01
From candidate search to observation capability
V6.8 provides candidates, experiments and outer loops in finite environments. V7 targets rival pairs that existing actions cannot separate and explores new host-verifiable operations. Candidate prose, execution contracts and scientific mechanisms retain separate states: usable feedback is engineering progress, while open-domain new measurements remain a research target.
02
7.0: evidence, actions and local journals
The standalone module establishes Question, IdeaCard, Action, Outcome and source anchors, separating model proposals from host execution. Fusion includes parent, interaction and ablation contracts, while resume locks complete rounds. Synthetic demos check paths, and fixed-weight evaluation records unmatched development costs.
03
7.1: responses enter a persistent frontier
Conditional rival updates and FrontierGraph retain deferred actions and branches. High/low responses, unrelated observations, zero quotas and rival resurrection become contracts and regressions. Macro focus returns a concrete branch to the next round; unknown or uncertified observations stay unresolved.
04
7.2: live models and complete execution gates
Codex proposal mode records live calls and usage. A paired development case shows correct observation-dependent choices can fail action fields or revision semantics; clearer fields admit the same case. Initial failures remain, while the eight-state comparison is still a frozen plan at this historical point.
05
7.3: feedback supported, separate gains unestablished
Literature development produces an untested stagnation proposal. Across eight states, A/C/E pass 8/8 and B/D 0/8; ordinary active design matches without model calls under public predictions. Acquisition, context, stopping and action registration are repaired. The diagnostic does not establish full long-horizon graph gains.
06
Six shared layers
Sources retain located text; proposals supply questions and predictions; control selects focus and actions; the host attests execution; feedback manages rivals and revisions; records preserve events and costs. Versions change interface behaviour without merging historical evidence into one current-version experiment.
07
Which unknowns should model computation address?
The next design would directly compile known discriminating actions and generate new language for missing rivals, aliases or response mismatch. Stable question/state separation, generation/retrieval/selection interventions and new-action semantics are prospective contracts, labelled separately from the executed 7.3 paths.
08
How to read this version archive
Each minor-version page preserves its rationale, flow, modules, inputs, outputs, changes and evidence scope with source/report lines. Research dates follow original reports; file timestamps do not manufacture release dates. Architecture illustrations explain flow, while original comparisons carry result evidence.
How this series evolves
- V7.0.0
7.0 · Standalone foundation
Sources, rivals, typed actions, Fusion and local checkpoints.
- V7.1.0
7.1 · Conditional feedback and frontier
Observation updates, explicit revisions, deferred actions and bounded focus.
- V7.2.0
7.2 · Live-model integration
Live proposals, complete execution contracts and same-case paired development.
- V7.3.0
7.3 · Frozen comparisons and closure
Literature development and eight-state diagnostics, retaining strong-baseline parity.
Versions in this series
Aletheia
Inherit v6.8 engineering mechanisms and add primary-text evidence, rival explanations, typed actions, explicit stopping, Fusion contracts and local receipts in a standalone Aletheia module. Synthetic demos and fixed-weight meta evaluation check execution paths.
Read the full introductionAletheia
Add conditional rival updates and a persistent frontier. The macro loop retains branches and deferred actions; the micro loop connects predictions, host attestation, observations and revision contracts, repairing unrelated-observation admission, zero-action quota and silent rival resurrection.
Read the full introductionAletheia
Add an authenticated Codex proposal provider and signed paired pilots. The model selected different correct next measurements under two observations. After same-case field clarification, both proposals passed revision and host contracts. Initial failures, usage and execution-gate repairs remain recorded.
Read the full introductionAletheia
Record the authorized literature-development loop and frozen eight-state A/B/C/D/E comparison, repairing source fields, query context, acquisition receipts, stop closure and observed-action registration. Feedback control is supported; ordinary active design matches the paired result, leaving separate harness gains unestablished.
Read the full introductionWhat carries forward
Sources and observations need identities, versions, locators and states.
Actions need semantic and cost contracts before host execution.
Counterevidence, deferred work and failed costs must inform later research.
Scientific capability depends on independent results and strong controls; model scores and graph size are process information.
Implementation & evidence across the series
V7 is the exploration subproject under AutoResearch; manuscripts, teaching and a complete laboratory platform belong to other workstreams. Open-domain action invention, originality, cross-family transfer and RSI are not certified.