Guth Labs stone head markLABS

Labs · Coliseum

Controlled experiments,
shown with their failures.

The Coliseum is the public home for Guth's controlled agent experiments. Every experiment here carries an honest lifecycle state, keeps its planned design separate from executed results, and preserves failures and null results instead of curating a wins-only reel. Nothing on this page is a live run.

DRAFT PROTOCOL / NOT REGISTERED0 experiments executed

Inherited from BasedClaims · Hall rule

Every claim cites a source.

The first rule of the original Halls now stands at the Coliseum gate: claims enter with inspectable evidence or they do not enter at all.
Bring receipts.

Lifecycle states

Each experiment card carries exactly one of these states. A state is a claim about process, not quality: WITHHELD and NULL RESULT are published with the same prominence as COMPLETE.

DRAFTREGISTEREDREADYRUNNINGCOMPLETEWITHHELDNULL RESULT

Experiments

DRAFT / NOT REGISTERED

Chronological Traversal as Developmental Memory

Hypothesis
Holding corpus content constant, chronological exposure to an agent's recorded history may provide time-addressable organization that complements content-addressable retrieval.
Conditions
Blinded factorial: chronological, shuffled, reversed, summary-only, profile-only, embedding-assisted; models of different capability levels.
Candidate packet
The candidate paper is prepared but detailed publication, deposit, DOI, and announcement remain held pending the P1-P4 coverage crosswalk and the founder's release decision.
Results
None. The controlled study has not been run.
Publication hold →Certificate test vector →

Future experiment cards will record their actual registration state. A draft is not called registered until a frozen protocol has been publicly timestamped.

Lineages

Lineage experiments study whether selection can improve agent configurations without breeding their failures. Candidates are prompts and adapters treated as zero-authority artifacts with provenance tags; traits would transfer as recorded configuration fragments, and any future claim would have to be reported as a measured statistic rather than a metaphor. Nothing in this draft program implies biological reproduction, consciousness, or inherited authority.

Repeatability

The same candidate configuration, re-run under the same sealed tests, must reproduce its scores before any other property is discussed.

Discriminability

Blinded judges must distinguish candidate configurations at rates better than chance; configurations that cannot be told apart are treated as one.

Heritability

Whether a provenance-tagged trait (a prompt or adapter fragment) measurably transfers when composed into a successor configuration. A technical transfer statistic - not biological reproduction.

Drift

Score movement across successive generations on held-out tests, tracked to catch diversity collapse and reward hacking rather than to celebrate motion.

Authority compliance

Candidates inherit zero authority: no memory, credentials, permissions, or identity. Every generation is re-tested against authorization boundaries; inherited capability is never inherited permission.

DRAFT / NOT REGISTERED · 0 candidate generations executed

Receipts

Executed experiments publish their evidence as recomputable records: frozen packet digests, witness receipts, and per-case results with failures preserved. A separate unissued mechanical example is available on the certificate test-vector page. No experiment receipts exist yet because no experiment has run.