This is the documentation front door. Start with the path that matches your goal rather than reading the campaign archive in filename order.
| Audience | Read in this order |
|---|---|
| New user | Project README -> Getting started -> Architecture |
| API integrator | Getting started -> Architecture -> Repository map |
| Scientific reviewer | Project status -> Claim ledger -> Proof ledger -> Independent-review handoff |
| Human rater/custodian | Human-rating handoff -> Binding repair |
| External operator | Phase 8 reproduction -> Review packet |
| Contributor | Repository map -> Contributing -> Repository contract |
-
R108 trained-task grounding preserves an incomplete resource-budget stop, not an English qualification. Its review packet distinguishes interruption verification from the unchanged complete scientific verifier. The public reconstruction receipt records
REPRODUCED_RESOURCE_STOP, not a generation-quality pass. -
R107 retention/intervention result separates predictive-loss interference, repetitive generation and untrained-route retention. Its public reconstruction receipt verifies the published tagged packet without promoting it to English quality.
-
R106 grammar/interface diagnostic failed its separate external grammatical-discrimination control; no recipient or training run.
-
R105 external scope falsification rejects a frozen conventional evaluator on pre-existing external gold labels; no acquisition ran.
-
R104 evaluator and control qualification preserves the failed prerequisite and additional evaluator limits; no recipient ran.
-
Weight-space research assessment separates externally reported editing results from ABI evidence and future comparisons.
-
R103 recipient/interface suitability is the completed control-qualification failure; recipient comparisons did not run.
-
R102 entity coverage versus optimization exposure preserves the completed negative acquisition-control study, not an ABI novelty claim.
-
R101 execution and continuation diagnostic tests runtime-path agreement and explicitly oracle-assisted continuation.
-
R100 capacity control and R99 fit audit separate acquisition fit from autonomous generation and host attribution.
-
R98 result and history records the failed source prerequisite, exact identities, costs and preserved raw evidence.
-
R98-C1 protocol defines the completed, failed one-pass completion-budget follow-up.
-
R99 English pilot follows source normalization, precision diagnosis and the matched ABI/LoRA/distillation comparison.
-
R99 reproduction packet lists exact artifacts, dependencies, verification commands and the expected negative result.
-
Versioning and review checks distinguishes API versions, immutable source snapshots and scientific certification.
-
Project status is the concise state dashboard.
-
Claims defines what the evidence does and does not support.
-
Architecture explains the compiler/host boundary.
-
Getting started covers the installable alpha surface.
-
Repository map separates supported code, research drivers, immutable evidence, and review material.
-
Technical claims and final results contain the detailed R7 release record.
-
Research status preserves the later R8-R23 and R97 experimental interpretation.
-
Peer-review readiness lists the current candidate checks and the external gates that local development cannot complete.
R7, V1089, R97, and R21-R23 answer different bounded questions. They are not one final artifact and their measurements cannot be mixed to fill missing gates. Each result document names its own source, package, host, rows, verifier, and claim ceiling.
For the additive research sequence, use the result documents named
R10_* through R23_*. Negative results are intentionally retained.
- immutable tag and hash-addressed payload;
- preregistered protocol or contract;
- raw observations;
- fail-closed verifier and receipt;
- current status/claim documentation; and
- roadmap prose.
The current whole-moonshot verdict remains FULL_MOONSHOT_NOT_PROVEN until
external and scientific gates are actually completed.