01
From idea count to the next research action
Aletheia 7.0 centres observation capability: when rival explanations predict the same responses to available actions, how can a researcher distinguish them? It aims to identify that gap, propose a bounded-cost operation with verifiable semantics and let observed results redirect exploration. Progress on unseen real tasks is the scientific goal; 7.0 first makes the engineering interfaces inspectable.
02
Inherit v6.8 while keeping distinct identities
A separate Aletheia module is built beside the existing project, reusing budgets, candidate records and finite experiment interfaces. An explicit v6.8 entry remains; its events and reports retain their original format. New records have distinct event identities, preserving which implementation produced each run and separating inherited engineering experience from the new research goal.
03
Primary text, supplements and missing-source states
Inputs comprise a research direction, source manifest, budget and optional host adapter. EvidenceStore indexes local PDF, HTML, Markdown and text, distinguishing body, abstract and supplementary material. Spans retain versions, page or section locators and hashes; missing, unreadable and partially extracted sources remain explicit. Indexing establishes source management, while claim support still requires reading.
04
Separate questions, ideas and predictions
Question holds the valuable question and at least two Rival records. IdeaCard stores a mechanism, predicted difference, strongest precedents and parents. Action specifies its type, definition, per-rival predictions, cost and admission state; Outcome holds the host response. Predictions and observations are separate objects, preventing self-evaluation from registering a confirmed discovery.
05
How one bounded round runs
The controller packs bounded evidence and requests a structured proposal. After object, source and parent checks, policy selects an action from public predictions, aliases and costs. Reading executes locally; a probe requires host attestation before observation. Results, rejection reasons and budgets inform review and the next context. This is primarily a proposal chain; the persistent multi-branch frontier arrives in 7.1.
06
The host verifies executable semantics
An alias witness concerns indistinguishability relative to the current rivals and action set. A proposed probe needs an operation, differing predictions and a cost. The host, rather than the proposing model, implements ActionAdapter. A finite-feature adapter can check a declared response table; without a real-domain adapter, experiments remain proposals or human-review items. Contract validity, observed response and scientific validity are separate judgements.
07
Fusion needs a load-bearing relationship
Fusion proposals record two parents, their relationship, an interaction prediction, component-only nulls and parent-removal ablations. They explain a response that the components alone may not produce. Missing fields or failed contracts remain recorded. Structural admission filters some invalid combinations; substantive interaction still requires host experiments, strong controls and independent review.
08
Outputs and resume retain version boundaries
JSON, Markdown and HTML reports retain proposals, choices, sources, observations and costs. Content-addressed objects and a local HMAC chain check record continuity. Resume occurs at complete-round boundaries and checks sources, runtime code, report projections and remaining time. Runs with executed probes reject resume because no general adapter-restoration protocol exists; literature runs can resume under the contract.
09
What the synthetic demonstrations establish
The historical report records finite-probe demonstrations, revision after a literature conflict and frozen policy-weight evaluation. Meta evaluation searches five preset weights. Test action costs match, but construction and development costs are not fully matched to the default baseline. The release reports 2,056 passes, 41 skips and no failures, plus packaging checks. These are historical reported counts; this website review reruns no research experiments.
10
Engineering foundation and scientific goal
At this historical point, there are no paid live-model calls, real scientific-domain probes or external independent reviews. Finite synthetic hosts check code paths rather than open-domain action invention, originality, real transfer or recursive self-improvement. The next contract first checks improvement headroom, then freezes real tasks, models, action language and full costs against active design, same-grammar search and self-editing baselines.
Reading example · Explaining the workflow
Illustration: two rivals fit the old measurement
Two rivals, matching predictions for an existing action, an anchored text and a finite host.
- 01
Identify the pair that existing actions cannot separate.
- 02
Propose differing responses and request host attestation of a new operation.
- 03
The host returns an observed label, which the next proposal must cite.
- 04
Record revisions, rejections and costs before continuing or stopping.
A follow-up proposal linked to an observation, or an explicit unverified or abstained outcome.
A conceptual walkthrough, adding no experimental or discovery claim.
Changes in this version
Exploration entry
- Earlier approach
A direction-only v6.8 input can remain ideas-only.
- This version’s approach
v7.0 connects sources, rivals, actions and outcomes in a bounded loop.
Make the next research action inspectable.
Evidence identities
- Earlier approach
Inherited finite-world results use their original reports.
- This version’s approach
New events and runtime identities remain distinct; the old entry is retained.
Interpret historical evidence under the implementation that produced it.
Fusion admission
- Earlier approach
A combination may remain prose-level.
- This version’s approach
Require relationships, interaction predictions, nulls and parent-removal tests.
Give structural claims concrete tests.