SARAL Archive: When rules meet discretion
Scope of this page
This page documents the SARAL pilot system, its deployment setting, and the design observations that motivated the next experimental stage. It does not present formal empirical results or publish inferential claims from the exploratory field deployment.
- Design observation: structured rule outputs did not fully determine downstream human decisions in the pilot workflow.
- Design observation: unstructured notes, contextual cues, and local administrative constraints appeared to shape operator behavior.
- Design observation: non-binding model outputs may have affected how operators justified discretionary action.
- Boundary: these observations are retained here only as exploratory context for subsequent experimental validation.
SARAL v1.3 · System brief
What the v1.3 build was, and what it did
SARAL v1.3 is a decision-support probe developed to examine how human discretion interacts with rule-based eligibility logic in street-level welfare triage. It is the build deployed during the January 2026 pilot documented on this page.
Early prototype. Concept only; not field-tested.
Field-deployed in Jan 2026. Frozen and retained as an archived system build and design reference.
What v1.3 does
- Intake: Accepts structured attributes (age, income, document presence).
- Rule engine: Evaluates inputs deterministically against formal policy criteria.
- Output: Provides an interpretable recommendation (Eligible, Ineligible, Escalate).
- Visibility: Displays a rule trace showing why a rule fired, alongside conflict cues.
- Workflow variation: Surfaced an optional, non-binding AI risk score in specific pilot conditions.
What v1.3 does NOT do
- No auto-approval: It does not make automated final decisions.
- Not a service: It is not a live, public-facing production system.
- No outcome tracking: It does not track long-term welfare disbursement or poverty alleviation outcomes.
- No citizen-facing interface: Operator-facing only; citizens never interact with the rule trace.
System at a glance
Where SARAL sits in the decision chain
Final outcomes are determined at the discretion layer; SARAL is advisory. The deployment was designed to probe how formal rule presentation interacts with human judgment under realistic administrative conditions.
Methodology
Deployment design & measurement context
Deployment design
The SARAL deployment used an operator-facing workflow to examine how formal rule presentation interacted with human discretion. Outputs were advisory and non-binding throughout.
- System role: decision-support interface with rule trace and conflict cues
- Workflow variation: operator-facing configurations varied across intake conditions
- Constraint: final authority remained with human operators and local gatekeeping rules
This page records deployment structure and instrumentation context rather than formal evaluative claims.
Unit of archival description
The archived workflow centers on a triage instance: one intake interaction producing a SARAL recommendation and a final operator action.
- Setting: Maharashtra, PMAY verification context
- Scheme: PMAY (housing eligibility)
- Archive focus: workflow structure, information environment, and decision dynamics
Information environment
SARAL consumed policy rules and structured fields, while additional context often appeared in unstructured notes and discretionary review.
- Structured: tabular intake fields used by the rule engine
- Unstructured: field notes with contextual evidence
- Discretion layer: operator judgment and local administrative constraints
Instrumented workflow features
- Final action states: Approve / Reject / Escalate
- Rule presentation: recommendation plus rule trace
- Conflict surfacing: disagreement cues between encoded rules and note content
- Latency capture: triage timing as a process marker
These are presented as features of the archived workflow, not as validated outcome measures on this page.
Telemetry and logging
- Latency: wait and processing time stamps
- Decision states: recommendation plus final action
- Reason capture: override reason codes when provided
- Sequence: exposure order to notes and conflict cues when instrumented
Logs describe operator behavior inside the pilot workflow; they are archived here as system artifacts.
Operational definitions (archival use)
Archived cohort: the frozen pilot workflow retained for documentation and system interpretation.
Override: recommendation ≠ final operator action.
Schema gap: decision-relevant information exists outside structured fields evaluated by rules.
Binding constraint: the factor determining whether tool output translates into action.
Unstructured evidence: field notes or contextual cues not formally evaluated by the rule engine.
Data handling
- Identifiers: names removed; direct identifiers excluded from archive
- Notes: excluded or heavily sanitized prior to export
- Contact data: phone numbers converted to irreversible citizen hashes and not retained post-session
Coding and design development
- Purpose: identify recurring decision patterns for future controlled testing
- Use: taxonomy-building, interface interpretation, and Phase 2 design refinement
- Boundary: observations retained here are not framed as formal inferential findings
Operator review structure (exploratory archive only)
Allocation in the pilot workflow was operational rather than designed for formal identification. This archive therefore treats operator variation as contextual background, not a basis for comparative claims.
Archive note: The retained logs capture how the workflow functioned in practice, but this page does not use them to make formal statistical, causal, or publication-style claims.
Acknowledgement: Operator review in the archived pilot workflow involved remote volunteer operators without direct field access. Identities are withheld to keep the archive non-personal and publication-safe.
Archive observations
What the deployment surfaced
1. Schema limits in practice
The deployment indicated that structured eligibility logic did not fully capture the information operators treated as decision-relevant. In practice, unstructured notes and contextual cues appeared to influence how formal recommendations were received.
Design implication: a rule engine can remain internally coherent while still omitting decisive practical context.
2. Salience of non-binding model outputs
The pilot suggested that non-binding risk indicators may shape operator attention and justification, even when the system does not automate final action. This observation motivated closer experimental study of how displayed signals affect discretionary review.
Design implication: advisory signals may influence behavior through salience and institutional cover, not only through formal decision logic.
3. Recurring unencoded cues
The archived workflow surfaced recurring categories of unencoded cues, including material-affluence proxies, contextual spillovers, normative heuristics, and administrative feasibility constraints. These categories now serve as inputs to the Phase 2 synthetic profile design.
Design implication: the relevant explanatory unit may be the unencoded signal rather than the formal rule alone.
4. Friction and discretionary scrutiny
Process logs suggested that some cases attracted greater scrutiny than others, pointing to the possibility that review effort is not neutral but selectively allocated. This informed the need for a controlled design that can separate judgment, attention, and institutional alignment.
Design implication: process friction itself may be part of the decision mechanism under study.
Emergent taxonomy used for Phase 2
The exploratory deployment informed a provisional taxonomy of failure points and unencoded signals. These categories are presented here as design scaffolding for future controlled testing.
| Category | Description | Role in Phase 2 |
|---|---|---|
| SCHEMA_GAP | Decision-relevant facts appear outside structured fields evaluated by rules. | Signal family for synthetic profile manipulation |
| NOTE_SALIENCE | Free-text or contextual notes appear to shift attention or justification. | Narrative cue treatment |
| ADMIN_CONSTRAINT | Local caps, process bottlenecks, or implementation constraints shape practical decisions. | Institutional alignment test |
| RISK_SIGNAL | Displayed risk or conflict cues may affect discretionary review even when non-binding. | Salience and override experiment |
Limits
Archive limits
Exploratory status
The archived deployment was exploratory and is retained as system context, not as a finalized human-subject evidence base.
Single-region context
The deployment setting was localized, which limits general interpretation even at the descriptive level.
Non-binding system role
SARAL did not make final decisions. Any archive interpretation must therefore center human judgment and institutional process.
No formal inferential claims here
This page is intentionally bounded to deployment description, archival documentation, and design motivation for later experimental work.
Interpretation
Bounded interpretation
What this archive supports
- A clear description of SARAL as an advisory decision-support system.
- A documented gap between encoded rules and broader information environments.
- Motivation for a controlled Phase 2 study on unencoded signals and discretion.
What this archive does not claim
- Not a formal results page for a completed human-subject study.
- Not a causal or statistical evaluation of welfare outcomes.
- Not a claim that the archived deployment alone validates the taxonomy or downstream theory.
Theoretical framing
The archive is situated within broader questions about street-level discretion, administrative decision-making, and the limits of algorithmic formalization. Its role here is conceptual and design-oriented: to motivate a more rigorous next-stage experiment rather than to stand in as definitive evidence.
Data & privacy
Release status
This archive hosts instruments, protocols, interface structure, and sanitized system artifacts from the SARAL pilot. Names and direct identifiers are removed. Free-text verification notes are excluded or heavily sanitized prior to any export. Phone numbers used during intake were converted to irreversible citizen hashes and were not retained after the session.
SARAL suggestions are non-binding. Operator decisions remain authoritative. Released artifacts are sanitized and archived for documentation, transparency, and future research development.