Aletheia

What the project explores

Aletheia advances a research direction into questions, mechanisms and discriminating next actions. It retains the strongest prior, actual sources and unresolved conditions so a researcher can inspect why a step was chosen and what evidence is missing.

From primary text to a proposal

Local body text, abstracts and supplements enter through a source manifest. The first round retrieves passages by a scientific query; later rounds prioritize strongest priors, conflicts, acquired passages and recent observations. Proposals cite exact spans or acquired receipts, while missing sources and context omissions remain visible.

Six-layer core architecture

Sources, proposals, control, execution, feedback and persistent records have explicit boundaries. The model proposes candidates while the host controls semantics, execution, actual results and budgets; typed contracts connect them.

1. 01

   Source layer

   Build a source store from an operator-supplied local manifest. Body, abstract and supplementary text retain distinct roles, page/section locators, versions, hashes, availability and query windows.

   EvidenceStore / EvidenceSpan

2. 02

   Proposal layer

   The model reads a bounded evidence packet and returns a question, rivals, an idea and candidate actions. A strict structure contract bounds fields and size; predicted outcomes cannot become observed results.

   Provider / PROPOSAL_SCHEMA / Question / Rival / IdeaCard / Action

3. 03

   Controller layer

   Micro policy ranks declared response separation and action cost. Macro policy retains branches and pending actions, selecting affordable dependency-ready obligations and rotating stale untested branches.

   AliasReport / PolicySpec / FrontierGraph

4. 04

   Host execution layer

   Selected actions pass identity, source, semantics and cost gates. Literature actions execute locally; experimental probes observe only after host domain-adapter attestation. Without an adapter, experiments remain proposals or human checkpoints.

   ActionAdapter / AdapterAttestation / Outcome

5. 05

   Feedback layer

   Observed responses conditionally update rivals; anomalies, noise and uncertified mappings remain unresolved. Revisions explain introduced, revived and removed rivals. Review separates discriminating question observations from independently validated idea mechanisms.

   ObservationAnalysis / QuestionRevision / IdeaReview

6. 06

   Persistent record layer

   Proposals, selections, rejections, budgets and outcomes enter content-addressed objects and a host HMAC chain. Complete rounds save state checkpoints, and reports provide readable projections. Source and runtime identities stay locked across resume.

   EventJournal / content-addressed objects / ROUND_CHECKPOINT / report.json / REPORT.md / REPORT.html

Research-software architecture outline, not an experimental result.

Micro and macro loops

The micro loop links response predictions, action attestation, observations and conditional rival updates. The macro loop retains idea lineage, evidence and deferred actions, then selects the next focus from unresolved obligations. Old branches can return to context; pending actions preserve their specifications and pass the contract again.

Make actions and costs inspectable

Reading, prior challenges, probes, human review and stopping have distinct action types. Model calls, retrieval, execution, rounds and context are bounded; reservations precede calls and failure or overage costs remain recorded. New probes require executable semantics attested by the host.

Exploration records and evidence levels

JSON, Markdown and HTML reports include proposals, reviews, observed responses, frontier and budgets. An HMAC journal checks local record continuity. A discriminating observation about question rivals and independent validation of an idea mechanism are separate conclusions; evidence acquisition has its own receipt type.

Progress recorded in v7.3

Authorized literature development used six live-model calls and three local acquisitions, producing an untested proposal on exploration stagnation. In the eight-state diagnostic, the conditional ledger, flat presentation of the same information and ordinary active design passed 8/8; fixed actions and disconnected feedback passed 0/8. Feedback-driven follow-up decisions are supported, while separate harness gains remain unestablished.

Four polynomial-response states and four graph-search-trace states. Each world evaluates one follow-up decision; the formal batch used 40 model calls.

Literature proposal status · PROPOSED_UNTESTED

Four engineering steps within v7

7.0 establishes evidence and action foundations; 7.1 adds conditional updates and a persistent frontier; 7.2 connects live Codex proposals and paired development pilots; 7.3 records the literature loop and frozen diagnostic, using strong-baseline results to shape the next architecture study.

Next stage: spend model computation on explanation and action gaps

The next design would compile admitted actions directly and separate stable question specifications from changing state, then construct new measurement language when existing actions cannot distinguish rivals. Selective repair and stagewise generation/retrieval/selection interventions remain prospective contracts.