Waits for M-of-N incoming branches to complete before enabling the outgoing flow. The remaining N-M branches complete normally and are not cancelled. Must be within a structured block where all branches originate from the same split.
A quorum review system assigns a research paper to 5 domain experts. Each reviewer independently assesses the paper and submits a structured review. Once 3 of the 5 reviews are submitted — a quorum — the editorial decision agent is unlocked and produces an accept/revise/reject recommendation. The remaining 2 reviewers still complete their reviews; their outputs are appended to the editorial record for completeness, but they do not delay the decision.
The key insight: scientific peer review operates on quorum logic by design. Three independent assessments are statistically sufficient for a reliable editorial decision. Forcing all 5 to complete before proceeding penalizes the process when one reviewer is slow or unavailable — a common occurrence. The partial join captures the governance benefit of multiple independent opinions while removing the tail-latency penalty of waiting for the slowest participant. Remaining reviewers contribute to the archival record, not the critical path.
| Metric | Signal |
|---|---|
| Quorum fire latency | Time from assignment to 3rd review arrival — the critical-path latency for this pattern |
| Late review completion rate | Fraction of cycles where all 5 reviewers complete — tracks whether archival value is actually realized |
| Decision reversal rate | How often later reviews (4th, 5th) would have changed the editorial decision — measures quorum quality |
| Reviewer completion distribution | Variance in completion times across reviewers — high variance validates the latency benefit of partial join over AND-join |
| Node | What it does | What it receives | What it produces |
|---|---|---|---|
| Assign Reviewers | Selects 5 domain experts and dispatches the paper to each via AND-split | Paper + reviewer pool | 5 parallel reviewer activations |
| Reviewer 1..5 | Each reviewer reads the paper and produces a structured assessment: score, recommendation, key concerns | Paper text + scoring rubric | Structured review object |
| 3-of-5 Quorum (partial join) | Fires after 3 reviews arrive. Passes the 3 collected reviews to the editorial agent. Remaining 2 reviewers continue unaffected. | 3 review objects (first arriving) | Quorum review set for editorial decision |
| Editorial Decision | Aggregates quorum reviews and produces a structured editorial recommendation with reasoning | 3 reviews from quorum join | Accept / revise / reject decision + rationale |
| Origin of Value | Where it appears | How it is captured |
|---|---|---|
| Future Cashflow | Editorial decision latency | Decision latency is determined by the 3rd-fastest reviewer, not the 5th. For human reviewers with high variance, this can reduce median decision time by 40..70%. |
| Governance | Quorum threshold (M) | M encodes the governance policy — how many independent assessments are required before a binding decision is made. Changing M changes the quality-latency tradeoff without restructuring the workflow. |
| Conditional Action | All 5 reviewer branches | All 5 reviews consume compute. The 2 post-quorum completions have archival value only. If cost is a constraint, consider whether reducing N to 3 (eliminating the 2 archival reviewers) is acceptable. |
| Risk Exposure | Quorum selection bias | The first 3 reviewers to complete may be systematically similar (same institution, same methodology). Early-completers may not be representative of the full reviewer panel. Fix: track reviewer identity in the quorum join to detect demographic or methodological concentration. |
Quorum as governance primitive. Partial joins are a precise encoding of quorum rules: "N agents are asked; M must agree before action is taken." This maps directly to committee decision-making, distributed consensus, and multi-model validation. The join threshold M is the single parameter that controls the quality-speed tradeoff — making it externally configurable is a high-leverage design choice.
Reviewers with a strong positive prior about the topic complete faster and tend to submit early. The quorum fires on 3 favorable reviews, and the 2 skeptical (slower) reviewers are excluded from the decision. The editorial agent receives a skewed sample. Fix: add reviewer metadata to each review object and run a diversity check in the quorum join — if the first 3 completions are too homogeneous, hold for one more arrival before firing.
A reviewer submits an incomplete review (e.g., only filled in the score field, not the rationale). The partial join fires when this empty review is included in the quorum. The editorial agent receives a 3-review set with insufficient content for a decision. Fix: each reviewer branch validates its output against a schema before emitting the completion token. Incomplete outputs are rejected and the reviewer is re-prompted.
The editorial decision completes and is communicated. Two hours later, the 4th and 5th reviewers submit reviews that strongly contradict the decision. These reviews are stored in the archival record but the decision is not reopened. This may be the correct behavior — or not, depending on policy. Fix: define explicitly at design time whether late reviews can trigger a decision review cycle, and if so, route them to a secondary evaluation node rather than directly to the archive.
| Variant | Modification | When to use |
|---|---|---|
| Weighted Quorum | Each reviewer has a weight; the join fires when accumulated weight exceeds threshold, not raw count | Reviewers have heterogeneous expertise levels; senior reviewer votes should count more |
| Diversity-Constrained Quorum | Join fires only when the quorum set satisfies a diversity constraint (e.g., at least 2 different methodological backgrounds) | Systematic bias from homogeneous early completers is a known risk |
| Rolling Quorum | After initial decision, subsequent completions can trigger a re-evaluation node if they disagree significantly | High-stakes decisions where late-arriving evidence should be able to reopen the question |
| Pattern | Relationship |
|---|---|
| 40.47 Blocking Partial Join | Same M-of-N logic but blocks re-entry until all N complete — use when cycle isolation is required |
| 40.48 Generalised AND-Join | Waits for all currently active branches — use when N is not fixed at design time |
| 60.66 Cancelling Partial Join MI | Cancels remaining instances after M complete — use when remaining work has no archival value |
| 20.25 Consensus | Higher-level pattern that uses partial join logic to reach agreement across multiple evaluating agents |