Platform · Models & self-improvement

Models trained on how business work actually gets done

One3 plans to ride frontier models for general intelligence and build its own models where only One3 has the data: Action schemas, Indian business documents and languages, and verified outcomes.

The thesis in five points

Don't race the frontier labs on general intelligence. Ride them. They will out-train One3 on that for the foreseeable future.
Win on intelligence only One3 can produce: models that know One3 Action schemas, tenant business context, Indian business documents and languages, and verified business outcomes.
The moat is the data engine, evals and business gyms, not the weights. Weights are superseded every 6–18 months. A verified-outcome dataset and a trusted eval suite compound.
"Self-improving" means a governed pipeline. The system proposes improvements, proves them on evals, and people approve promotion. It never silently rewrites its own permissions, policies or guardrails.
Compute follows evidence. Single build machine, then rented bursts, then owned nodes, then clusters, each step triggered by measured utilisation, cost and quality gaps.

Four model tiers: buy, adapt, build, maybe pretrain

Tier	What	When it is the right answer
T-A Frontier APIs	Closed models via API, multiple providers, BYOK	Hard reasoning, long-horizon planning, rare or low-volume tasks
T-B Open-weight, off the shelf	Open models served by One3 without training	Data residency, predictable volume, good-enough quality on evals
T-C One3-tuned	Open base plus One3 fine-tuning, preference optimisation or RL	High-volume narrow tasks, One3 schemas, Indic quality gaps, cost pressure
T-D One3-pretrained	Pretraining or large continued pretraining	Only if strict criteria are all met. Not planned before Year 6.

Tiers move on triggers, not ambition

Every task class sits in one tier at a time and moves only when measured triggers fire: eval quality within margin, cost above the serve-cost break-even, data-class policy, or enough consented training data. A monthly Model Refresh Review checks task class × tier × eval score × cost, and new open-weight releases are evaluated automatically against every relevant suite.

The planned One3 model family

Codename	Job	First wave
One3-Route	Task class, intent, language and complexity → routing decision	W2
One3-Guard	PII, prompt injection, unsafe content and exfiltration signals	W1 off the shelf, W2 tuned
One3-Embed	Multilingual and Indic retrieval embeddings	W0 off the shelf, W2 tuned
One3-Rerank	High-precision ordering of retrieved candidates	W0 off the shelf, W3 tuned
One3-Extract	Structured data from invoices, GST documents, bank statements, resumes and contracts	W2
One3-Act	Choosing and filling One3 Actions correctly, sequencing and refusing	W3
One3-Plan	Goal → plan decomposition and re-planning; starts on frontier APIs	W4 distilled candidate
One3-Critic	Verifier and reward model against rubrics, schemas and policies	W3
One3-Indic	Adapters for Indian-language business understanding	W3
One3-Voice	Indic and code-mixed speech recognition and consented TTS	W3
One3-Vision	Document layout, retail and industrial vision	W4–W5
One3-Forecast	Demand, cash, pipeline and workforce forecasts	W4
One3-Code	Pack development assistant trained on One3's own code	W3
Tenant adapters	Per-tenant style, terminology and procedures	W4

Built-in rules for the models

Model output is never the final authority on its own.

Guard detects, policy decides

One3-Guard is a detector. Permission decisions stay with deterministic policy and risk tiers. A model never grants permission.

Validators after extraction

GSTIN checksum, tax arithmetic and date checks run after One3-Extract. Nothing posts to the ledger without a validator pass and a risk-tier gate.

Forecasts must beat naive

A forecasting model must beat seasonal-naive and classical baselines on backtests before use.

Consent for voices

TTS fine-tuning happens only for voices with consent registry records.

Licence checks on every base

Base models are verify items. Licence notes decide whether a candidate can be used commercially or shipped in One3 Box.

Native-speaker eval sets

Indic eval sets use human-verified translations. Unverified synthetic data never enters evals.

What "self-improving" honestly means

It is	It is not
Continuous capture of outcomes, edits, failures and costs	Models editing their own weights in production
Automatic generation of candidate improvements	Agents changing their own permissions, budgets, autonomy, policies or guardrails
Automatic evaluation on versioned suites	Promotion without evidence
Human-approved promotion, canary rollout, automatic rollback	Training shared models on tenant data without opt-in
Measurable gains in cost, quality and outcomes	A claim that cannot be shown on a dashboard

Six improvement loops

Loop	What improves	Initial cadence	Who approves
L1 Prompt/Skill	Prompts, examples, tool descriptions, output schemas	Daily	Skill owner
L2 Retrieval	Chunking, query rewriting, embedder and reranker versions	Weekly	Knowledge admin
L3 Model	Tuned models and adapters	Monthly	AI lead + governance
L4 Process	Workflows suggested by process mining	Monthly	Tenant admin
L5 Agent policy (RL)	One3-Act and Plan behaviour trained in business gyms	Quarterly, W4+	AI lead + governance + tenant
L6 One3 codebase	One3's own code, tests, Packs and performance	Continuous	Tech lead

Immutable safety rules for every loop

The component being improved is never its own sole evaluator, and held-out sets are hidden from optimisers. Policies, permissions, risk tiers, autonomy thresholds, guardrails, eval suites and the Company Constitution change only through human-authored or human-approved changes with audit. Every promotion is reversible, every loop has a budget cap and a kill switch, and a loop for one tenant cannot read another tenant's data. Rented Digital Employees improve only as versioned releases through the Agent Training Gym.

Business gyms: safe practice environments

Reinforcement learning needs an action space, a state and a verifiable reward. One3 has all three: typed Actions, the object and event model, and outcome events.

1
Define the environment
A gym declares its objects, the allowed Actions, a simulator with scripted and persona customers, and deterministic world rules such as refund policy and SLA.
2
Score with verifiable rewards
Rewards favour resolved, policy-compliant outcomes and penalise violations. Deterministic policy checks outrank model-judged rewards.
3
Train on a curriculum
Start from real approved trajectories, then move from easy to harder seeds with ambiguous policies and noisy documents.
4
Defend against reward hacking
Held-out seeds and simulators, human reviews of top-reward episodes, frozen critic models during a run.
5
Prove it in shadow mode
Gains count only if they reproduce in shadow mode on consenting tenants, where agents propose and people act.
6
Canary, never auto-raise autonomy
Promotion runs at current autonomy levels. A gym result never raises an agent's autonomy by itself.

Consent and data-use tiers

Tier	Data	Default
T0 Operational	What is needed to run the service, debug and secure it	All tenants
T1 Tenant-private learning	Tenant data used only for that tenant: memory, retrieval tuning, adapters	Opt-in; adapters off until the tenant enables them
T2 Shared learning, de-identified	Scrubbed or aggregated examples for shared One3 models	Off by default; explicit opt-in with a contract clause
T3 Public, licensed, synthetic, internal	Licence-compatible public data, purchased data, synthetic data, One3's own usage	—

Evaluation science

Promotion depends on evidence that holds up to scrutiny.

Seven eval layers

Capability, One3-specific, safety and security, language, robustness, economics, and online outcomes.

Contamination control

Eval sets live in a restricted store with no training access. Overlap checks fail any training job that leaks eval items.

Calibrated LLM judges

Every judge rubric has a human double-labelled calibration set, bias tests and monthly recalibration.

Promotion statistics

Paired comparisons with confidence intervals. Safety must not regress; primary quality must improve or match.

Cost/latency Pareto

The router picks from the quality-versus-cost frontier. Frontier movement over time is the visible proof of improvement.

Public benchmarks as sanity checks only

Public benchmarks are regression signals, never promotion criteria.

The model plan in numbers

model tiers, T-A to T-D: 4
planned model family members, including tenant adapters: 14
governed improvement loops: 6
consent and data-use tiers: 4
compute stages, C0 to C5: 6

Questions about One3 models

Will One3 train on my company's data?

Not for shared models unless you opt in. Shared learning (T2) is off by default and needs an explicit opt-in with a contract clause. Tenant-private learning uses your data only for your own tenant.

Does One3 build its own foundation model?

Not in the near term. Pretraining is tier T-D and is not planned before Year 6. It would need a persistent capability gap, a legally clean corpus, dedicated funding and a clear commercial need.

Can the AI change its own rules?

No. Policies, permissions, risk tiers, autonomy thresholds, guardrails and eval suites change only through human-authored or human-approved changes with audit.

What are business gyms?

Simulated business environments, such as ticket resolution or invoice posting, where agents can practise with verifiable rewards before any change reaches real tenants. The first gyms are planned for W3.

How are Indian languages handled?

One3-Indic and One3-Voice are planned for W3 to close measured gaps in languages such as Hindi, Tamil, Telugu, Kannada, Malayalam, Marathi and Bengali, including code-mixed input.

Talk to us about AI that improves under governance

Learn how One3 plans to build models around your business outcomes while keeping your data under your control.