ObliTracker Contract Intelligence
Building a governed AI extraction engine that turns complex commercial contracts into structured, actionable intelligence — designed from the ground up for professional services delivery.
Contract intelligence that holds up under scrutiny.
ObliTracker is a governed AI system built by Changeable to extract, structure, and deliver commercial contract intelligence across three service tiers. It reads contracts the way a senior commercial lawyer would — methodically, precisely, and with a clear understanding of what matters operationally.
The design goal was not simply automating a reading task. It was building a system that produces outputs a client can act on, a practitioner can stand behind, and an organisation can use to manage its contractual obligations reliably over time.
That requires discipline. Not just capability.
Tiered service model
Clarity, Insight, and Intelligence tiers deliver graduated depth — from essential obligations tracking through to full risk analysis, anomaly identification, and negotiation notes.
Extraction rules applied every time
A locked rule set governs every extraction — covering business day arithmetic, liability cap precision, remedy chain mapping, and more. No shortcuts. No drift.
Mandatory human review loop
The system is engineered to stop. Part A is produced and reviewed before Part B is ever generated. The practitioner cannot be bypassed.
Full deliverable pack per client
Every engagement produces a polished PDF report, an interactive HTML dashboard, and a calendar-importable obligations file. Intelligence that flows directly into client operations.
Contract review is slow, inconsistent, and expensive.
Commercial contracts are where organisations take on risk, commit to performance standards, and lock in financial exposure — often for years at a time. But the process of actually understanding what a contract requires is almost universally ad hoc. Important obligations get missed. Key dates pass unnoticed. Liability positions are assumed rather than verified.
AI tools can read contracts quickly. The problem is that speed without discipline produces confident-sounding output that may be wrong, incomplete, or misleading. For a client acting on that output, the consequences are real.
What was needed was not a faster reader. It was a governed extraction system — one where the quality of the output is enforced by the design of the system, not by the vigilance of the person using it.
The constraints that shaped the build
- Every extracted item had to carry a source clause reference — no unsupported claims
- Inferred consequences had to be clearly labelled — never presented as express contractual terms
- Financial figures had to be verified against source schedules, not body text descriptions
- The system had to be incapable of producing Part B without human review of Part A
- Obligations, prohibitions, rights, and warranties had to be extracted as distinct legal categories — not merged
- Business day arithmetic had to be exact, with workings shown
- Liability caps had to be stated with full legal precision — no acceptable shorthand
- The deliverables had to be client-ready, not practitioner-ready
System architecture
The design challenge was not building something capable enough to read contracts well. It was building something disciplined enough to read them consistently — and honest enough to show its work.
Two-pass extraction methodology
Every contract goes through two structured passes before a single output item is written. The first maps document structure, cross-references, dependency chains, and remedy pathways. The second performs the full extraction against that structural map. Skipping the first pass is not possible — the methodology enforces the sequence.
Tiered scope control
Three tiers — Clarity, Insight, and Intelligence — define precisely which extraction layers are included or excluded. A Clarity engagement does not produce risk flags. An Intelligence engagement does not omit negotiation notes. Scope is determined by tier instruction, not by practitioner discretion, ensuring every client at a given tier receives the same analytical standard.
Locked classification rules
Obligations, prohibitions, rights and controls, warranties, and ownership and licensing positions are extracted into distinct categories. A “must not” is never placed in the obligations list. A “may” is never written as a duty. The classification rules are applied before any output is written — not as a style preference, but as a structural constraint on what the system can produce.
Enforced human review gate
After producing Part A, the system stops. Completely. It does not proceed to structured JSON output until the practitioner has reviewed the human-readable summary and given an explicit instruction to continue. A question, a correction, or an acknowledgement does not constitute approval. The gate is structural — it cannot be bypassed by implication.
NZ and AU norms reference layer
Intelligence tier extractions include a cross-check against a curated reference framework drawn from NZ and Australian contracting norms — covering liability cap benchmarks, privacy obligations, auto-renewal conventions, data residency standards, restraint enforceability, and more. Every norms finding carries a confidence label separating extraction fact from market commentary.
Living lessons architecture
Every extraction run produces lessons. Classification gaps, arithmetic errors, missed obligation categories, and process improvements are logged with a structured identifier and propagated to the core instruction set. The system gets more precise with every engagement — not through retraining, but through deliberate, versioned improvement of the governing rules.
What ObliTracker delivers
A contract intelligence engine that a practitioner can stand behind and a client can act on — reliably, every time.
Obligations you can manage
Every obligation is extracted with its trigger, timeframe, consequence, and source clause. Nothing is left to interpretation. Clients know exactly what they must do, by when, and what happens if they don’t.
Dates you can trust
Fixed, conditional, and calculated dates are all extracted — including notice windows, renewal deadlines, and post-termination obligations. Business day arithmetic is shown and verified. The calendar file imports directly to any calendar application.
Risk that’s named and rated
Intelligence tier extractions surface risk flags rated High, Medium, or Low — with the source clause, the risk description, and the operational consequence. Anomalies are typed and sourced. Nothing is buried in prose.
Liability position you understand
Aggregate caps, carve-outs, excluded loss types, material breach triggers, and indemnities are each extracted as discrete structured items. No simplification to “per event.” No assumptions about what the cap applies to.
Intelligence that flows downstream
Every extraction produces a structured JSON output that feeds directly into client-facing deliverables — PDF report, HTML dashboard, and calendar file. The practitioner downloads, quality checks, and delivers. No reformatting or rebuilding required.
A system that improves with use
Forty-six lessons captured from live extractions are applied to every subsequent run. The system knows what it has missed before and checks for it systematically. Accuracy compounds over time rather than resetting with each new engagement.
Speed was never the hard part.
AI can read a contract faster than any lawyer. That has been true for some time. The hard problem is not speed — it is discipline. Getting an AI system to apply the same analytical standard to every contract, every time, without drifting, without taking shortcuts, and without presenting inference as fact, requires deliberate engineering.
The most important design decisions in ObliTracker are not technical. They are structural. The enforced human review gate. The locked classification rules. The precision requirements on liability and arithmetic. The requirement that every item carry a source reference. These are the mechanisms that make the output trustworthy — not the capability of the underlying model.
Any AI can produce confident-looking contract analysis. The question is whether it is correct, complete, and honest about what it does not know. ObliTracker is designed around that question — not as an afterthought, but as its founding constraint.
What made this work
- Treating the instruction set as a legal instrument, not a style guide
- Engineering the human review gate so it cannot be bypassed by implication
- Separating extraction fact from operational inference at every level of output
- Building a living lessons system that propagates improvements forward
- Defining a locked classification vocabulary — obligations, prohibitions, rights, warranties — with no overlap permitted
- Requiring full legal precision on liability caps rather than accepting shorthand
- Designing the deliverable output as a client operations tool, not a practitioner reference document
- Accepting that a system that stops itself when uncertain is more valuable than one that produces confident output
Have a question about contract intelligence systems?
Common questions about building governed AI extraction engines for professional services delivery.
What makes ObliTracker different from simply using an AI to read a contract?
Any general-purpose AI can summarise a contract. ObliTracker applies a governed methodology — a locked extraction rule set, a mandatory two-pass reading process, a structured classification system, and an enforced human review gate. The output is not a summary. It is a structured extraction that every item traceable to a source clause, with inferences clearly labelled as such.
Why does the system stop after Part A?
Because the practitioner’s judgement is part of the product. The human-readable review summary (Part A) is produced first and reviewed by the practitioner before any structured output is generated. This is not a workflow preference — it is a quality control mechanism. Errors caught at Part A cost nothing. Errors that flow into client deliverables cost considerably more.
What does the three-tier model mean in practice?
The tier determines the analytical scope, not the extraction quality. Clarity delivers obligations, dates, and financial terms. Insight adds risk flags, anomaly identification, and a cross-check against NZ/AU contracting norms. Intelligence adds strategic recommendations and negotiation notes. Every tier applies the same extraction rigour to the sections it covers — the difference is what is included, not how carefully it is done.
How does the system handle contracts it has never seen before?
The extraction methodology is contract-agnostic. The two-pass structural read, the 25 extraction rules, and the classification system apply regardless of contract type, jurisdiction, or complexity. The system does not rely on pattern-matching against prior contracts — it reads each one from first principles against a consistent analytical framework.
What happens when the system is uncertain about something?
Uncertainty is surfaced, not hidden. Every extraction includes an items-flagged section where the system identifies anything ambiguous, any clause that requires verification against the original document, and any inference it has drawn that the practitioner should review before delivery. The system is designed to be wrong visibly rather than confidently.
Can a system like this be built for other professional services contexts?
Yes. The design pattern — a governed extraction methodology, tiered scope control, enforced human checkpoints, and a living improvement system — applies wherever structured analysis of complex documents produces professional service value. Due diligence, regulatory compliance review, procurement analysis, and policy assessment all have analogous requirements.
How does the system improve over time?
Every extraction run that surfaces a new issue generates a structured lesson entry. Each lesson identifies the category of error, what went wrong, and the fix required. Lessons are applied to the governing instruction set and versioned. Every subsequent extraction benefits from every prior one — systematically, not incidentally.
Start with a free Decision Clarity Session.
A Decision Clarity Session is a no-obligation conversation where we listen to where you are, what you are trying to achieve, and what is getting in the way. If you are thinking about building a governed AI system for document intelligence, contract analysis, or professional services delivery, you will leave with a clearer view of what that actually takes.