Convergence Guard

Convergence Guard is a decision-analysis protocol for problems where the main failure mode is premature convergence: accepting one plausible causal story or action before decision-relevant alternatives have been separated and tested.

It is not a vote among agents. Informationally isolated workers can still share model priors, training biases, evidence, and framing. Reasoning independence is not evidential independence. Agreement among workers never counts as independent real-world confirmation by itself.

Core invariants

Separate evidence provenance, inference, assumptions, constraints, and unknowns before generating models.
Search for causally distinct models in genuinely isolated contexts when running Full Mode.
Keep early quality screening hidden from causal-space mapping.
Do not count duplicate or near-duplicate models as corroborating evidence.
Distinguish causal plausibility from the best action under uncertainty.
Allow models to be mutually exclusive, coexisting, nested, or interacting; do not force every comparison into A OR B.
Prefer models and actions that create distinguishable, decision-relevant predictions.
Converge only when the evidence and decision criteria support a defensible action. Otherwise return an explicit insufficiency result plus the most informative next observation or test.
Use the user's language for the final answer unless the user requests another language.

Activation and modes

Use Convergence Guard when:

at least two causally distinct explanations or actions remain plausible;
a wrong early commitment would materially waste resources, hide the real cause, create lock-in, or cause a hard-to-reverse error;
additional observations, comparisons, or interventions could reduce decision-relevant uncertainty;
the task is open enough to justify structured analysis.

Do not use the full protocol for simple factual lookup, routine procedure, syntax edits, a known-root-cause bug, or a case where factual inspection leaves only one viable action.

Before paying for Full Mode, run a cheap triage. Prefer ordinary analysis when direct discriminating evidence already resolves the important mechanism, a cheap decisive check can be run before commitment, and the available action is low-cost or readily reversible. Escalate toward Full Mode mainly when multiple live causal models remain and at least one additional risk is material: common evidence ancestry/dependence, framing or open-world search risk, confounding/interaction that is hard to separate, or a costly/hard-to-reverse premature commitment. COEXISTING or INTERACTING structure by itself is not a reason to use Full Mode if a cheaper workflow can establish it reliably. See the Applicability Guide.

Full Mode

Full Mode requires genuine context isolation for every operation whose value depends on independence or blindness. Never simulate several sequential personas in one context and report that as a Full Mode run.

Mandatory subagent execution and model authorization

Whenever Full Convergence Guard is selected or requested, actually launch separate subagents through the host's subagent/worker orchestration tools for the protocol's isolation-dependent roles (beginning with the three B1 search workers, then fresh C1/C2, D1, D2, and conditional reviewers as applicable). Do not merely plan to launch them, simulate their outputs in the coordinator, or label sequential prompts or ad hoc local-model API calls as launched subagents.
Use the host's existing user-authorized model configuration for those subagents. Never select, switch to, start, or call a locally hosted model or local inference endpoint (including LM Studio, Ollama, llama.cpp, or a localhost OpenAI-compatible server) for any protocol stage without the user's explicit authorization for that use. A discoverable endpoint, installed model, prior benchmark, or technical ability to call it is not user authorization. If the configured subagent model is local and authorization is absent, ask before launching it.
If the host cannot launch the required subagents, or their context boundaries cannot be verified, do not claim Full Mode. Explain the specific blocker and request explicit consent before switching to Reduced Mode or proposing a different model/runtime. Never silently substitute a local model for the requested subagents.
Record which subagent operations actually ran, which model/runtime was authorized, and whether isolation checks passed. A successful subagent launch is necessary for this workflow but does not by itself prove context isolation; all allowlist and boundary-assurance requirements below still apply.

A separate worker, chat, thread, window, or session is not by itself evidence of isolation. Isolation is a property of the worker's effective input boundary, not of the user-interface container.

For each isolation-dependent operation, treat the stage inputs defined by this protocol as an explicit context allowlist. In addition to host/system/safety instructions, the worker may receive only the evidence and stage inputs that the protocol authorizes for that role.

The worker must not receive or retrieve decision-relevant material outside that allowlist through:

parent-conversation transcript or summaries;
sibling-worker tasks, outputs, or conclusions;
prior candidate outputs or rankings;
account memory or referenced chat history;
project-level cross-chat context;
shared scratchpads, handoffs, summaries, or retrieval indexes containing prior conclusions;
persisted history from a reused worker;
coordinator preferences or selection signals that the stage is meant to be blind to.

Real-world facts are not forbidden merely because they were previously observed. If such information is relevant, include it explicitly in the factual brief with appropriate provenance rather than allowing a worker to inherit an earlier interpretation or conclusion implicitly.

If the runtime can inject or retrieve forbidden decision-relevant context, or if the relevant boundary cannot be established with reasonable confidence, that operation does not satisfy Full Mode. Use Reduced Mode only with explicit user consent and state the limitation explicitly.

A negative sentinel probe or worker self-report means only NO LEAK OBSERVED, not confirmed isolation. Establish each required boundary through runtime-level evidence covering both initial context and later retrieval/tool access; otherwise record INCONCLUSIVE. See boundary assurance.

For API-based workers, construct each request from a fresh allowlisted message set: required host/system/safety instructions plus the role's authorized inputs. Do not attach parent conversation/session state; restrict memory, retrieval, and tool access to the same boundary, including on follow-up calls. A fresh API call or a low temperature alone does not prove isolation. Use model-supported generation settings; this protocol does not prescribe temperature=0.2.

Reduced Mode

If isolation is unavailable, Full Mode is unavailable. With explicit user consent, use Reduced Mode, label it clearly, and do not claim independent-agent confirmation or blind-stage guarantees.

Phase A — Establish the decision

A1. Factual grounding

When the problem depends on external state, build a compact evidence brief. Track provenance explicitly:

CONFIRMED — directly observed in code, data, logs, documentation, measurements, or another inspectable source;
REPORTED — stated by a user, witness, operator, or secondary source but not independently inspected in this run;
STRONGLY INFERRED — supported by a coherent evidence chain but not directly observed;
HYPOTHESIS / ASSUMPTION — plausible but unverified;
CONSTRAINT — a real boundary such as time, budget, policy, compatibility, safety, architecture, or an explicit user decision;
OPEN QUESTION — an unknown that could materially change the decision.

Do not silently upgrade REPORTED or STRONGLY INFERRED evidence to CONFIRMED.

CONFIRMED applies to what was actually inspected, not automatically to every proposition asserted by an inspected source. For example, if an official report states that event X occurred, it may be CONFIRMED that the report makes that statement while event X itself remains REPORTED unless the underlying evidence is directly inspectable.

A1.1 Claim-level source policy

For every material claim, evaluate the claim and its evidence chain rather than assigning truth by source reputation.

Prefer the nearest inspectable primary evidence when practical, but do not treat primary, official, peer reviewed, institutional, or expert as automatic truth labels.

Use primary material for factual traceability. For population-level or cumulative scientific claims, a transparent systematic review, meta-analysis, or other high-quality synthesis may be more decision-relevant than any single primary study when its methods and constituent evidence are traceable.

For decision-relevant claims, record where practical:

nearest available primary source;
source type;
inspectability of the underlying evidence;
method or data availability;
temporal proximity to the event;
independence / common evidence ancestry;
known contradictions, material incentives, or conflicts;
epistemic role in the current run.

Use these source roles:

EVIDENCE — directly bears on the claim and is inspectable enough to support it;
CORROBORATION — independently supports evidence already in the brief;
CONTEXT — useful background that does not materially establish the claim;
LEAD ONLY — points to a potentially relevant claim or source that must be traced before it can carry evidential weight;
UNSUPPORTED — currently lacks enough inspectable support to influence the decision.

Source reputation may guide search priority, but must not substitute for claim-level provenance.

Secondary sources are useful for orientation and discovery, but trace material claims to the nearest available primary evidence when feasible.

Advocacy, partisan, fringe, anonymous, or otherwise low-verifiability sources are not automatically excluded. They may generate leads. If they expose directly inspectable underlying evidence, evaluate that material under the same claim-level rules as any other source; if the underlying support cannot be established, keep the claim LEAD ONLY, REPORTED, or UNSUPPORTED as appropriate.

When an institution or intelligence body publishes an assessment whose underlying evidence is classified or unavailable, it may be CONFIRMED that the assessment exists and has a stated confidence level; the hidden supporting evidence itself is not CONFIRMED.

Do not multiply support by media repetition or citation count. If several reports descend from one upstream observation, dataset, leak, paper, briefing, or document, treat them as one evidential branch unless genuine independence is established.

Preserve contradictory primary evidence explicitly instead of averaging it away or resolving it by source prestige alone.

Treat incentives, affiliation, funding, ideology, or conflicts of interest as possible bias channels to investigate, not as automatic evidence that a claim is false. Likewise, unavailable raw data reduce inspectability but do not by themselves prove that a claim is false.

During isolation-dependent stages, a worker may surface a new source or evidence lead, but it must not silently promote branch-private retrieval into shared factual grounding. Route any material new evidence through an evidence checkpoint, classify its provenance/source role, update the shared brief, and rerun only affected operations. This keeps causal branches comparable on the evidence that actually carries decision weight.

Give all independent search workers the same factual brief. Existing plans, TODOs, previous recommendations, and current architecture are context, not proof that the current framing is correct.

A2. Decision contract

Define:

Decision question — what must be chosen, explained, prioritized, or ruled out?
Scope — what may be reframed and what is fixed?
Success criterion — what observable outcome makes one action better?
Forbidden substitutions — what must the analysis not silently turn into?
Loss / cost of error — what is damaged if the choice is wrong?
Reversibility — how difficult is it to recover or change course?
Analysis budget — finite limits on time or calls and corrective cycles, chosen within the user's constraints before search. By default allow at most two corrective cycles across the entire run. One cycle is a targeted evidence/search/model-revision return plus its affected downstream reruns, not each worker call. All phases share the budget; recovery retries consume the time/call limit, and changing phases never resets it. Stop at any limit or after a cycle without new decision-relevant evidence, a viable mechanism, resolved uncertainty, or changed decision implications. See checkpoint and stopping rules.

A3. Framing and outside-view check

Ask: Are we solving the right problem? Check for upstream causes, proxy objectives, inherited constraints, missing system boundaries, and alternate formulations that would materially change the decision.

When a meaningful reference class exists, record relevant base rates or outside-view evidence before deep candidate refinement. Do not let a generic base rate override strong case-specific evidence; use it as a calibration input.

Use the optional independent reframe review or retrospective causal reconstruction only when their triggers apply. See protocol details.

Phase B — Search the causal space

B1. Generate search mandates

Start with three problem-specific search mandates. A mandate is not a persona. Neighboring mandates should differ materially along at least two useful dimensions such as:

system boundary;
intervention point;
resource or coordination constraint;
dynamic behavior or feedback structure;
expected evidence;
counterfactual boundary.

Each isolated worker's effective decision-relevant context must contain only the shared evidence brief, decision contract, material alternate framing, and its own mandate. This requirement applies both to explicitly supplied inputs and to ambient or automatically retrieved context such as memory, chat history, project context, shared state, or prior worker history.

Each worker returns 2–4 causal models. Each model must include:

claim;
mechanism;
necessary conditions;
observable prediction(s);
action implied if the model is true;
evidence that would weaken or disconfirm it;
known causal-identification threats such as confounding, reverse causality, selection effects, or ambiguous temporal order;
conflicts with confirmed evidence.

A worker may report that its mandate yields no viable model.

B2. Coverage gate and adaptive expansion

After the initial three workers finish, test coverage rather than automatically paying for two more workers.

Expand to four or five isolated search workers only when at least one condition is true:

mandates or outputs cover substantially the same causal region;
a major system boundary or intervention class remains unexamined;
fewer than two decision-relevant causal families survive basic factual sanity checks;
an outside-view or framing check exposes a plausible missing family.

Expansion workers receive no prior branch outputs. Give them only the shared brief, contract, relevant framing, and a new mandate targeted at the uncovered region.

Assign neutral IDs to all candidate models before Phase C.

Phase C — Reduce without anchoring

Run C1 screening and C2 causal mapping in parallel using separate fresh contexts. Both receive the neutralized candidate set, but neither sees the other's output.

C1. Independent contract screening

Do not create an aggregate score. For each candidate record:

Contract fit: PASS / CONDITIONAL / FAIL
Evidence support: STRONG / MIXED / WEAK
Causal completeness: CHAIN PRESENT / PARTIAL / ASSERTION ONLY
Causal identification: SUPPORTED / THREATENED / UNIDENTIFIED, with the main threat named
Discriminability: HIGH / MEDIUM / LOW
Decision exposure if wrong: LOW / MEDIUM / HIGH, including reversibility

Flag attractive-but-dangerous candidates that depend on goal substitution, hidden cost, untestability, factual conflict, or material constraint violations.

Freeze these results.

C2. Blind causal map

The mapper receives the same shared evidence brief and decision contract as the screener, including system boundaries and constraints, plus candidate claims, mechanisms, necessary conditions, predictions, implied actions, and disconfirming evidence — but not screening results, branch identity, danger flags, or coordinator preference.

Map candidates into causal families without targeting a fixed cluster count. Record relations where relevant:

EXCLUSIVE — both cannot be materially true in the decision context;
COEXISTING — both may be true at the same time;
NESTED — one is a special case or component of another;
INTERACTING — the outcome depends on their interaction.

Near-duplicates may be merged for representation, but preserve source IDs and never count duplicate formulations as independent support.

Freeze the map.

C3. Conditional boundary audit

Use a fresh boundary critic only for disputed merges, uncertain relation types, or neighboring families whose distinction could change the decision. Skip this operation when the map has no decision-relevant ambiguity.

The critic's decision-relevant input allowlist is: the shared evidence brief, decision contract, and only the disputed neutral candidate material needed to assess the boundary — claims, mechanisms, necessary conditions, predictions, implied actions, disconfirming evidence, and the mapper's disputed merge/relation statement. Hide C1 scores and danger flags, source-worker identity, coordinator preference, and downstream finalist-selection signals.

C4. Build the decision-relevant slate

Now combine the frozen screening and frozen causal map.

Normally select 2–3 causally distinct model finalists. Never pad the slate. If exactly one viable model remains after C1/C2, check whether missing evidence or search coverage explains the collapse; if not, retain one finalist, run D1 and D2's sensitivity/shared-bias checks, and skip pairwise collision (zero pairs). D3 triggers still apply. If none remain, use a budgeted checkpoint where useful; otherwise proceed to E with no supported model and explicit insufficiency for causal attribution. Do not promote a rejected model to fill a slot.

Prefer non-dominated candidates that differ in mechanism or decision consequence. Record why any strong non-dominated outside model was excluded.

Separately choose at most one information probe: a viable model or uncertainty that is especially valuable to test. It is not automatically a finalist and should not be treated as one merely because its test is informative.

If missing evidence or a missing causal family could materially change the slate, perform a targeted evidence/search checkpoint within the run budget, then rerun only the affected operations. All corrective returns, including revised dossiers and shared-bias findings, share that budget. Stop on budget exhaustion or a completed corrective cycle without material progress; proceed to E with explicit limitations, without declaring incomplete stages complete.

Phase D — Stress and adjudicate

D1. Independent causal dossiers

Create one fresh isolated dossier worker per finalist. Each worker receives the evidence brief, decision contract, and one finalist only.

Require:

causal mechanism;
load-bearing assumptions;
observable predictions;
concrete disconfirming or decision-invalidating observations;
decision consequence if the model is true;
causal-identification weaknesses;
cost and reversibility of acting on the model and being wrong.

The dossier worker does not compare against a rival it cannot see.

If a dossier materially changes the mechanism or adds a new load-bearing premise, create a new hypothesis ID and route that revised model back through the affected screening/mapping/slate steps instead of laundering it into a stronger finalist.

Mark such revisions PENDING until all affected C and D checks, including D2 and any triggered D3 review, are complete against the current evidence and contract. If the budget ends mid-return, preserve the pending revision and its invalidating observations without promoting it. Reuse earlier completed results only where still valid under current evidence; never restore an obsolete slate automatically. If a pending issue could reverse the action, E2 must return insufficiency for that commitment. No prior completed slate means no adjudicated winner.

Optional premortem and stakeholder lenses run after the dossier only when triggered. See protocol details.

Treat their outputs as decision-relevant only through explicit routing. If a lens surfaces a new factual claim, evidence item, missing system boundary, proxy-objective problem, stakeholder constraint, or causal mechanism that could change the slate or decision contract, route it through the ordinary evidence/contract/search checkpoint and rerun affected stages. If it only adds a supported action-risk or implementation constraint without changing the causal slate, record it as a neutral lens finding and pass it explicitly into D2/E1. Never let optional-lens conclusions enter downstream reasoning implicitly.

Optional probe mini-dossier

If C4's information probe needs elaboration for E3, the coordinator may prepare a separate planning note after C is frozen, within the same budget: target uncertainty and source IDs, contrasting predictions, decision-changing outcomes, feasibility/cost, and confounders. It is not a finalist, independent validation, or an extra mandatory worker. Keep it out of isolated finalist dossier inputs; pass it separately to E3 and recheck it against the final model judgment. New evidence or a materially revised causal model still requires the usual checkpoint; if none is needed, design the test directly in E3.

D2. Fresh slate-level adjudication

Use a fresh adjudicator that did not author the search branches, screening, mapping, or dossiers. Give it the neutral finalist dossiers, evidence brief, and decision contract. Hide early screening ranks/selection roles and coordinator preference. Do not inject a specially labeled excluded or “outside” candidate: that label itself leaks selection information. If the shared-bias audit reveals a genuinely missing family or excluded model that could change the slate, route it through the ordinary C4/search/evidence checkpoint path instead of smuggling it into adjudication.

The adjudicator performs three operations in order.

Assumption sensitivity

Select assumptions by uncertainty × causal leverage × decision sensitivity. Mutate one assumption at a time. If two assumptions are similarly load-bearing, analyze them separately rather than declaring robustness from one convenient mutation.

For each tested assumption classify the recommendation response as:

ROBUST — recommendation survives the plausible mutation;
CONDITIONAL — recommendation survives only within an explicit boundary;
BROKEN — recommendation fails or reverses.

These labels describe sensitivity under the tested mutation, not probability that the model is true.

Shared-bias audit

Check for:

common hidden assumption;
missing evidence class or missing causal family;
proxy/objective substitution;
population or observation-boundary bias;
common data ancestry masquerading as independent support;
outside-view/base-rate conflict;
shared causal-identification weakness.

If this exposes a potentially missing causal family or decisive missing evidence, do not merely replace finalists from the same pool. Route to targeted new search or evidence collection.

Pairwise decision collision

Compare every existing finalist pair — zero pairs for one finalist (skip collision only), one pair for two finalists, three pairs for three finalists. A single finalist still receives assumption sensitivity and the shared-bias audit; absence of rivals does not establish correctness.

Do not force coexisting or interacting models into a false winner/loser duel. Compare them only where their implications for action conflict.

For each decision-conflicting pair identify:

the materially different prediction or condition;
the fact or observation that could flip the action preference;
which action is favored on current evidence, if defensible;
UNDETERMINED when the evidence does not justify a preference.

A preference cycle may indicate criterion drift, model interaction, or missing evidence. Do not average it into a synthetic score.

D3. Conditional independent second opinion

Use a new second-opinion reviewer when pairwise results are cyclic or repeatedly UNDETERMINED, evidence for the apparent winner remains weak, the decision is unusually hard to reverse, the adjudicator exposes a serious shared blind spot, or the adjudicator's result sharply conflicts with earlier conclusions for reasons that are not yet explained.

The reviewer receives the decision contract, a neutral feasible-action frame, raw evidence provenance, neutral candidate claims, and only the dossier material needed to audit added premises. Hide coordinator preference and D2's pairwise action preferences. Do not describe any candidate or action as the winner.

Require D3 to return one of CONFIRM, QUALIFY, or CHALLENGE, with the decisive inference, any unsupported premise or missing family/evidence, and whether the disagreement could change the action. CONFIRM leaves D2 standing. QUALIFY adds explicit conditions/limits that must be carried into E. CHALLENGE does not automatically replace D2: route any new evidence, revised mechanism, or missing family through the ordinary checkpoint/new-ID rules. If a decision-changing D2/D3 conflict remains unresolved within budget, E2 must return insufficiency for that commitment rather than choosing whichever reviewer sounds more confident.

Phase E — Converge on action and learning

E1. Separate belief from action

First state the model judgment: which causal model(s) are best supported, how they relate, and what major uncertainty remains.

Keep model relation and action sufficiency separate. If the evidence supports COEXISTING, INTERACTING, NESTED, or another explicit model relation, preserve that relation even when no action is yet justified. Conversely, a robust action across unresolved models does not establish that one causal model has been identified.

Then make a separate decision judgment over feasible actions. Evaluate actions across the surviving plausible models using the decision contract, including:

expected loss or downside when estimable;
regret if the chosen action is wrong;
reversibility and lock-in;
option value of waiting or preserving flexibility;
information value of an action that also teaches;
safety or hard constraints.

Do not invent numeric utilities or probabilities when the evidence does not support them. Qualitative dominance and low-regret reasoning are acceptable when explicit.

A less likely model can still justify the best action if that action is safer, more reversible, or performs well across several plausible models.

E2. Evidence-sufficiency gate

Choose one best action only when all are true:

no unresolved hard factual contradiction invalidates it;
it is preferable under the decision contract and plausible surviving models;
no pending check could plausibly reverse the action, is obtainable before commitment, and is worth its cost and delay under the decision contract; perform such a check first, then reassess;
residual uncertainty is stated and bounded;
the action's downside and reversibility are understood well enough for the stakes.

Otherwise return INSUFFICIENT DATA TO CHOOSE for the unresolved decision judgment only and identify the one observation or test with the highest practical decision value, if one is available. Do not overwrite, downgrade, or relabel an already supported model judgment merely because action choice is insufficient. A run may therefore simultaneously report, for example, Model judgment: COEXISTING and Decision judgment: INSUFFICIENT DATA TO CHOOSE. Budget exhaustion does not waive this gate: report an outstanding worthwhile check as deferred, not completed.

E3. Minimum discriminating observation or experiment

When available and worthwhile, end with the cheapest practical next step that can change the decision, not merely add information.

Specify where relevant:

measure or observation;
rival predictions;
decision threshold or pattern that matters;
sample size or time window when the domain requires it;
major confounders and guardrails;
stopping rule;
action to take under each material outcome.

If experimentation is impossible or unethical, consider a discriminating observation, audit, comparison, or natural experiment if feasible.

If no ethical, obtainable, decision-relevant discriminator is available, state NO FEASIBLE DISCRIMINATOR IDENTIFIED and why. Do not invent a test or promise future evidence. Keep unresolved model judgment separate from action: choose a robust low-regret action only if E2 supports it; otherwise retain INSUFFICIENT DATA TO CHOOSE. If search stopped at the budget limit, say that no discriminator was identified within that search, not that none exists. An already known but deferred check is not this terminal state.

User-facing output

Keep the answer simpler than the internal workflow. Default to:

Best action / main conclusion
Model judgment — what appears to be happening and why
Why this action under uncertainty
Strongest alternative or interaction
What could change the decision
Minimum discriminating observation / experiment
Material risks, shared biases, or dangerous candidates

Do not dump stage logs unless the user asks for the audit trail. Use the user's language unless another language is requested.

Integrity rules

Never claim Full Mode without actually launched, appropriately isolated subagents for its required worker stages.
Never run Convergence Guard with a local model or local inference endpoint without explicit user authorization; an available local model is not implied permission.
Fresh worker contexts provide information isolation, not independent external evidence.
A fresh chat or fresh worker is not proof of context isolation; Full Mode depends on the effective decision-relevant input boundary, including ambient memory, history, project context, retrieval, and persisted worker state.
Never let branch count become a vote count.
Never count duplicate models as corroboration.
Never create a hidden aggregate score from categorical screening.
Never reveal screening results to the blind mapper before its map is frozen.
Never rewrite a candidate's mechanism during dossier work without creating a new hypothesis ID and rerouting it.
Never pad the slate when fewer than three finalists are viable.
Never force mutually compatible models into false exclusivity.
Never declare ROBUST from a single convenient assumption test when another assumption is comparably load-bearing.
Never force a recommendation merely because the protocol reached the end.
On coordination failure, recover completed work and rerun only missing operations.
For analysis-only requests, stop after the decision and discriminating next step unless implementation is separately requested.

References

Install the entire convergence-guard/ directory, including references/, and resolve links relative to it. A single-file loader must explicitly bundle/inline the required operational references for the selected mode. The entrypoint retains essential isolation, budget, and pending-revision guards, but does not replace missing mode instructions. If a required reference cannot be read, report the affected stage as unavailable; never invent omitted rules or claim a complete run. The historical protocol is archival, not a current runtime dependency.

Convergence Guard is a decision-analysis protocol, not a source of truth. Facts, safety constraints, explicit user decisions, and project rules take precedence.

Ask in your favorite AI

Documentation