RP002A · CLOSED
When history helps prediction
In finite partially observed stochastic processes, when does recursively updated history recover predictive distinctions unavailable in current observation?
Autorite / Research Program
Begin with questions.
Follow the evidence.
An open research program on continuation, representation, reachability and realized history. The first Cycle investigates the testable boundaries of these questions.
Questions before architecture.
Autorite keeps the scope of claims, their assumptions and the evidence that could change them visible. Negative results are part of the research record. The program makes no claim to a completed theory.
The audited baseline is frozen. RP002A completed its separately authorized second confirmation: 60 units and 120 learned fits, with seven conditional decisions under the frozen analysis. The first incomplete attempt remains preserved. RP002A is CLOSED after its scoped internal review; Cycle 01 remains ACTIVE. Read the package review.
Cycle 01 research
RP003 · PLANNED
From local to global preservation
When do local continuation and preservation judgments support a global judgment?
RP004 · PLANNED
Branching lineages
Which distinctions are required for continuation claims under copying, branching and merging?
RP005A · PLANNED
Constraints and viable futures
Can targeted restrictions steer horizon-limited decisions away from states with fewer viable future targets?
Recent changes
2026-10-10
RP002A package reviewed and completed
D-0004 accepts the scoped finite commitment after an explicit internal review. All seven frozen outcomes remain unchanged, including the indeterminate recurrent comparison. RP002A is CLOSED; Cycle 01 stays ACTIVE with 1/13 obligations resolved. Q-3 and Q-4 remain ACTIVE. Next: RP003 protocol preparation; no new execution.
2026-10-10
Second confirmation completed and independently replayed
The freshly authorized technical retry completed all 60 units and 120 fits. Seven frozen conditional decisions: two benefit, four equivalent, one indeterminate. 484 hashes and saved-checkpoint/analysis replay passed; no trace warnings. First incomplete attempt remains preserved. No broad Claim or Cycle closure.
2026-10-10
First confirmation attempt interrupted; partial artifacts preserved
A wall-time/host-suspension clock incident stopped the one attempt. 36/60 complete units and 73 saved fit traces are preserved and audited. No aggregate decisions or VALID Result for that attempt. A separate retry was later freshly authorized. The timer guard is repaired.
2026-10-09
Confirmatory count and analysis frozen before outcomes
The calibrated eligible counts and fresh SD-only selection chose 20 replicates per world. Models, training budget, seven intervals, .01-nat margins and one four-hour attempt were fixed before confirmation.
2026-10-09
Fresh target-size variance stage completed
Five independent replicates per world and 30 learned fits completed and replayed. Only their SDs informed the frozen count rule; means remain descriptive.
2026-10-09
Fixed synthetic statistical calibration completed
All 90 fixed cells and 1.8 million trials completed. Counts 20,30,50 qualified under the declared simulation gate; simulated coverage does not guarantee actual-loss coverage.
2026-10-09
Uncertainty design and replicate-count sensitivity reviewed
An exact resampling-variance audit identifies an added within-episode term. Bootstrap tail resolution, assumed precision and runtime sensitivity are reproducible; no new samples, fits or final statistical design.
2026-10-09
Target-data-size feasibility pilot completed and audited
Three units and six fits completed in 5 minutes 48 seconds, with about 251 MiB lifetime peak process memory and no declared warning. All 50 outputs replay; n=1 per world supports no variance or confirmatory decision.
2026-10-09
Target-data-size pilot execution preparation verified
The separately gated six-fit pilot and no-variance artifact audit are implemented. Fifty-seven software tests pass; prior outputs replay. No new research training or pilot approval.
2026-10-09
Remaining budget warnings reviewed; target-size pilot proposed
A replayable validation-only diagnostic leaves all 12 prior warnings unresolved. A fresh six-fit feasibility pilot at the intended data sizes is proposed; no new observations or training.
2026-10-09
Longer-budget exploratory stage completed and audited
The separately approved 40 fits completed in about 6 minutes 24 seconds. All 322 outputs and saved predictions replay. Twelve near-cap warnings remain; no scientific Claims or Cycle resolutions.
2026-10-09
Stage-002 execution preparation verified
The longer-budget entry point and artifact audit are ready. Fifty software tests pass and previous outputs replay. The separate one-run approval gate remains closed; no new research training.
2026-10-09
Longer-budget convergence proposal prepared
A separate 300-epoch, 20-epoch-patience plan keeps models and data sizes fixed with 140 fresh seeds. Proposal checks pass; no new training or confirmatory evidence.
2026-10-09
Exploratory convergence stage completed and audited
All 20 independent units and 40 model fits completed in about 83 seconds. Artifact replay passed. Many validation traces improved near the epoch cap, so settled optimization is not established. No confirmatory decisions or Cycle resolutions.
2026-10-08
Separate convergence and replicate-variability stage prepared
A bounded exploratory plan and executable authorization gate are ready for review. Forty learned-model fits are proposed across 20 independent units; no new run has been executed.
2026-10-08
Literature overlap and confirmatory design reviewed
The finite-history benchmark overlap is documented. Deterministic design calculations show that the ideal window-eight gap is very small; five training replicates lack a power justification. No new model runs, scientific Claim promotions or Cycle resolutions.
2026-10-08
Baseline frozen; first engineering pilot completed
The audited bootstrap and bounded pilot were approved. Baseline v0.1 is tagged; the RP002A pilot completed with finite outputs and verified artifacts. No scientific claim or Cycle obligation was resolved.
2026-10-08
Executable RP002A harness and isolated pilot proposal
Generators, reference filtering, model causality, episode resets and seed schedules pass software checks. Thirty tests pass; no research pilot or confirmatory run has been executed.
2026-10-08
Validated research core and proposed experiment configuration
Typed research records and the 13-obligation Cycle manifest now validate before site generation. RP002A has a numerical configuration proposal; it remains unfrozen and no experiments have run.
2026-10-08
Canonical GitHub repository and RP002A design
The source repository is omasim/Autorite. Baseline source bytes are preserved and checked; a proposed operational protocol is available. No experiments have been run and release approval remains pending.
2026-10-08
RP002A prior-art investigation started
The first literature review and research-start workflow are documented. Experimental protocol choices and GitHub setup remain open; no experiments have been run.
2026-10-07
Both English sites are live
The knowledge site and open laboratory are publicly available at autorite.org and autorite.net, with verified HTTPS.
2026-10-07
Publish the process before running research
The knowledge and laboratory sites are being prepared before research execution. Both sites will be in English.
2026-10-07
Implementation contracts completed
Research records, Cycle obligations, closure and pre-execution readiness rules were specified.
2026-10-07
The updated baseline selected
The updated publishing architecture takes precedence. The older package remains historical source material.