Researching AI that can remember, investigate, verify, act, and adapt across long-running work.
This is an active public-safe research record. It explains the direction, the evidence currently available, and the questions that still need proof.
Current AI can be capable within one session and still lose continuity: important decisions, source evidence, corrections, and the reason a tool action was taken.
Forget decisions between interactions
Lose useful evidence in large contexts
Repeat mistakes that were already corrected
Act without a clear record of why
Humaniform is an integrated research architecture, not a bundle of disconnected features. Each area below has a distinct technical question and a practical outcome to test.
Keep important claims connected to their origin with source-linked memory, freshness tracking, contradiction handling, and separate confidence and evidence estimates.
An AI system that can explain what supports a conclusion, what is missing, and what would change it.
Investigate layered memory for working context, decisions, source material, project history, semantic knowledge, skills, and scoped adaptation.
Continuity across long software projects, changing codebases, research programs, and operational workflows.
Observe entities, events, changes, anomalies, and unresolved contradictions before a question arrives, then investigate deliberately when a goal or uncertainty appears.
A system that is prepared to reason about a living environment rather than only reacting to the latest prompt.
Explore fast paths for familiar work and deeper computation for ambiguity, novelty, risk, or long-horizon planning.
More useful work per unit of compute without treating every input as equally difficult.
Represent entities, state, events, actions, dependencies, and possible futures so the system can compare outcomes before acting.
Better planning, side-effect prediction, recovery after failure, and investigation of complex systems.
Give tool-using systems explicit identity, capabilities, data scope, risk budgets, evidence requirements, reversibility, and approval thresholds.
Increasingly capable agents that preserve human control, traceability, and clear boundaries.
Separate fast episode state, scoped reversible changes, and persistent updates accepted only after verification, replay, and retention testing.
Learning from experience without treating every interaction as permanent truth.
Treat external context as an environment to investigate through metadata, bounded retrieval, subproblems, specialists, deterministic tools, and source-backed synthesis.
Effective operation over large repositories, document collections, histories, and research corpora while active memory stays bounded.
Turn wrong sources, failed patches, missed contradictions, unsupported actions, and repeated strategies into structured replay and correction signals.
A persistent system that makes fewer repeated mistakes, not simply one that stores more conversations.
Study waste across low-value examples, redundant supervision, uniform computation, long-context activation cost, and unnecessary updates.
A research path toward more efficient training and post-training, evaluated with complete cost accounting.
The current model is an integrated stack. It keeps context and evidence close to the work while letting the system investigate uncertainty and act through controlled tools.
Streams, documents, code, interfaces, and tools
Entities, events, goals, uncertainty, and predictions
Exact evidence, important events, and project state
Retrieval, recursion, specialists, and tools
Alternatives, causality, world models, and planning
Evidence, tests, contradictions, and permissions
Controlled tools, software changes, and workflows
Replay, scoped adaptation, and validated consolidation
System layers
Humaniform Core
General reasoning, prediction, uncertainty, and adaptive computation.
Aegis Evidence
Source-linked memory, provenance, verification, and claim support.
Continuum
Persistent context, project memory, and long-running state.
AURORA
Retrofitting Humaniform capabilities onto existing models.
FORGE
Research into more efficient training, post-training, and continual improvement.
The following describes controlled prototype evidence and its limits. It is not a claim of broad general intelligence, universal reliability, or an independently validated production result.
Source-grounded answer behavior survived in the learned Humaniform path and collapsed in no-anchor and transformer controls.
Reproduced controlled prototype result
The model learned to produce source spans and citations instead of answer text alone.
Demonstrated on controlled grounding fixtures
A separate high-recall source bank raised supported source exposure and supported answering to 100% on the current fixture.
Controlled-fixture architectural validation
Removing, masking, shuffling, or corrupting evidence caused target grounded behavior to collapse or degrade.
Direct causal evidence
Joint answer + citation
0.3984
Source-span accuracy
0.6680
Confident-wrong rate
0.0000
Directly documented · causally supported
Joint answer + citation
0.0000
Controlled comparison
Joint answer + citation
0.0000
Controlled comparison
Correct source exposed top-3
1.0000
Correct source used by support
1.0000
Supported-answer accuracy
1.0000
Supported over-refusal
0.0000
Exact learned-anchor presence
0.3594
Compare optimized alternatives with the same hardware, data opportunity, time budget, tools, context, and evaluation suite.
Reject gains that materially damage language quality, reasoning, coding, grounding, calibration, safety, or generalization.
Remove or disable components to test whether they actually cause the claimed benefit.
Include training, processing, retrieval, verification, replay, storage, communication, failed runs, and runtime overhead.
Evaluate on new repositories, altered environments, hidden future questions, counterfactual changes, and adversarial inputs.
Pause or remove a subsystem when overhead, instability, or capability regressions outweigh its value.
Broad general-intelligence superiority
Human-equivalent comprehension
Safe autonomous self-improvement
Unlimited or zero-rot context
Universal agent reliability
Guaranteed training-compute reductions
Fully autonomous software development
Scientific or industrial autonomy
Consciousness or AGI
Source-linked memory, exact pointers, evidence selection, contradiction handling, and grounded action control.
Fast paths, fixed-shape routing, fused evidence operations, useful-work measurement, and matched compute accounting.
Event and entity tracking, future-state prediction, deferred-query memory, and proof-preserving compression.
Layered memory, large external context, bounded active state, exact source fallback, and context-integrity evaluation.
Tools, software workflows, permissions, verification, recovery, audit, and rollback.
Temporary neural memory, scoped adapters, verified local learning, retention, poisoning resistance, and safe consolidation.
Architecture scaling, specialist computation, efficient training, multimodal capability, and long-running deployment.
Confirm the owner-reported C2h, Gate D, and Gate E results from raw artifacts.
Test wrong entity, wrong field, wrong source, stale, contradictory, ambiguous, and corrupted bindings.
Raise complete human-auditable source-span recall from the reported zero baseline.
Improve exact top-one selection while preserving high recall.
Measure target quality, grounded work per second, latency, FLOPs, energy, cost, and memory.
If you are researching a serious AI problem, governed agent capability, or specialised business model, start with the evidence you need and the risk you need to control.
Start a research conversation