for identities and coverage, then an evidence-based rubric for conclusions that code cannot settle.

How it works

Build an accepted input set before synthesis

Treat accepted worker outputs as evidence records, not as paragraphs ready to concatenate. Preserve which original requirement each record addresses, the source version and location, and any limits on its claim. Anthropic’s research-system account recommends persistent artifacts rather than repeatedly passing large outputs through the lead agent, reducing information loss across stages.

A useful application design maintains three linked artifacts: an expected-work list, an accepted-evidence set, and an integration decision. Code attaches trusted task and input identities; workers supply the supporting passages and interpretations. These are application artifacts, not an Anthropic integration API.

For a corpus pipeline, reconcile accepted outputs with the expected map units described in 1.5.3. For a heterogeneous investigation, reconcile them with the original requirements. Do both where necessary: every document can be processed while one required comparison remains unanswered.

Check gaps at two levels

A mechanical gap is an expected unit or required output field with no accepted result. Detect it through set comparison or contract validation. A semantic gap is an unanswered obligation despite apparently complete records. Detect it through an explicit rubric that asks what evidence establishes each required conclusion.

Anthropic’s research evaluation uses completeness, factual accuracy, citation accuracy, source quality, and tool efficiency. Adapt those dimensions into the integration prompt. Require the evaluator to identify the unmet criterion and supporting record IDs, rather than return only a global quality score. The resulting feedback must tell the coordinator what action can change acceptance.

Do not let synthesis silently convert unavailable evidence into a negative finding. An application status such as unresolved remains unresolved until an accepted follow-up settles it. A complete explanation can disclose uncertainty, but a mandatory decision criterion cannot disappear because no worker answered it.

Detect duplication without discarding corroboration

Define the identity of a claim before counting claims. In an example document pipeline, an application key can combine source version, passage location, subject, and review question. If overlapping shards return the same passage for the same question, link both worker records to one evidence item before computing totals.

Do not deduplicate solely by text similarity. Two documents can contain the same wording while establishing separate obligations. Conversely, paraphrases of one passage can look different while describing one event. Keep the original records and their attribution so the integration stage can explain which items it merged and why.

Anthropic distinguishes sectioning from voting: independent scope coverage and repeated perspectives serve different purposes. Apply the same distinction to evidence. Two workers interpreting one passage are not two independent sources. Two genuinely distinct sources supporting one conclusion can be retained as corroboration without inflating the number of underlying events. Code handles stable identities; rubric-based review resolves ambiguous semantic overlap.

Check compatibility across accepted claims

Recall the comparability and conflict-record contract: retain attributed incompatible claims until their definitions and treatment are settled.

At integration, test how those claims interact across team outputs. A supplier-risk conclusion might depend on an incident count from operations and a cost estimate from finance. Keep references from that conclusion to both accepted records and their versions. If deduplication changes the incident count, mark the dependent conclusion for verification; do not replace either team’s source record with the synthesized wording.

Use Anthropic’s evaluator-optimizer pattern to check these relationships against explicit criteria. Classify the defect before re-delegating, then re-check the whole revised answer. The professional control is the dependency trail: the integration decision identifies which accepted records and merge decisions support the released conclusion.

Integration runs exact checks, then the rubric; a defect is classified before re-delegating, then the whole revised answer is re-checked.

Assign ownership across teams

When workers belong to different teams, name one integration owner for the expected-work list and release decision. Each contributing team owns its output contract and source-version declarations. The integration owner accepts changes to required coverage and records which team must supply any missing contribution. A team finishing its assignment cannot remove another team’s outstanding obligation.

Give the integration owner authority to hold release when a required comparison lacks accepted evidence. A domain reviewer decides consequential interpretations; the integration owner records that decision beside the affected requirement and evidence versions. This responsibility split is an application design: it makes the research account’s outcome-based evaluation actionable across organizational boundaries.

Evaluate integration as its own stage

Anthropic’s research account recommends starting evaluation with small representative samples rather than waiting for a large dataset. Apply that advice to labelled integration cases whose inputs are already accepted worker outputs. Include a missing required contribution, duplicate reports of one event, similar wording for distinct obligations, and a conflict that changes the recommendation.

For each case, label the expected gap, merge or separation decision, conflict, and release outcome. Measure missed defects and incorrect holds separately: rejecting every result would catch defects while making the service unusable. Track duplicate-detection misses against the labelled duplicates and false merges against distinct evidence items. These are proposed application metrics, not built-in Anthropic telemetry.

Rerun those cases when a contributing team’s contract, the merge key, or the integration prompt changes. A missed duplicate becomes a regression even if the final prose remains persuasive. Keep clean cases and cross-team version changes in the set, and expand it with defects found during human review. The acceptance test concerns correct recombination and release, rather than a prescribed sequence of evaluator calls.

Key distinctions and trade-offs

Aggregation versus integration: Aggregation collects and counts records. Integration establishes that the combined records support the required conclusion.

Schema completeness versus evidence completeness: Required fields can all exist while the evidence fails to answer a required question. Use both contract checks and a semantic rubric.

Deduplication versus suppression: Deduplication removes repeated counting while retaining attribution. Suppression hides an inconvenient source or interpretation. Preserve the decision trail for each merge and unresolved conflict.

Verification adds latency and reviewer effort. Use deterministic checks for exhaustive accounting and exact identities, then focus model or human review on interpretation. Anthropic’s research account notes that human testing finds edge cases automation misses; a fluent evaluator response is not a reason to remove that review channel.

Architecture / decision scenario

A government technology team must prepare a procurement comparison before a committee meeting. Four workers return accepted reports covering deployment, operating cost, accessibility, and support. The final recommendation must address all four criteria and identify evidence the committee can inspect.

The integration stage compares evidence against the original request. It finds that the accessibility report describes documentation quality but supplies no evidence for the required interface behavior. It also finds that deployment and support workers counted the same incident twice, and that two cost estimates assume different annual workloads.

Code reconciles the shared incident identity and recomputes its count while retaining both worker references. The coordinator requests targeted accessibility evidence and workload-aligned cost analysis. It preserves the accepted deployment findings. Synthesis then receives the repaired evidence and explicit instructions on the remaining uncertainty.

The final evaluator checks every criterion, incident totals, comparable assumptions, and agreement between the rationale and recommendation. If accessibility remains unresolved at the deadline, the committee receives a comparison that defers that judgment. The stage cannot describe the procurement recommendation as fully verified merely because all four workers finished.

Common exam traps

• Checking that every assigned task finished without checking the original requirements.
• Counting repeated worker reports as independent evidence.
• Merging similar text from different documents as though it were one obligation.
• Resolving incompatible estimates by averaging them without a valid analytical basis.
• Repairing the evidence table while leaving the final recommendation unchanged.
• Changing a contributing team’s output contract without rerunning integration cases.

What to remember

Integrate against the original outcome, not only the decomposition plan. Account for missing work, deduplicate by evidence identity, and test whether accepted claims remain compatible in the final answer. Repair the specific defect and verify the revised whole before release.

Knowledge check

References

• Anthropic, Building effective agents
• Anthropic, How we built our multi-agent research system