How does a JTMS differ from an ATMS?

A JTMS maintains a current acceptance context, while an assumption-based TMS records sets of assumptions supporting propositions across alternative environments.

Can a proposition have multiple justifications?

Yes. This is essential when a conclusion has more than one independent route of support.

Does a circular chain of reasons establish belief by itself?

A simple unsupported cycle should not create well-founded support. The system needs appropriate treatment of dependencies and grounding premises or assumptions.

AI Glossary / Graphs and symbolic AI / Advanced

JTMS

Definition

JTMS stands for justification-based truth maintenance system. It maintains the current accepted status of propositions by recording their supporting justifications and updating dependent conclusions when assumptions or other supporting information change.

What it means

A reasoning system can derive a conclusion correctly from assumptions that later cease to be accepted. Without dependency tracking, the old conclusion may remain in the database even though its support has disappeared.

A JTMS addresses this bookkeeping problem. Nodes represent propositions, and justifications record dependencies between supporting nodes and conclusions. A node labeled IN is currently accepted; OUT means it is not currently accepted. OUT is not automatically a proof of the proposition's logical negation.

The word truth should not suggest an external fact-checking service. The system maintains support relative to supplied premises, assumptions and justifications. Incorrect premises or missing dependencies can still produce an unsuitable belief state.

How it works

The problem-solving component supplies propositions and justifications. A simple positive justification says that a conclusion is supported when all its antecedents are IN. An enabled assumption or a premise can provide a starting point for supported reasoning.

When a support changes, the JTMS propagates the consequences through the dependency network. If a conclusion loses its selected support, the system checks whether another valid justification remains. A proposition with independent surviving support should not be permanently withdrawn merely because one reason was removed.

Implementations can flag contradiction nodes and invoke a handler when they become supported. Deciding which assumption to revise requires a policy or problem-solving decision. Terminology varies: nonmonotonic variants additionally allow support conditions involving nodes being OUT. Such conditions must be distinguished from classical statements that those propositions are false.

Three practical examples

A configuration reasoner revises derived compatibility conclusions when a selected component assumption is withdrawn.
A diagnostic exercise tracks which candidate conclusions depend on a tentative observation and which have independent support.
A rule-based assistant explains why a conclusion remains accepted after one supporting source is removed.

Think of it this way

Think of a set of conclusions attached to their supporting notes. Removing a note should trigger review of conclusions that rely on it, while a conclusion backed by another sufficient note can remain.

The analogy explains dependency maintenance. A JTMS uses explicit formal relationships; it does not assess a note's credibility or invent missing reasons merely because a conclusion sounds plausible.

Worked example

Retracting one reason without losing another

Consider a fictional configuration model with assumptions M = MainSupplyAvailable and B = BackupSupplyAvailable. There are separate justifications M supports P = PowerAvailable and B supports P. A further justification P supports R = ReadyForTest. There are no other supports.

Initially M is IN and B is OUT. The main-supply justification makes P IN, which makes R IN. Three of the four nodes are currently accepted: M, P and R.

Next enable B. Now both supply assumptions are IN. P has two possible supporting justifications, although a simple implementation may select one as its current support. All four nodes are IN.

Retract M. P must remain or become IN through the valid B justification, and R remains supported through P. The resulting accepted set is B, P and R. An implementation that permanently removed P and R would have ignored the alternative support.

Finally retract B too. Neither justification for P is satisfied, so P becomes OUT; R also becomes OUT because its only antecedent is unsupported. All four nodes are now OUT.

This final state says that the system currently lacks support for PowerAvailable and ReadyForTest. It does not, without additional represented rules or facts, prove NoPower or NotReady. The trace illustrates settled states after propagation, not every temporary label used inside a particular implementation.

Illustrative case study

The fictional Willow Reasoning team tests a rule-based configuration assistant. A conclusion unexpectedly disappears when one assumption is disabled, even though another independent justification remains available.

Reviewers trace the recorded reasons and compare the settled labels with the four-stage example. They find that their surrounding application treated removal of the selected reason as removal of the proposition itself. They revise that integration so that alternative support is considered.

Willow also changes its user-facing explanation from false to unsupported when a node becomes OUT. This fictional case illustrates two different errors: losing valid alternative support and confusing belief status with logical negation. It does not report a deployed product's behavior.

Common misconceptions and mistakes

OUT means proven false. It means not currently accepted in the relevant JTMS labeling scheme.
Losing one justification always retracts the conclusion. Another valid support may remain.
A JTMS independently discovers all domain rules. The surrounding reasoner supplies the propositions and dependencies.
A contradiction tells the system exactly which premise is wrong. Resolving it requires additional assumptions or a revision policy.

Business use

The concept is useful where conclusions must respond predictably to changing assumptions and remain explainable through recorded dependencies. It can support symbolic configuration, diagnosis and exploratory reasoning.

Before applying it, define which items are premises, which assumptions can change and how contradictions are handled. Check that external updates are translated into the correct enable, retract or justification operations. Audit explanations against the stored support graph rather than generating plausible prose afterward. The mechanism improves dependency accountability within its model; it does not establish that the domain model or input evidence is correct.

Check your understanding

After M is retracted while B remains enabled, why should ReadyForTest remain IN?

Answer: B still supports PowerAvailable, which supports ReadyForTest. Removing M eliminates one reason but not the surviving support chain.

Related terms

Justification
Dependency network (graphical model)
Belief revision
Contradiction
ATMS
Truth maintenance system
assumption
dependency network
nonmonotonic reasoning

Learn next

Replace the ReadyForTest rule with one requiring both PowerAvailable and SensorAvailable. Mark SensorAvailable as an assumption, then trace what happens when it is withdrawn while BackupSupplyAvailable stays IN. Keep unsupported conclusions separate from explicit negative propositions.

Sources

Forbus describes JTMS labels, well-founded support and the distinction from ATMS in this primary research paper: https://www.qrg.northwestern.edu/papers/Files/QRG_Dist_Files/QRG_1988/Forbus_1988_QPE_journal_article.pdf

Forbus and de Kleer's author-hosted implementation shows justification, retraction and alternative-support operations: https://www.qrg.northwestern.edu/BPS/jtms/jtms.lisp

Their companion examples illustrate dependency changes: https://www.qrg.northwestern.edu/BPS/jtms/jtms-ex.lisp

The supply trace and Willow Reasoning case are original examples. Implementation details are not presented as a universal interface for every JTMS.

Previous: Jonckheere's trend test Next: Judgment sample