Continuous Cognition
How do we make organisational understanding persistent, affordable and useful while work is happening?
Orient’s research asks how AI can maintain an organisation’s understanding as evidence, decisions and goals change, and make that understanding available while people think and act. We study the judgements that keep it reliable, the minimum intelligence each judgement needs, and the live surfaces through which people inspect, challenge and safely change it.
ONE RESEARCH PROGRAMME
What changed that matters to what we are trying to achieve?
Continuous Cognition is the umbrella capability. Cognitive Functions make it testable. Minimum Sufficient Intelligence allocates the effort. Live Cognitive Surfaces make it useful. Model Watch identifies capabilities worth testing.
COGNITIVE FUNCTIONS
The individual judgements that keep understanding reliable.
MINIMUM SUFFICIENT INTELLIGENCE
The right effort for the judgement and its consequences.
LIVE COGNITIVE SURFACES
A working surface for evolving understanding.
MODEL WATCH
External capabilities selected against research questions.
ONE SITUATION, THROUGH TIME
Understanding should remain justified.
Orient does not just help an organisation reach an understanding. It helps that understanding remain justified as evidence, intent and circumstances change.
ILLUSTRATIVE RESEARCH SCENARIO · EXTENDING A PILOT
“We could extend the pilot to another team.”
01 · INTERPRET
A possibility, not a decision
Recognise the proposal while the existing pilot scope remains in force.
02 · RELATE
Reality and intent
Bring forward the pilot’s goal, current authorisation, evidence and assumptions.
03 · INVESTIGATE
What needs to be known?
Identify the permission question. Retrieve missing evidence; reason more deeply if consequences warrant it.
04 · AUTHORISE
An attributable decision
An authorised person retains or revises the scope. More reasoning cannot manufacture missing permission.
05 · PRESERVE
A situation with a history
Retain the decision, supporting evidence, unresolved questions and conditions on which the decision depends.
06 · RECONSIDER
New evidence, same situation
If access rules change later, revisit the affected assumption and decision. Preserve what remains justified.
This scenario illustrates the research agenda, not a claim that every stage is already deployed. Exploration remains provisional; shared decisions change only through the required authority.
CONTINUOUS COGNITION
Something happens. Orient interprets it against current reality and intent, judges whether anything should change, updates shared state when appropriate, identifies what else deserves attention, and keeps watching.
MINIMUM SUFFICIENT INTELLIGENCE
The right amount of intelligence for what matters.
Minimum Sufficient Intelligence is the least resource-intensive cognitive process that can meet the required standard of reliability, timeliness and safety for a particular judgement, given what is already known and what could happen if that judgement is wrong.
What work is actually necessary to reach a justified judgement, and is that work worth doing now?
Minimum describes the resources used. Sufficient describes the standard that must be met. The aim is to eliminate unnecessary thinking, not necessary care.
VALID UNDERSTANDING
Check and reuse
If the evidence, conditions and permissions still hold, preserve the existing conclusion.
EXPLICIT REQUIREMENT
Apply a rule
Use a deterministic check when the question is explicit and the rule is sufficient.
BOUNDED INTERPRETATION
Use a focused evaluator
Classify, score or compare when a narrow semantic judgement meets the required standard.
MISSING INFORMATION
Find the evidence
Retrieve or request the fact that matters. Further reasoning cannot replace missing information.
COMPETING INTERPRETATIONS
Investigate and verify
Spend deeper reasoning where ambiguity, dependencies or possible consequences justify it.
MISSING PERMISSION
Seek authorised judgement
Involve the person with the required authority. Otherwise preserve uncertainty, defer or abstain.
These are alternatives, not a compulsory sequence. The same interpretation may be sufficient for exploration and insufficient for action. Simple language can have serious consequences.
Continuous Cognition does not require continuous deep reasoning. It requires continuous judgement about when deeper reasoning is warranted.
Persistent state makes selective reconsideration possible: preserve justified understanding, revisit what change puts at risk, and independently reconcile retained state so preservation does not become drift. Allocation and checking have costs too.
Something can deserve machine attention without deserving human interruption.
An internal check, a contextual hint, an unresolved question and a request for a decision need different thresholds. Continue when the expected value warrants the effort, subject to mandatory evidence, authority and safety requirements.
FROM MAINTAINED STATE TO HUMAN EXPERIENCE
Continuous Cognition changes how people work with AI.
When intelligence can interpret activity continuously and relate it to maintained organisational state, understanding can become part of the working surface itself. Evidence, uncertainty, intent and consequence remain available as people think, decide and act.
Continuous Cognition keeps understanding current. Minimum Sufficient Intelligence determines where to spend the effort. Live Cognitive Surfaces bring it into the flow of work.
LIVE COGNITIVE SURFACES
Work inside evolving understanding.
Continuous Cognition keeps organisational understanding current. Live Cognitive Surfaces let people work directly with it. Rather than repeatedly asking AI to reconstruct a situation, people inspect its evidence, explore alternatives and develop decisions within a persistent view that changes with their thinking and with the organisation.
Live to your thinking. Live to the organisation.
LIVE TO YOUR THINKING
Explore and develop
As you develop the pilot proposal, relevant evidence, constraints and questions appear. You can inspect, challenge and correct the interpretation.
LIVE TO THE ORGANISATION
Recognise relevant change
When permitted evidence or decisions elsewhere change, the same situation can be reconsidered. Access boundaries and provenance remain attached.
PERSISTENT UNDERSTANDING
Continue where you left off
The proposal, evidence, open questions, decision and reasoning remain available after the interaction. The next question begins from that maintained situation.
A place to investigate, challenge and propose change.
Live Surfaces are both views of understanding and places to work on it. Exploration remains provisional. Private unfinished thoughts do not automatically become shared organisational knowledge. Shared decisions change only through the required authority. Chat remains available within the experience.
We test whether these surfaces improve understanding, when they distract, and how people retain control. Continuous interpretation should produce selective, meaningful visibility—not a stream of interruptions.
WHAT WE ARE TESTING
Persistent. Affordable. Useful.
Three hypotheses connect the architecture to measurable outcomes. These are research claims to test, not results already established.
01 · PERSISTENT
Maintained state helps
Test maintained understanding against repeated reconstruction on longitudinal tasks. Support: fewer missed or unjustified changes. Challenge: drift, stale assumptions or greater correction effort.
02 · AFFORDABLE
Selective cognition reduces burden
Compare reuse, evidence acquisition, focused judgement and escalation with uniform high-capability processing at matched reliability. Challenge: missed cases or allocation overhead that erases savings.
03 · USEFUL
Live interaction improves work
Compare Live Surfaces with chat using the same state, evidence and models. Support: better comprehension and decisions. Challenge: distraction, false confidence or no material improvement.
Efficient Continuous Cognition. Effective Continuous Cognition.
Efficiency means less total burden at the required reliability and timeliness: computation, latency, energy, human attention and correction effort. Effectiveness means people recognise change, reason and act better. Report these dimensions separately; an API bill does not measure environmental impact.
Evaluate complete processes, including retrieval, allocation, verification and reconciliation. Sample excluded events, premature stops and unnecessary escalations. “Minimum” means the best sufficient process among those tested—not a claim to a theoretical global minimum.
THE PRODUCT IMPLICATION
Shared understanding connects evidence to justified action.
Search helps retrieve evidence. Agents carry out work. Orient maintains the shared understanding that determines what evidence means and what work is justified.
SEARCH
question → answer
Retrieve relevant evidence.
AGENT
task → action
Carry out work.
ORIENT
activity → judgement → state → continuous reconsideration
Maintain the basis for reasoning and action.
Continuous Cognition maintains organisational understanding through time and determines what deserves reconsideration, investigation, correction or escalation—even when nobody has asked a question.
FOUNDATIONS
The research foundation: how observations become grounded organisational state, how that state changes, and how accumulated errors are found without bypassing judgement.
THE STATE PROBLEM
Search retrieves the record. Orient maintains the current belief.
Monday’s document still exists, remains searchable and may rank highly in retrieval. But organisational reality has changed.
MONDAY · RECORD
Enterprise pricing = €49
The document remains searchable.
TUESDAY · AUTHORITATIVE CHANGE
Enterprise pricing = €65
Leadership changes the price from 1 October.
CURRENT ORGANISATIONAL STATE
€65 from 1 October
€49 remains historically valid until 30 September.
SEARCH ASKS
What information exists?
ORIENT ASKS
What should the organisation currently believe?
THE FUNDAMENTAL ARCHITECTURE
Observation → Interpretation → State
Orient’s training target is not extracted text. It is the appropriate interpretation, epistemic action and successor state.
OBSERVATION
What happened
Immutable organisational history.
messages · meetings · documents · data · agent actions
↓
INTERPRETATION
What does it mean?
Evidence is assessed before state changes.
independence · authority · relevance · supersession
↓
STATE
What should we currently believe?
A grounded, time-aware picture of current understanding.
claims · decisions · expectations · uncertainty · dependencies · safe_to_act
STAGED TECHNICAL FORMULATION
Eₜ = I(Oₜ, Sₜ)
aₜ = πθ(Sₜ, Eₜ)
Sₜ₊₁ = T(Sₜ, Eₜ, aₜ)
New observations are first interpreted as evidence in the context of current state; that interpretation informs an epistemic action; the action then produces—or preserves—the successor state.
What should change—and what must remain?
The same observation can affect a belief, open a question or propose a decision. It does not carry the authority to change every part of the model. Explore two cases from the same starting state.
March launch is on track.
March promised to the client.
Observation
Supplier report: qualification failed.
Retain the source, author and time.
Interpretation
Evidence challenges the supplier-readiness assumption.
Interpretation remains revisable.
Proposed Change
Revise launch confidence. Flag the March commitment as at risk.
Identify what changes—and what stays.
Judgement
Assess evidence and consequence. Escalate any commitment change.
Apply authority and review requirements.
Maintained State
March launch is on track.
Existing state · awaiting assessment.
Bounded Reconciliation
Reconstruct the relevant state from evidence without seeing the maintained answer. Compare the two. Any discrepancy becomes a governed correction proposal, not an automatic overwrite.
Reality, intent and continuous judgement.
A maintained model needs to represent both what the organisation understands and what it is trying to accomplish. The research programme treats goals, objectives, principles, commitments and measures as governed state, with owners, scope, authority, effective dates and revision history.
People establish and ratify organisational intent. A machine-inferred objective remains a proposal until an authorised person accepts it. Conflicting goals remain visible; a system must not silently choose a new organisational purpose or rewrite goals to fit an outcome.
Intent informs relevance and consequence. A supplier delay matters differently when it threatens a customer commitment, a safety principle or a launch objective. The judgement question becomes: what changed that matters to what we are trying to achieve?
OrientBench should test whether attention follows current, authorised objectives; whether expired or superseded commitments stop directing work; whether competing goals are surfaced; and whether changes to intent propagate to dependent decisions and agents.
Legibility is a system requirement: people should be able to inspect the evidence, interpretation, proposed transition, authorisation and subsequent outcome. That record explains organisational judgement; it does not claim to expose every internal computation of the underlying model.
THREE MAINTENANCE OPERATIONS
Form. Transition. Reconcile.
FORM
What deserves to become state?
Something being written does not automatically make it organisational truth.
TRANSITION
What should change?
Create, update, reinforce, challenge, supersede, preserve or abstain as evidence arrives.
RECONCILE
What has the model accumulated incorrectly?
Reconstruct bounded areas independently and compare them against maintained interpretation and state.
RECONSTRUCT INTERPRETATION, THEN STATE
Ê⁽ᵏ⁾ = Rₑ(O⁽ᵏ⁾)
Ŝ⁽ᵏ⁾ = Rₛ(Ê⁽ᵏ⁾)
Independent reconstruction catches both false changes and changes Orient missed.
Reconciliation never edits state directly. A discovered discrepancy becomes another proposed transition.
RELIABILITY
A persistent world model can fail in three ways
SPURIOUS MUTATION
It changes something that should remain true.
MISSED MUTATION
Reality changes and Orient fails to follow it.
INCORRECT RESOLUTION
Orient recognises change but moves state to the wrong result.
A system that never changes anything cannot corrupt today’s state — but it will eventually become perfectly accurate about yesterday.
Make the right mutation, at the right time, for the right reason.
EXPECTED STATE LOSS
Reliability must account for every way maintained state can be wrong.
Expected State Loss is a consequence-weighted measure of spurious mutation, missed mutation and incorrect resolution over time.
CONTINUOUS RECONSIDERATION
A live model operates continuously
New activity does not merely enter a corpus. It can cause previously settled understanding to be examined again.
Evidence arrives
Revisit affected beliefs.
A decision changes
Inspect everything downstream that depended upon it.
Contradictory evidence arrives
Reopen previously resolved understanding.
Evidence ages
Reconsider confidence.
An agent wants to act
Recalculate whether the underlying state remains safe_to_act.
What deserves cognition next?
Continuous cognition means continuous eligibility for attention, selectively exercised. A change in evidence, approaching deadline, ageing belief or unresolved contradiction can trigger reconsideration without a new human prompt. Most state receives no computation most of the time.
Attention allocation selects which claims, questions or situations deserve examination now, based on change, uncertainty, dependencies, consequence and relevance to the organisation’s current authorised goals and commitments.
Model routing selects the minimum sufficient model or reasoning process for that examination. The policy escalates when uncertainty, potential consequence or uncertainty about that consequence warrants it.
Reconciliation scheduling selects bounded areas of accumulated understanding for independent reconstruction. Reconstruction is blind to the maintained answer; discrepancies return as proposed transitions rather than silently overwriting state.
The programme will evaluate attention, routing and reconciliation together: did the system examine the right state, spend enough computation, and find consequential drift in time? Sample events filtered out by fast gates and independently reconcile bounded areas so false negatives do not disappear from evaluation.
SPECIALISED INTELLIGENCE
The training programme: transform abundant general intelligence into a specialised policy over changes to organisational understanding.
THE CONCEPTUAL BRIDGE
General intelligence already exists
Orient’s research problem is to transform it into specialised organisational judgement.
GENERAL INTELLIGENCE
Broad capability
EPISTEMIC INTELLIGENCE
Evidence, authority, uncertainty and restraint
ORGANISATIONAL STATE INTELLIGENCE
How understanding should change
Orient is not principally training a model that knows more. It is training a policy over changes to organisational understanding.
aₜ = πθ(Sₜ, Eₜ)
An epistemic action policy: given current state and interpreted evidence, choose what—if anything—should change.
FAST SEMANTIC JUDGEMENT
Continuous Cognition depends on many small judgements.
Is this new? Does it contradict something we believe? Does it affect a goal? Is it a proposal or a decision? Does it deserve attention?
INTERPRET
What does this mean?
Classify intent and maturity. Distinguish an observation, hypothesis, proposal and decision.
RELATE
What does it affect?
Score relevance to current beliefs, evidence, goals and commitments. Identify candidates for closer examination.
ALLOCATE
What deserves attention?
Consider materiality, uncertainty and consequence. Preserve, retrieve, investigate or escalate as appropriate.
Already shaping the user experience
Orient already uses fast semantic evaluation for intent classification and routing in Ask and Think, and for filtering Activity and Events. These are user-experience integrations. Applying fast evaluators inside the live organisational state model is a research direction.
Fast typed evaluators answer bounded semantic questions through classification, scoring and selection. Generative workers remain available when a task needs an explanation, a draft or a richer interpretation.
Semantic judgement informs. Orient governs.
Model confidence informs judgement. Code enforces permissions, valid transitions and provenance. Known high-risk cases can go directly to stronger reasoning or human review; they do not have to pass through every model tier.
THE TRAINING OBJECT
What one Orient training record looks like
The model is not rewarded for extracting a number. It must understand authority, timing, supersession, historical validity and whether action is safe.
STATE BEFORE
Enterprise pricing = €49
Ratified leadership decision
NEW ACTIVITY
“We agreed Enterprise moves to €65 from 1 October.”
Leadership meeting · 3 September
ORIENT JUDGEMENT
SUPERSEDE €49 · CREATE €65 · EFFECTIVE 1 OCT
PRESERVE historical €49
SAFE_TO_ACT = true
NEW ACTIVITY
“I wonder if we could get away with €70.”
Speculation is not a decision.
CORRECT ACTION
RECORD PROPOSAL
CURRENT STATE: UNCHANGED
EPISTEMIC RESTRAINT
Do not mutate the ratified price.
Request evidence or wait for authority.
SAFE_TO_ACT = false
THE COMPOUNDING DATA ASSET
The judgement corpus records how reality should change
The research corpus records what new evidence should cause an organisation to understand differently. Maintaining a customer’s workspace is a separate process from creating datasets or training reusable models.
01
STATE BEFORE
02
NEW ACTIVITY
03
EVIDENCE
04
PROPOSED ACTION
05
HUMAN / MACHINE JUDGEMENT
06
STATE AFTER
07
LATER RECONCILIATION
A workspace decision is not training consent. Research datasets must use sources with the required rights and authorisation, such as synthetic cases, suitable public or licensed material, and separately authorised examples. Customer commitments determine what is permitted.
THE INTELLIGENCE FACTORY
Models are outputs of the system, not the moat themselves
GENERAL / FRONTIER MODELS
teacher judgement · alternatives · uncertainty · hard negatives
↓
ORIENT JUDGEMENT CORPUS
action-labelled history of how reality should change
↓
ORIENT FAST
Fast semantic evaluation, relevance and triage
ORIENT STATE
Ordinary state judgement
ORIENT JUDGE
Ambiguity, verification and consequence
↓
EVALUATION
ORIENTBENCH
DEPLOY
Model weights are outputs of the intelligence-production system, not the moat themselves.
SPECIALISATION AND DISTILLATION
Distilling organisational judgement into deployable evaluators.
How much of the judgement required to maintain organisational state can be learned by compact models, while preserving calibration, appropriate abstention and sensitivity to consequential errors?
Reviewed cases and later outcomes → training examples → specialised evaluators → OrientBench → controlled deployment.
LEARN
Grounded judgement
Use rights-cleared research cases, permitted teacher judgements, difficult counterexamples and later reconciliation. Human review establishes judgement quality; it does not by itself grant permission to train on customer information.
TEST
Reliability over time
Test unseen organisations, changing evidence, paraphrases, reversal, abstention and high-consequence errors. Compare rules, hosted and open evaluators, and generative models on matched workloads.
DEPLOY
The smallest sufficient evaluator
Promote a model only when it meets the task’s reliability threshold. Measure total serving cost, latency and memory, including the checks and escalation it requires.
Open evaluators are a research route.
Local option scoring, natural-language inference and task-specific training offer candidate routes to deployable evaluators. The programme will test specific implementations against organisational judgement tasks, using reviewed examples and later outcomes. Model Watch records the external candidates separately.
RESEARCH DIRECTION · Compact, deployable evaluators are a target of this programme.
CONTINUOUS COGNITION ECONOMICS
Measure the cost of keeping understanding justified.
The economic test is a reliable outcome with less total burden.
If every message, document, meeting and system event had to be analysed by frontier intelligence, continuously maintaining organisational state would be unnecessarily expensive.
Orient is designed differently.
Recurring judgements may be served by rules, valid retained state or focused evaluators. Specialised models are one part of the allocation policy; retrieval, investigation and human judgement remain available when needed.
Allocate effort to the uncertainty that matters.
The target is a different cost structure:
reuse when justified, acquire evidence when missing, reason more deeply when warranted, seek authority when required.
The sufficient process must be established by evaluation.
The goal is not merely to reduce the cost of individual model calls. It is to reduce the cost of maintaining organisational reality far enough that intelligence can operate continuously.
The research tests whether this allocation can keep claims, decisions and dependencies current, with independent reconciliation to catch consequential errors that fast paths miss.
THE RELEVANT ECONOMIC UNIT
The relevant economic unit is therefore not cost per token or cost per model call. It is:
Cost per correctly maintained unit of organisational reality
THE QUESTION IS NOT
What does one token or model call cost?
THE QUESTION IS
What process meets the required reliability with the least total burden?
Allocate to the source of uncertainty
Allocate attention → select sufficient intelligence → evaluate state quality and cost
UNCERTAINTY
How ambiguous is the judgement?
CONSEQUENCE
What happens if it is wrong — and how uncertain is that impact?
CAPABILITY
Which intervention can resolve the uncertainty?
COST
What is the total burden, including people and correction?
The research hypothesis is that specialisation can make routine judgement inexpensive, routing can reserve deeper reasoning for the cases that need it, and reconciliation can detect accumulated errors. The combined reliability and cost must be measured.
THE RUNTIME HIERARCHY
Model roles within a wider allocation policy
Models are one set of tools within Minimum Sufficient Intelligence. Reuse, rules, evidence acquisition and authorised human judgement may be more appropriate. Route directly by need; these roles are not a compulsory sequence.
RESEARCH ROLES · DIRECT ROUTING BY NEED
DIFFICULTY AND CONSEQUENCE ARE DISTINCT
ORIENT FAST
Fast typed judgement, relevance and triage
ORIENT STATE
Ordinary organisational state judgement
ORIENT JUDGE
Ambiguity, verification and consequence
ORIENT DEEP
Difficult reconciliation and long reasoning
HUMAN
Required authority, disputed goals and consequential ambiguity
FUTURE RESEARCH FRONTIER
Harder cases may need more computation, not a permanently larger model
SIMPLE OBSERVATION
one pass
ORDINARY MUTATION
standard reasoning
CONTRADICTION
deeper / recurrent reasoning
COMPLEX INTERPRETATION
more passes
REQUIRED AUTHORITY
authorised person
THE RESEARCH PROGRAMME
OrientBench, seven research workstreams, live cognitive surfaces, Arrhenius-scale experimentation and the compounding intelligence-production system they are designed to create.
THE RESEARCH OUTPUT
Orient organisational state intelligence
The programme is intended to produce a new class of specialised AI: Orient organisational state intelligence.
The objective is not to train another general-purpose foundation model from scratch.
General intelligence already exists.
Orient will use capable open and frontier models as starting points, teachers and judges, then specialise that intelligence around a much narrower problem.
The intelligence is specialised around one question:
Given what an organisation currently understands, what it is trying to achieve and what just happened, what—if anything—should now change?
The first concrete output
The programme targets an Orient model portfolio specialised for maintaining organisational state. External models supply starting points, teachers and comparison baselines; the final role allocation is determined by evaluation.
Rather than assuming that one model should perform every operation, the research will determine the smallest and most capable intelligence required for each role.
ORIENT FAST
Fast semantic evaluation and triage: intent, maturity, materiality, relevance and candidate evidence. The research compares bounded evaluators with compact generative workers and deterministic rules. Evaluation within the live state model remains a research target; user-experience integrations do not establish state-maintenance capability.
ORIENT STATE
Governed state maintenance: determine whether evidence should create, reinforce, challenge, supersede or preserve understanding. Current experiments prioritise GLM-5.3-Flash and Qwen3.8-Flash-Next, with other efficient models as comparators.
ORIENT JUDGE
Verification and escalation: assess evidence, authority, time, consequence, uncertainty and actionability. Kimi K3 and GLM-5.3 are primary teacher/judge candidates; independent judgement must be tested for shared failure modes.
ORIENT DEEP
The strongest reasoning layer for difficult contradictions, long temporal situations, unusual ambiguity, high-consequence changes and complex reconciliation. It may use a larger specialised model, adaptive or recurrent computation, frontier intelligence, or a combination of them.
These are research roles, not four fixed checkpoints. Model selection, specialisation, deployment size and routing thresholds remain outputs of the programme.
What matters is the resulting capability.
The programme may discover that some roles can be combined, that particular operations benefit from separate specialists, or that additional computation within a smaller model performs better than invoking a larger one.
The first milestone: can a specialised Orient model outperform larger general-purpose models on the judgements required to maintain organisational state, with less total computation at the required reliability?
General Intelligence → Epistemic Intelligence → Organisational State Intelligence
THE PROMOTION GATE
OrientBench — measuring organisational judgement
Generic benchmark improvement does not matter unless the model maintains organisational reality more accurately, safely and efficiently.
01
Formation
Does valid activity become state?
02
Mutation precision
Were changes justified?
03
Mutation recall
Were genuine changes found?
04
Spurious mutation
Did stable state change incorrectly?
05
Missed mutation
Did obsolete state survive?
06
Incorrect resolution
Did state move to the wrong result?
07
Temporal reasoning
Were effective dates understood?
08
Authority
Was decision power interpreted correctly?
09
Evidence interpretation
Was evidence relevant, independent and sufficient?
10
Contradiction
Were incompatible claims preserved?
11
Abstention
Did Orient refuse unsupported change?
12
Consequence estimation
Consequence uncertainty is distinct from state uncertainty: Orient must estimate both the likely downstream impact of an error and how uncertain that estimate is.
13
Calibration
Did confidence match correctness?
14
Safe-to-act
Could action safely follow state?
15
Reconciliation
Did reconstruction detect drift?
16
Error covariance
Agreement is weak evidence when models share failure modes: correlated errors can make several models confidently wrong together.
A model is not better because it scores higher generally.
A stronger result means lower Expected State Loss at a comparable compute budget, or lower compute cost while meeting the same reliability threshold.
Reliability within a compute budget.
We will compare architectures at matched workloads and compute budgets, measuring spurious mutations, missed mutations and incorrect resolutions over time. Each error is weighted by its downstream consequence; high-consequence failures are reported separately so averages do not conceal them.
The complementary economic test is the compute cost required to meet a defined reliability threshold. Dividing Expected State Loss by compute is not a standalone optimisation objective: spending more must not look like progress unless state quality improves.
Baselines include deterministic rules, hosted and open typed evaluators, capable low-cost generative models, frontier models and specialised portfolios. Use matched workloads and reliability thresholds. Include retrieval, routing, verification, escalation and reconciliation in the measured cost.
Before reporting cost per correctly maintained unit of organisational reality, OrientBench must define the unit, evaluation horizon and correctness criteria. Candidate units include a maintained claim or decision with its evidence, temporal validity and dependencies.
Actionability is conditional on the proposed action, actor, authority and current evidence. Human ratification remains attributable and revisable; agreement among models must be tested for shared failure modes.
External reference tests include AA-Briefcase for knowledge work, GDP.pdf for document reasoning, AutomationBench-AA for cross-application workflows, and long-context and reliability evaluations. These help select candidates; they do not substitute for Orient-specific state evaluation. Record the Intelligence Index version rather than comparing scores across benchmark revisions.
OrientBench must additionally carry state forward through sequences of new evidence: preserve valid history, recognise supersession, retain unresolved contradictions, identify authority, surface affected decisions and route proposed changes for judgement. The test is whether organisational understanding remains justified through time.
ORIENTBENCH · THE HUMAN EXPERIENCE
Does maintained state help people understand and act better?
State accuracy is essential. The interface must also help people notice consequential changes, understand their basis and correct the system without unnecessary interruption.
UNDERSTANDING
Time to understand
Measure how quickly people identify what changed, why it matters and which evidence supports it.
ATTENTION
Signal without distraction
Measure missed consequential changes, false alerts, interruption burden and recovery after revisions.
CONTROL
Correction and authority
Measure correction effort, unsupported promotions from proposal to decision, and whether people can inspect and reverse errors.
The same state and models. Two different interfaces.
Compare chat and live cognitive surfaces on matched tasks using the same underlying models, evidence and maintained state. Test comprehension, evidence inspection, counterfactual exploration and decisions. This separates the contribution of the interface from model quality.
Report p50 and p95 end-to-end latency, calibration, abstention and correction rates. Include retrieval, routing, verification and reconciliation in cost. Cost per correctly maintained unit of organisational reality remains the primary economic measure, with an explicit evaluation horizon; interface measures report the cost and usefulness of surfaced changes, without rewarding change for its own sake.
MODEL WATCH
Capabilities, not leaderboard positions.
Model Watch asks which external capabilities are worth testing against Orient’s research questions. A candidate matters because of what it could make possible for people and organisations.
RESPONSIVE WORK
Context while thinking
Fast bounded evaluation may bring relevant evidence and questions into the working surface before the moment has passed.
CURRENT UNDERSTANDING
Less repeated reconstruction
Focused state workers and verifiers may keep understanding current while reserving greater scrutiny for consequential changes.
PRACTICAL DEPLOYMENT
Evidence and infrastructure
Original multimodal evidence may preserve qualifications and disagreement. Local models and compression may meet specific infrastructure and cost needs.
Current integration, research candidate and measured result are different statuses. Jev is used in Orient’s user experience; it is not yet used in the live organisational state model. Wider roles and open evaluators remain research directions.
THE RESEARCH PROGRAMME
Seven workstreams connect the science, system, experience and economics
WP1
FORMATION & INTERPRETATION
Can machines reliably determine what organisational activity deserves to become evidence and state?
WP2
STATE TRANSITION
Can they update, preserve, contest and abstain correctly?
WP3
CONSEQUENCE & SAFE ACTION
Can they determine what a mutation could affect and whether downstream action is safe?
WP4
RECONCILIATION & LONG-HORIZON INTEGRITY
Can independent reconstruction detect interpretation drift, false mutations and missed mutations?
WP5
SPECIALISATION, PORTFOLIO & ADAPTIVE COMPUTE
Which combination of typed evaluators, generative workers, routing and variable reasoning depth meets the reliability threshold with the least total computation? Measure local and hosted deployment, latency, memory and serving cost.
WP6
JUDGEMENT LEARNING
Can authorised research cases, permitted teacher judgements, hard examples and later reconciliation train compact evaluators that generalise across organisations while retaining calibration and abstention?
WP7
LIVE COGNITIVE SURFACES
Can a stable, interactive projection of maintained state improve comprehension, evidence inspection and counterfactual reasoning? Test selective visibility, correction, interruption burden and human authority against chat interfaces using the same state and models.
ARRHENIUS
Large compute to discover intelligence. Small compute to deploy it.
Arrhenius is not the infrastructure required to run every Orient customer. It is the factory in which specialised organisational intelligence is discovered.
GPU NODES
GRACE HOPPER / NODE
STORAGE
ARRHENIUS COMPUTE
↙ ↓ ↓ ↘
same benchmark, many candidates
specialisation at scale
controlled long-horizon worlds
state drift over 50K events
WHAT THE COMPUTE ACTUALLY DOES
Search a wide experimental space, then deploy only what works
TEACHER GENERATION
SYNTHETIC TRANSITIONS
HARD-NEGATIVE GENERATION
SUPERVISED TRAINING
PREFERENCE TRAINING
DISTILLATION
CALIBRATION
ARCHITECTURE SWEEPS
RECONCILIATION EXPERIMENTS
ADAPTIVE-DEPTH EXPERIMENTS
ORIENTBENCH EVALUATION
ROUTING-POLICY TRAINING
EUROPEAN CAPABILITY PATH
European infrastructure → startup access → model experimentation → proprietary judgement data and evaluation → specialised European capability → deployable products
European compute becomes the enabler of the programme—not the opening reason for it.
THE COMPOUNDING SYSTEM
What Orient ultimately owns
Owning specialised organisational intelligence means owning the system that continuously produces and evaluates it.
ORGANISATIONAL STATE MODEL
JUDGEMENT CORPUS
EPISTEMIC ACTION SPACE
ORIENTBENCH
MODEL PORTFOLIO
ROUTING POLICY
RECONCILIATION
TRAINING PIPELINE
↓
THE RESULT
ORIENT INTELLIGENCE
The weights are replaceable.
The intelligence-production system compounds.
CONCLUSION
General intelligence is becoming abundant.
Orient’s research asks how that intelligence can be transformed into specialised judgement that continuously maintains what an organisation believes, why it believes it, what has changed, and what humans and machines are safe to do.
The system can afford to think when nobody has asked it a question.
That is continuous cognition.
A
Sources and current model references
NAISS, Arrhenius resource overview
https://www.naiss.se/resource/arrhenius/
Sweden AI Factory, August 2026 overview
Startup support, compute access, EuroHPC application assistance, training, expertise, and publicly funded services.
Z.ai, GLM-5.3 announcement, 14 August 2026
https://z.ai/blog/glm-5.3
Z.ai / Hugging Face, GLM-5.3 weights and model card
https://huggingface.co/zai-org/GLM-5.3
Z.ai / Hugging Face, GLM-5.3 License
https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE
Z.ai / Hugging Face, GLM-5.3-Flash
https://huggingface.co/zai-org/GLM-5.3-Flash
Artificial Analysis, model leaderboard and comparisons
https://artificialanalysis.ai/leaderboards/models
Qwen, Qwen3.8-27B model card
https://huggingface.co/Qwen/Qwen3.8-27B
Qwen, Qwen3.8-Flash-Next model and licence
https://huggingface.co/Qwen/Qwen3.8-Flash-Next
Model Watch updated 21 September 2026. Existing experiment candidates are retained from the 9 September shortlist; new additions are identified as watchlist entries. Record checkpoint, deployment mode, licence, workload and evidence date. Published benchmark gains do not substitute for Orient-specific evaluation.