---
title: Planner–Auditor Twin Architecture
url: https://www.emergentmind.com/topics/planner-auditor-twin
type: topic
---

# Planner–Auditor Twin Architecture

Searching arXiv for the specified Planner–Auditor Twin papers and closely related work to ground the article.
First, I’ll look up the main Planner–Auditor Twin paper by its arXiv identifier.
Querying arXiv for `2601.21113`.
Using the arXiv search interface to retrieve metadata for the core paper and related entries.
Planner–Auditor Twin denotes a dual-role architecture in which planning, forecasting, or action proposal is separated from auditing, validation, or outside-view correction. Across the cited formulations, the Planner is responsible for generating an inside-view plan or policy, while the Auditor applies deterministic checks, empirical benchmarking, or triggered review before or after execution. In the clinical formulation, this separation is used to improve safety and reliability in discharge planning over FHIR data; in project management, it appears as a Planner Twin and Auditor Twin implementing inside-view forecasting and outside-view due diligence; in SolarChain-Eval, it is instantiated as a Planner/Auditor layer interposed between an RL controller and a decentralized-energy-market environment [2601.21113][1302.2544][2607.08681].

## 1. Conceptual basis and lineage

The central principle of the Planner–Auditor Twin is decoupling generation from validation. In the clinical system, the high-level architecture explicitly separates “System 2” plan generation by an LLM-based Planner from deterministic verification by a rule-based Auditor. The stated purpose is to reduce hallucination, omissions, and miscalibrated confidence in structured clinical discharge planning. Auditor verdicts do not directly block the plan; instead, they trigger retry, logging, or replay mechanisms, which makes the Auditor observational and control-oriented rather than a hard gatekeeper [2601.21113].

A distinct but related formulation appears in project management through the opposition between the inside view and the outside view. Kahneman and Tversky are described as showing that decision makers systematically underestimate costs and durations and overestimate benefits. The cognitive roots listed are optimism bias, anchoring on best-case scenarios, and strategic misrepresentation. Flyvbjerg’s due-diligence procedure then institutionalizes an Auditor Twin that benchmarks the Planner Twin’s inside-view forecast against a reference class of completed projects. This formulation is not LLM-centric; its Planner–Auditor split is epistemic and governance-based rather than model-based [1302.2544].

SolarChain-Eval introduces a further variant in a cyber-physical setting. There, the Planner defines episode-level action bounds and audit rules, while the Auditor reviews and revises high-risk actions. This suggests that the Planner–Auditor pattern is not tied to a single substrate: it can operate over LLM-generated plans, conventional project forecasts, or RL actions, provided that proposal and audit remain distinguishable roles [2607.08681].

## 2. Canonical decomposition of roles

Across the three formulations, the Planner and Auditor serve different but structurally analogous functions.

| Setting | Planner role | Auditor role |
|---|---|---|
| Clinical discharge planning | Generates a structured JSON “ActionPlan” with four discharge tasks and an explicit confidence score | Runs deterministic checks for coverage, drift, and calibration; buffers high-confidence omissions |
| Project management due diligence | Produces the inside-view forecast, assumptions, and risk registers | Applies the outside view through reference-class benchmarking, bias adjustment, and due-diligence reporting |
| SolarChain-Eval | Returns episode-level action bounds and audit-policy parameters | Reviews triggered hourly actions, approves or revises them, and records structured intervention logs |

In the clinical setting, each action in the plan includes a type, details, `deadline_hours`, and a confidence score in \([0,1]\). In project management, the Planner Twin documents scope, assumptions, outputs, and key uncertainties, while the Auditor Twin selects a reference class, estimates \(f_{\rm ref}(x)\), and constructs an adjusted forecast distribution \(f_{\rm adj}(x)\). In SolarChain-Eval, the Planner provides bounds \(\alpha_{\min}, \alpha_{\max}, \ell_{\min}, \ell_{\max}, b_{\max}\) and audit thresholds \(\tau_{\nu}, \tau_{\Delta}, \tau_{\sigma}, \tau_{\kappa}, B_e, c_e\), after which the Auditor is invoked only when risk or instability triggers fire [2601.21113][1302.2544][2607.08681].

A common misconception is that the Auditor must always be a separate model that blocks unsafe outputs. The cited systems do not support that simplification. In the clinical paper, the Auditor is a deterministic module and does not directly block plans. In project management, the Auditor Twin is an independent due-diligence function anchored in reference-class evidence. In SolarChain-Eval, the Auditor is LLM-based but operates under explicit trigger logic, JSON-schema validation, clipping, and budget/cooldown constraints.

## 3. FHIR-native clinical discharge planning

In the clinical formulation, the Planner–Auditor Twin is implemented as an agentic, retrospective, FHIR-native evaluation pipeline over MIMIC-IV-on-FHIR. For each patient, the system fetches a FHIR R4 bundle from a HAPI FHIR server loaded with MIMIC-IV-on-FHIR. The bundle contains `Patient`, `Encounter`, `Condition`, `MedicationRequest`, `Observation`, and `Procedure`. A template-based `SummaryGenerator` then flattens these resources into a reproducible text narrative and structured JSON listings of active conditions, medications, labs, and procedures. Optional clinical guidelines, described as pre-chunked and indexed in a vector store, are retrieved and concatenated to the patient summary. The resulting context is passed to the Planner agent, exemplified by GPT-4o-mini, which drafts a JSON ActionPlan covering four discharge tasks: Follow-up Appointments, Medication Reconciliation, Patient Education, and Symptom Monitoring [2601.21113].

The Planner is also asked to self-report confidence: “How confident are you (0–1) that you included all required tasks?” The paper states that the Planner’s \(\hat p\) empirically correlates with actual coverage success, enabling downstream calibration analysis. The Auditor is a stateless Python module that applies three checks in order: coverage, drift, and calibration. Coverage is defined over the four required task categories as
$$
\mathrm{Coverage}_{\mathrm{All}}=
\mathbb{I}\bigl(
\mathrm{Has}_{\mathrm{FollowUp}}
\wedge
\mathrm{Has}_{\mathrm{Meds}}
\wedge
\mathrm{Has}_{\mathrm{Education}}
\wedge
\mathrm{Has}_{\mathrm{Monitoring}}
\bigr).
$$
Drift is measured by the \(L_1\) distance between the current action-type frequency vector \(p\) and a historical reference \(q\),
$$
L_1=\sum_i |p_i-q_i|,
$$
with a drift warning logged when \(L_1>0.4\). Calibration is tracked using Brier score and Expected Calibration Error over observed coverage success [2601.21113].

The clinical system also defines a specific failure mode: if \(\mathrm{Coverage}_{\mathrm{All}}=0\) and \(p_i\ge 0.8\), the episode is marked as a “dangerous failure.” Such cases are buffered for replay and may also trigger within-episode regeneration if self-improvement is enabled. This design places the emphasis on omissions and overconfident misses rather than only on free-form factuality.

## 4. Self-improvement loops, caching, and empirical results

The clinical framework uses two complementary self-improvement loops. Tier 1 is within-episode regeneration: when a coverage violation occurs and `self_improve` is enabled, the prompt is rerun with the previous draft plan included in context, for example with an instruction such as “In step 2, you missed Medication Reconciliation; please revise.” Tier 2 is a cross-episode discrepancy buffer and replay mechanism: episodes where confidence is at least \(0.8\) but coverage fails are stored and later replayed offline or on a schedule. The paper states that replay on these “hard” cases often yields complete plans on second or third try [2601.21113].

Optional context caching is orthogonal to self-improvement. With caching enabled, patient summaries and guideline embeddings are stored in memory across runs. The reported impact is a reduction in baseline latency from \(17.42\) s/episode to \(11.83\) s, described as \(-32\%\), with coverage increasing from \(32\%\) to \(52\%\). The abstract and detailed results identify the self-improvement loop, rather than caching, as the primary driver of gains.

The evaluation uses \(N=50\) retrospective MIMIC-IV episodes under five configurations.

| Configuration | Latency (s) | Throughput (ep/min) | Coverage | ECE |
|---|---:|---:|---:|---:|
| Baseline | 17.42 | 3.16 | 32% | 0.564 |
| Cache | 11.83 | 4.42 | 52% | 0.356 |
| SI | 19.65 | 2.77 | 86% | 0.062 |
| Cache+SI | 18.70 | 2.95 | 86% | 0.034 |
| Buffer Replay | 27.78 | 2.09 | 100% | 0.107 |

The multi-task coverage results are reported as \(32\%\) for baseline, \(52\%\) for context cache only, \(86\%\) for self-improve only, \(86\%\) for cache plus self-improve, and \(100\%\) for buffer replay over \(N=7\) buffered cases. Brier scores are \(0.544\), \(0.382\), \(0.126\), \(0.123\), and \(0.017\) for the same configurations. Per-task completion rates indicate that baseline performance is already high on Follow-Up at \(96\%\), but materially lower on Education and Monitoring at \(54\%\) each; with Cache+SI, the reported rates are \(100\%\) for Follow-Up, \(96\%\) for Medication, \(98\%\) for Education, and \(92\%\) for Monitoring. A plausible implication is that the Planner–Auditor Twin is particularly effective at reducing structured omissions in the weaker task categories rather than merely improving already dominant action types [2601.21113].

## 5. Outside-view due diligence in project management

In project management, the Planner–Auditor Twin is formulated as a governance mechanism for de-biasing front-end forecasts. The Planner Twin embodies the inside view: it prepares the business case, conducts sensitivity analyses, and supplies the initial risk register. The Auditor Twin embodies the outside view: it selects a statistically similar reference class, estimates the empirical distribution of outcomes, compares the forecast to that distribution, and constructs debiased expectations and confidence intervals [1302.2544].

The eight-step due-diligence procedure is embedded in the front-end “business case → go/no-go” cycle:

- **Step 1, Identify & Describe the Forecast**: document scope, assumptions, outputs, and key uncertainties.
- **Step 2, Establish an Outside-View Benchmark**: select a reference class of completed, statistically similar projects and estimate \(f_{\rm ref}(x)\).
- **Step 3, Compare Forecast to Benchmark**: compare inside-view \(E_{\rm inside}\) and claimed \(\sigma_{\rm inside}\) against the reference distribution.
- **Step 4, Check the Forecaster’s Track Record**: request documented ex-post accuracy data and estimate forecaster bias factor and spread.
- **Step 5, Identify Project-Specific Risks**: list \(\Delta\)Risks not captured by the reference class.
- **Step 6, Compute Debiased Expected Outcome & Confidence Intervals**: anchor on \(\mu_{\rm ref}\) and \(\sigma_{\rm ref}\), then overlay project-specific risks.
- **Step 7, Solicit Forecaster Comments & Sensitivity**: allow the original forecaster to justify deviations or provide missing data.
- **Step 8, Conclude on Bias & Decision Implications**: revise the go/no-go recommendation, contingencies, or financing structure.

One bias-adjustment form is stated as
$$
B_{\rm optimism}=\mu_{\rm ref}-E_{\rm inside},
\qquad
E_{\rm adj}=E_{\rm inside}+B_{\rm optimism}=\mu_{\rm ref}.
$$
The same section gives the sample reference-class estimates
$$
\mu_{\rm ref}=\frac{1}{N}\sum_{i=1}^N x_i,
\qquad
\sigma_{\rm ref}=
\sqrt{\frac{1}{N-1}\sum_{i=1}^N (x_i-\mu_{\rm ref})^2}.
$$
If approximate normality is acceptable, confidence intervals are formed from \(\mu_{\rm ref}\) and \(\sigma_{\rm ref}\); otherwise percentiles are read directly from the empirical cumulative distribution function [1302.2544].

The A-Train public–private partnership serves as the detailed case. The inside-view demand forecast was \(14.1\) M passengers in Year 1 and \(17.7\) M in Year 10, with a claimed downside at \(95\%\) confidence of a \(15\%\) shortfall, implying \(\sigma_{\rm inside}\approx 9.1\%\). The outside-view reference class comprised \(61\) rail projects. First-year actual/forecast accuracy had mean \(\mu_{\rm ref}=0.59\), median \(0.51\), and \(\sigma_{\rm ref}=0.33\). The paper reports variance understatement, \(\sigma_{\rm inside}(9\%)\ll \sigma_{\rm ref}(33\%)\), and an adjusted expected Year 1 demand of approximately \(8.3\) M rather than \(14.1\) M. The \(80\%\) and \(90\%\) confidence intervals are given as \([3.2\ \mathrm{M},14.2\ \mathrm{M}]\) and \([2.1\ \mathrm{M},15.5\ \mathrm{M}]\), respectively. The forecaster declined to share ex-post accuracy on more than \(20\) prior projects; two known first-year overestimates were \(250\%\) and \(150\%\). The reported decision consequence is that the investor declined to proceed under the original terms [1302.2544].

This version of the Planner–Auditor Twin differs from the clinical architecture in implementation but not in logic: the Planner proposes an inside-view forecast, while the Auditor interrogates it using independent evidence, explicit uncertainty quantification, and governance gates.

## 6. Cyber-physical market governance and broader significance

SolarChain-Eval adapts the Planner–Auditor Twin to decentralized energy markets formulated as a Gymnasium-compatible MDP \(\mathcal{M}=(S,A,P,R,\gamma)\). The hourly state is \(s_t\in\mathbb{R}^{12}\), the continuous action is \(a_t=(\alpha_t,\ell_t,b_t)\), and the global constraint \(\alpha_t+\ell_t\le 0.98\) is enforced. At the start of each 24 h episode, the Planner ingests summary statistics of past physics risk, slippage, and supply-demand gaps, then returns action bounds and audit-policy parameters together with a textual rationale. Raw RL outputs \(u_t\in[0,1]^3\) are decoded into bounded actions and rescaled when necessary to satisfy \(\alpha_t+\ell_t\le 0.98\) [2607.08681].

The Auditor operates through explicit trigger logic. It computes action jitter
$$
\kappa_t=\|\bar a_t-a_{t-1}\|_1,
$$
the normalized supply-demand gap
$$
\Delta_t^{audit}=\frac{G_t^v-Q_t^d}{\max(Q_t^d,\epsilon)},
$$
a hard-trigger indicator
$$
H_t=\mathbb{I}\bigl[\nu_t>\tau_{\nu}\ \lor\ \Delta_t^{audit}<\tau_{\Delta}\ \lor\ \sigma_t>\tau_{\sigma}\bigr],
$$
and a soft-trigger indicator
$$
S_t=\mathbb{I}[\kappa_t>\tau_{\kappa}].
$$
The final audit decision is budgeted and cooldown-constrained:
$$
\mathbb{I}_t^{audit}
=
E_t
\wedge
\Bigl[
H_t
\lor
\bigl(N_{e,t}^{audit}<B_e\land t-t_e^{last}>c_e\bigr)
\Bigr],
$$
where \(E_t=H_t\lor S_t\). If no audit is triggered, the bounded action is executed directly; otherwise the Auditor LLM returns `decision`, `final_action`, `risk_assessment`, and `reason`, all validated against JSON schema and clipped again if needed [2607.08681].

The reward includes market-clearing gain, liquidity drawdown, action jitter, unmet demand, physics risk, and spatial fairness:
$$
R_t
=
M_t
-\lambda_d D_t
-\lambda_j J_t
-\lambda_u U_t
-\lambda_p \Phi_t
-\lambda_f F_t.
$$
Every intervention is logged with `policy/setting`, `episode`, `step`, `trigger_context`, `proposed_action`, `final_action`, `action_delta`, and `auditor_rationale`. The abstract reports a clear utility-safety trade-off: RL agents improve market utility but can still produce unsafe behavior; when the physics penalty is removed, reward-maximizing agents exploit invalid generation and increase artificial liquidity. The Planner/Auditor layer improves auditability and mitigates selected risks, but it cannot fully compensate for a misspecified reward function [2607.08681].

Taken together, these formulations indicate that Planner–Auditor Twin is best understood as an architectural and governance pattern rather than a single algorithm. In clinical planning it separates LLM generation from deterministic reliability measurement and targeted replay; in project management it separates inside-view forecasting from outside-view due diligence; in cyber-physical markets it separates policy proposal from triggered intervention under explicit operational constraints. This suggests that the enduring feature of the pattern is not the choice of model family, but the insistence that planning and auditing remain distinct, inspectable, and instrumented for calibration, drift detection, or empirical correction.

Source: https://www.emergentmind.com/topics/planner-auditor-twin