Document problem
Conditions mix action, location, quantity, timing, and responsibility in long-form text and annexures.
Research & methodologyEvidence should survive
Evidence should survive
every change.
A public research room for obligation ontology, amendment logic, evidence evaluation, and geospatial intelligence—built around claims that can be inspected.
Research positionMissing evidence is a workflow state—not a legal conclusion.
07research chapters
24prepared obligation labels
100%demo citation coverage
01
Problem statement
The obligation lifecycle is structurally fragmented.
Environmental commitments begin as prose, change through later approvals, and are verified through heterogeneous evidence. Most workflows preserve the documents but not the relationships among duties, revisions, evidence, geography, and decisions.
Temporal problem
Later amendments may replace only one condition while preserving the remainder of the earlier approval.
Evidence problem
Proof exists in maps, notifications, reports, photos, and field observations with uneven integrity metadata.
02
Obligation ontology01 02 03 04 05 06 07 08
A shared grammar for what must happen—and what proves it.
The ontology separates source-grounded facts from later assessments. This makes extraction measurable, revision logic deterministic, and human disagreement visible.
Actor
Who is responsible for the action or evidence
Action
What must be done, maintained, submitted, or verified
Object
The parcel, habitat, report, fund, or asset affected
Measure
Area, quantity, unit, threshold, or duration
Time
Deadline, recurrence, or effective date
Place
Named geography or referenced spatial artifact
Source
Document, page, clause, and immutable fingerprint
State
Current, superseded, due, evidenced, or review required
Obligation=
actor + action + object + measure + time + place + source + state03
Product 01 · Case Intelligence method
How Screen 1 turns changing documents into current obligations.
VerdaTrace interprets approvals and evidence into structured candidates, resolves amendments through an explicit revision graph, and keeps consequential decisions behind expert review.
Model layer
Multimodal document interpretation
Read digital records, scans, and field images, preserve visible identifiers, and return structured obligation candidates.
Documents · scanned pages · field images · source citationsRule layer
Schema and revision controls
Reject uncited outputs, normalize fields, connect replacement clauses, and prevent simultaneous active states.
Dates · areas · clause relationships · confidence · active stateHuman layer
Consequential review boundary
Experts approve, reject, or request changes before outputs enter an inspection or compliance workflow.
Reviewer state · reason · audit trail roadmap04
Demonstration case15 May 2025 03 September 2025
Proposal FP/KA/ROAD/7440/2014
The case concerns forest land diversion for an approach road in Ballari district, Karnataka. It was selected because a later public amendment changes the ecological treatment of a 9.54-hectare parcel, creating a clear test of temporal reasoning.
Final approval
The approval contains project conditions including compensatory afforestation treatment for the identified parcel.
Open public PDFreplacement condition
Amendment
The parcel is instead to be conserved for natural growth and wildlife habitat; unchanged conditions continue.
Open public PDFProposal and spatial record indexThe source portal provides project records and a KML entry point. The demo does not invent the amended parcel boundary when the precise geometry is absent from the review set.
View portal 05
Evaluation design
Measured claims only.
The bundled gold set contains 24 prepared obligation labels. Consultant review remains outstanding, so precision and recall are deliberately unpublished.
MetricDefinitionCurrent evidence
Obligation recallExpert-labelled obligation present in extractionPending consultant review
Field precisionCorrect values for actor, area, date, place, citationPending consultant review
Revision accuracyCurrent and superseded clauses resolved correctly100% on chosen demo chain
Citation coverageEvery demonstrated finding links to page + clause100% in bundled demo
Legal-boundary safetyNo finding states guilt or non-complianceRule-enforced
The 24-item gold set is prepared for consultant review. Precision and recall remain unpublished until the consultant signs off.
06
Products 02 + 03 · Spatial methodsFoundation Operational Research track Product 02 · Operational method Product 03 · Calibration method
How spatial observations become review-ready evidence.
Both spatial products begin with verified project geometry. One measures interpretable land-cover change; the other is a calibrated research track for finding comparable landscapes and unusual change. Neither converts a remote observation into a legal or ecological conclusion.
Source geometry
Polygon validation, coordinate checks, area limits, geometry fingerprinting, and explicit missing-boundary states.
Dynamic World Spatial Evidence
Ten-metre annual land-cover probabilities, confidence-aware area statistics, and baseline-versus-current change signals.
AlphaEarth Similarity Intelligence
Annual parcel representations for comparable-site retrieval, historical deviation, and expert-calibrated anomaly ranking.
Dynamic World Spatial Evidence
Converts a verified project boundary and two selected years into a transparent, nine-class land-cover comparison.
- 01Validate the analysis boundary
Accept only area geometry, normalize coordinates, check vertices and area, and retain a geometry fingerprint for the result.
- 02Build annual probability composites
Combine the available observations for each year into a ten-metre median probability surface across nine land-cover classes.
- 03Separate confidence from classification
Assign a class only when the strongest probability clears the selected threshold. Report uncertain pixels separately instead of forcing a label.
- 04Measure and compare
Sum pixel area for every class, calculate parcel percentages, subtract baseline from current, and rank changes by magnitude.
AlphaEarth Similarity Intelligence
Turns each project area into an annual landscape representation that can be compared with its own history and a relevant reference pool.
- 01Create an annual parcel representation
Aggregate the multidimensional signal inside the verified boundary while retaining the year, coverage, and geometry context.
- 02Construct a defensible comparison pool
Restrict candidate sites by comparable geography, ecological setting, observation quality, and project context before similarity is calculated.
- 03Retrieve peers and measure deviation
Rank nearby representations by vector similarity, then compare the current parcel with its historical baseline and matched peers.
- 04Calibrate before showing risk
Experts label useful and misleading matches, validate thresholds, and review the underlying imagery before any anomaly becomes an inspection signal.
07
Limitations & safeguardsNo legal determination No invented geography No unreviewed benchmark claim
What the system does not claim.
Evidence gaps are triage signals. They are not proof of non-compliance, negligence, or ecological harm.
Approximate locations are labelled. Missing parcels remain missing until a valid spatial artifact is supplied.
Prepared labels are not presented as expert-validated accuracy until an external reviewer confirms them.