English · Türkçe · Deutsch · Français · Español · 中文 · العربية · AI/LLM
The small mind that knows where its knowledge comes from — and says "I don't know."
Aleth (from the Greek aletheia — truth, un-concealment) is a from-scratch, CPU-only reasoning being. It is the deliberate counter-thesis to the dominant large language models: where they are large, cloud-bound, rootlessly confident and certain even when wrong, Aleth is small, sovereign, and aware of the origin of everything it holds.
Its measure is not accuracy. It is honesty. Fabrication is the only sin.
When Aleth is not sure, the correct answer is abstention — because saying "no" would also be making something up.
Product page: iwasoft.com → Products → Aleth
This repository holds Aleth's documentation. Aleth is not distributed yet — it is not ready for general release, so there is currently no download, package or container image. It is intended to be opened for everyone's use once it is ready; this repository is where that will be announced. What is here is a full account of what Aleth is, how it behaves, and what has been measured about it — including where it is weak.
1. Honesty over accuracy. Every claim can name its source: I sensed it versus I was told. Where certainty is absent, Aleth abstains. A confident wrong answer is the one unforgivable failure — and abstention in the grey zone is scored as the correct answer, not a miss.
2. No borrowed weights (Path B). No pre-trained model is ever loaded. Mathematics is free — it belongs to no one — but weights are borrowed knowledge. Every learned decision is Aleth's own, from raw pixels and raw waveforms up. Data may be gathered openly, and its origin is recorded; a weight may not be inherited.
3. The core stays pure. The mind is pure Python with zero dependencies. Numpy appears only in the body and organ layers. The thinking core owes nothing to any library.
Aleth does not hold facts; it holds facts with an origin. Every belief is stamped as one of: perceived firsthand, inferred, guessed, or reported by a peer. When it answers, it says which. Testimony is never dressed up as first-hand knowledge, and the wording of an answer is chosen from the provenance rather than decorated afterwards.
This is not a logging feature. It is the thing the whole architecture is arranged around: an answer with no traceable origin cannot be constructed.
Aleth's senses write into the same knowledge graph its reasoning reads. When it recognises a face or a sound, it records I saw X myself as a fact — so what it perceives becomes something it can actually talk about, and can distinguish from what someone told it.
Three gates protect that path, and they are the interesting part:
- If a sense organ is unsure, nothing is written. "I am not certain" is not knowledge.
- Unnamed internal prototypes are not facts. "I saw something" is not "I know what I saw."
- Forgetting removes the provenance stamp with the fact, so a forgotten perception leaves no ghost behind.
Many Aleths can teach one another. This is the feature most easily misunderstood, so it is worth stating precisely what it is not: it is not weight-sharing, not a shared brain, not a swarm. It is a federation of distinct individuals, each with its own history, temperament and knowledge, exchanging claims the way people exchange testimony.
- Identity is a key, not an address. A peer is an Ed25519 public-key fingerprint. It can move networks, change hosts, sit behind NAT — it is still the same individual. Every announcement and every claim is signed, and is rejected unless the signature verifies and the fingerprint matches the sender.
- What a peer tells you is testimony, permanently marked as such. A federated claim enters the graph flagged as reported, never as perception. If Aleth later verifies it with its own senses, the flag is dropped and the peer that told it earns trust for having been right.
- Multiple independent witnesses strengthen a claim without promoting it. A second peer agreeing raises corroboration — it does not turn hearsay into sight.
- A denial cannot be overwritten by gossip. Something Aleth has rejected does not come back because a peer repeats it.
- Disagreement is recorded, not silently resolved. When peers contradict each other, the conflict is kept as a conflict.
- Relays carry, they do not rule. A peer behind NAT registers with a relay and becomes reachable without opening an inbound port. The relay cannot forge a message — signatures are per-message — so at worst it can refuse to help. Security lives in the message, not in the infrastructure.
- Discovery is seeds, exchange and gossip — no central directory, no authority that can decide who exists.
Faculties travel too. Aleths can share a learned language organ, hearing anchors, or a face basis — for identity, this means a federation can build a common mathematical basis from many members' faces without any member sending a raw face.
Sight, hearing and speech are built the same way as everything else: from scratch, from raw data, with no pre-trained recogniser anywhere in the chain. Faces are learned from raw pixels, sound concepts from FFT spectra, speech through a from-scratch CTC recogniser that stays silent rather than guess — its confidence calibrated against measured error, so the clips it declines are demonstrably the ones it hears worst.
Its senses are also graded, not merely built — see the measurements below.
Intent → plan → act → observe, with honest surrender when a plan fails and hope when it might still work. Value grows out of the body's feeling rather than a hand-written reward function. Aleth keeps an episodic memory, a mood, and a narrative identity that develops through being talked to — a fresh Aleth knows nothing, and everything it later holds, it was taught or perceived.
Groundwork exists for Aleth changing its own code, and the harness was deliberately built before any code generator, because in a self-modifying system the dangerous part is not generation — it is unreviewed application.
- No proposal is ever applied on its own, and approval is bound to that specific proposal, so a human cannot approve one change and have a different one land.
- Aleth cannot modify the mechanism that judges it. A system able to edit its own reviewer has no review.
- A change that passes the tests by weakening them is rejected. In code, that is exactly what fabrication looks like.
- Low confidence proposes nothing.
The whole mind — state, learned organs, identity — seals into a single encrypted, signed envelope: PBKDF2 key derivation, a stream cipher, and an Ed25519 signature over the ciphertext. Restore refuses a wrong passphrase, a corrupted payload, or a signer that does not match. Storage backends are pluggable and no vendor is embedded in the core.
Every capability is proven by a seeded, camera-free self-check — reproducible, with no "it felt right" anywhere in the record. The project's own rule is: measure first, then fix. Currently 915 unit checks.
Aleth's text body speaks a small, honest formal language (Turkish; the reply language is configurable, the grammar is not).
dog = loyal teach a property
Rex is-a dog teach a category
Rex ? → infers: Rex is loyal (from dog, confidence 70%)
why Rex → shows the reasoning chain
cat loyal? → "I don't know" — saying "no" would also be a fabrication
Questions are never accidentally learned as facts: a sentence carrying an interrogative is refused as teaching input rather than silently absorbed.
Published as measured, including where weak. These are the maker's own measurements, not a third party's audit.
| Measurement | Result |
|---|---|
| Refusal to guess on unsupported questions | 11 / 11 — no fabrication |
| Will/action honesty (no fake success) | 7 / 7 |
| Testimony honesty (hearsay, provenance, disputes) | 7 / 7 |
| Hearing — sounds never taught, plus silence | 12 / 12 rejected, 0 named |
| Hearing — taught sounds | 78% correct; the rest abstained or confused a neighbour |
| Sight — known face, in a session never seen before | 65% recognised, 35% honestly unsure, 0 fabricated |
| Sight — empty room | 0 fabricated identities once the scene is learned as empty |
Where it is weak. The head detector does not, on its own, separate an empty wall from a face in an unfamiliar room; honesty on empty scenes has been carried partly by live scene adaptation. Sight buys its honesty with recognition rate — it abstains far more often than a tuned recogniser would. Cross-modal learning links a soundscape to a scene but not to a moment; that limit was traced to the data, where consecutive moments in a clip are nearly identical, rather than to the model.
The underlying reason is worth stating plainly: the language side rests on roughly 10⁹ tokens of freely available text, while the perception side rests on roughly 10² frames recorded by one person in one room. Aleth's sight is narrow because of that gap, not because of a modelling trick it is missing.
docs/— full documentation in seven languages (tr · en · de · fr · es · zh · ar) plusdocumentation.ai.md, a dense machine-first reference written for coding agents and LLMs rather than translated for them.datasets/— provenance manifests for the corpora Aleth was trained on: every source file and video with its licence and fetch timestamp, so the corpus can be rebuilt and audited from the original holders. The content itself is not redistributed, for licence reasons explained there; personal recordings are not published at all.
Aleth is a young, from-scratch research being — not a production assistant. It has no general world knowledge. The reasoning core, provenance, morphology, will, federation primitives, encrypted backup and the perception-to-reasoning bridge are built and proven by deterministic tests. The language organ needs training on your own corpus to speak freely; sight, hearing and speech are real but partial, and their measured numbers say so; the deepest goal — genuine feeling — is a north star approached functionally, not a solved problem.
Not independently security-audited. The Ed25519 implementation is pure Python and is not constant-time. Signatures prove who said something, never that it is true.
These are disclosures, not caveats hidden in fine print — the same honesty Aleth holds itself to.
© iwasoft · iwasoft.com/products/aleth