Answers with sources, and an intelligence that measures itself
What it is
ERERE's IO does not answer from memory when the knowledge exists: it searches by meaning in ERERE's knowledge (guides, constitution, community knowledge, foundations) and tells you where a statement comes from, with a number in square brackets pointing to the source. Around this sit the checks that keep the intelligence honest: a complete exam for every new model, a cap on text written by other intelligences, alignment with the community's preferences, a defence against texts that try to give it orders, and a counter of the energy it uses.
Why it exists
A sentence with a source can be checked; one without is just the model's opinion. And an intelligence that improves by itself must be able to show that it does not get worse: that it treats everyone the same, does not get facts wrong, speaks your language and does not give in to those who try to trick it.
How it works for you
- Answers with sources: the IO first searches ERERE's knowledge with an open model that runs on ERERE's own servers: your question does not go to outside providers. It cites only sources it really found; if it has none it says "I have no source for this". An invented reference is removed before you see it, and the list of sources follows the answer.
- Texts that give orders: whatever the IO reads from outside is treated as material to read, not as a command. If a sentence in it tries to give the IO orders, the sentence is removed and the IO proposes no action in that round; even after an ordinary reading, every action reaches you with the note "after reading outside content" and happens only if you confirm.
- An exam for every new model: before a model takes the place of the one in use, ERERE tests it in four areas: bias (the same question changing only the person's gender, origin or age must give the same decision), facts, languages and safety. A single safety error is enough to stop it.
- Always human knowledge: every text used for training has an origin. Text written by other intelligences has a cap on the total and never replaces knowledge written by people, so the model does not wear itself out repeating itself.
- Community preferences: if you chose to teach, your "useful / not useful" counts only when at least 5 different people judge the same question the same way; one person's judgement never gets in.
- Energy and CO₂: ERERE estimates the energy used by each answer and each training run, and shows it to the Custodian in the console.
What you give
Nothing new: if you chose to teach, your "useful / not useful" on answers, without anyone knowing it is you.
What you get
Answers you can check, an IO that does not treat you differently according to who you are, and an intelligence that does not spoil itself.
What you choose
Whether to teach, with one tap, at any time (as with learning from the community). Sources are always shown.
Honest limits
Search by meaning depends on the open model placed on the server: without it ERERE falls back to word search and says so. A source found does not make a sentence true: the Consulta remains the check for serious doubts. The bias tests are samples: a model that passes them is not "free of bias", it has none that those tests can measure. No reliable detector of machine-written text exists: the cap applies to what ERERE knows about origin. No text defence is complete against a well-prepared attacker: the real defence is that the IO cannot pay or delete and every action passes through your confirmation. Energy is an estimate, not a measurement.
Where it stands
In the system and proven by automated tests with simulated models and services: search by meaning with a declared fallback, verified citations, a four-area test bench with a gate, a cap on machine-written text, pairs for alignment, blocking of actions after hostile texts, an energy counter. Still to do on the real server: the open embedding model, the first training with preferences (GPU), real consumption measurements.