| 0001 |
Derive the world seed instead of authoring it on the scenario |
| 0002 |
Exclude unconditionable personas at the index, not after sampling |
| 0003 |
Personas react to impressions, not to individual stimuli |
| 0004 |
Briefs reference their category ontology by version rather than embedding it |
| 0005 |
World identity derives from the whole scenario, not the variant |
| 0006 |
The trace records whole turns in self-verifying partitions |
| 0007 |
Digest metrics are computed from the weights they carry |
| 0008 |
Calibration claims carry measured evidence above fixed floors |
| 0009 |
Identities are derived, and every pin is verified where its object is present |
| 0010 |
A persona's view is recorded with its turn, verified, and sees only earlier ticks |
| 0011 |
Worlds resume by replaying their trace, and the runner is the partition's only writer |
| 0012 |
Fallback models are pinned, and every model call records the route that served it |
| 0013 |
Loading a brief touches no network; evidence is fetched by a separate command into a sidecar |
| 0014 |
Category ontologies and audiences are drafted from the corpus and confirmed by a person |
| 0015 |
A population is built once and carried, never rebuilt |
| 0016 |
The engine is a research instrument, and its default persona corpus is research-only |
| 0017 |
A grounded field says whether it was measured or extracted |
| 0018 |
Calibrate against measured marginals rather than fusing one person from several |
| 0019 |
The social graph is generated, and what cannot be measured is swept rather than fixed |
| 0020 |
The pre-flight has two stages: what the corpus holds, then what the draw achieved |
| 0021 |
The engine speaks one OpenAI-compatible endpoint and imports no provider |
| 0022 |
Telemetry is exported as OpenTelemetry, carries no persona content by default, and is not the trace |
| 0023 |
Model calls are requested in batches, answered in request order, and fail as outcomes |
| 0024 |
A completed attitude is sampled by the engine from a distribution the model gives |
| 0025 |
A cached sample is never shared across replicates |
| 0026 |
Elicitation computes the published SSR formula, not a softmax over cosines |
| 0027 |
Anchor statements are frozen before they are validated, and never tuned on the data that judges them |
| 0028 |
Elicitation validates the text-to-rating mapping on human data, and records the simulation claim as owed |
| 0029 |
Anchors are a study input with recorded provenance, and the gate judges every one of them |
| 0030 |
A persona's state travels with its job, and the trace is the only store |
| 0031 |
The agent answers in batches, in request order |
| 0032 |
An unscorable reaction keeps its verbatim, and is scored when a version passes |
| 0033 |
A tick is recorded whole, or not at all |
| 0034 |
One trace view over two backends, and one implementation of every derived shape |
| 0035 |
Run state is derived from the trace, and a checkpoint is only ever a cache |
| 0036 |
A resume refuses a run whose inputs or engine moved |
| 0037 |
A budget rung belongs to the run, and a world replays the rung it ran under |
| 0038 |
A digest reports what a run measured, and says plainly what it could not |
| 0039 |
A digest is one world's, and a scenario carries the spread between its worlds |
| 0040 |
Anomalies are rules over what was measured, and report what they cannot measure |
| 0041 |
Objection clusters are deterministic, and labelled by a verbatim rather than a summary |
| 0042 |
A report is derived from the record, and carries what produced it |
| 0043 |
The interface is a single-operator local application |
| 0044 |
An ontology is an immutable versioned study input, authored against the codebook |
| 0045 |
The web layer derives nothing |
| 0046 |
A prompt is reconstructed and verified, never stored |
| 0047 |
A study's description drafts its audiences, never its category's conditioning set |
| 0048 |
Channels, survey waves and launch reach are scenario content |
| 0049 |
A population's categorical gates are judged as one family |
| 0050 |
A running study shows decisions before they are record |
| 0051 |
A version shows its personas its own description and price |
| 0052 |
A version is named outside its identity |
| 0053 |
Objections are chosen from what turns recorded, and grouped by average linkage |
| 0054 |
Finished-world analysis reuses a scoped trace snapshot |