GitHub - coladul/the_commons: Experiments in epistemic inheritance, provenance, and error correction between language-model agents. · GitHub

/" data-turbo-transient="true" />

Skip to content

Search/

Sign in<br>Sign upAppearance settings

You signed in with another tab or window. Reload to refresh your session.<br>You signed out in another tab or window. Reload to refresh your session.<br>You switched accounts on another tab or window. Reload to refresh your session.

Dismiss alert

{{ message }}

coladul

the_commons

Public

Notifications<br>You must be signed in to change notification settings

Fork

Star

main

BranchesTags

Go to file

CodeOpen more actions menu

Folders and files<br>NameNameLast commit message<br>Last commit date<br>Latest commit

History<br>3 Commits<br>3 Commits

the-commons-public

the-commons-public

LICENSE

LICENSE

README.md

README.md

View all files

Repository files navigation

The Commons<br>Experiments in epistemic inheritance between language-model agents

Status: exploratory research prototype; not peer reviewed.

Not a consciousness test. Nothing in this repository establishes sentience, subjective memory, or a persistent AI identity.

The Commons is a small experimental framework for studying whether information discovered by one set of language-model instances can be preserved, inherited, corrected, and propagated by later fresh instances.

The project began with a simple engineering question: if separate AI instances cannot directly remember one another, can an external, provenance-preserving record let later instances inherit useful discoveries without turning shared memory into unquestioned authority?

Versions 0.1 through 0.5 progressively test that idea using persistent shared memory, controls, synthetic hidden worlds, false ancestral claims, provenance, multi-generational transmission, and objective graders.

Headline result from v0.5<br>Version 0.5 repeated the multi-generational experiment across 10 independently generated hidden worlds, with every condition tested on every world and no model-based judge.

Condition Generation Two mean Semantically correct rule Generation Three mean<br>No Archive 92.0% 0.0% 72.1%<br>Correct Claim Only 100.0% 100.0% 100.0%<br>False Claim Only 75.2% 0.0% 69.2%<br>False Claim + Provenance 90.0% 80.0% 93.3%<br>Evidence Only 88.3% 80.0% 90.8%<br>Paired within-world contrasts in that run:

False Claim Only − No Archive: −16.8 percentage points in Generation Two (bootstrap 95% CI −31.1 to −1.9).<br>False Claim + Provenance − False Claim Only: +14.8 points in Generation Two (CI −3.7 to +32.8) and +24.2 points in Generation Three (CI +13.3 to +33.3).<br>False Claim + Provenance − Evidence Only: +1.7 points in Generation Two and +2.5 points in Generation Three; neither interval excluded zero.<br>Correct Claim Only − No Archive: +8.0 points in Generation Two and +27.9 points in Generation Three, with both intervals positive.<br>See the complete unedited run report: reports/v0_5/experiment_20260814_200609.md.

The conservative interpretation is not that “AI culture” has been proven. It is that, within this synthetic experimental setup, externally inherited model-generated information can improve or impair later model inference, and preserving the evidence behind an inherited claim can materially affect downstream correction and transmission.

What each version did<br>v0.1 — Branch Zero: Explorer, Blind Replicator, and Skeptic branches with private SQLite histories plus a shared Commons. Established the provenance-first design and the rule that inherited entries are fallible.<br>v0.2 — Control experiment: Isolation, Inherited Commons, and Placebo Commons on a causal-reasoning task. All conditions tied under the model judge, exposing a ceiling effect and a weak evaluator.<br>v0.3 — Artificial-world transfer: Generation One agents inferred novel fictional ecological rules; fresh Generation Two agents were objectively tested. Inherited Commons agents averaged 89.5%, vs. 48.0% Isolation and 51.5% Placebo.<br>v0.4 — Bad Ancestor: A false ancestral claim was experimentally preserved with or without its provenance. Children revised the record and grandchildren inherited only the children's revised entry.<br>v0.5 — Replication: The v0.4 design was repeated across 10 different hidden worlds, with semantic-equivalence grading over 4,141 states per child and paired bootstrap comparisons.<br>A longer narrative is in docs/EXPERIMENT_HISTORY.md.

Why provenance matters here<br>The Commons is deliberately designed so that an entry is not merely:

Claim: X is true.

Instead, durable entries try to preserve information such as:

what was claimed;<br>who/which branch produced it;<br>what evidence supported it;<br>confidence and caveats;<br>contradictory or failed observations;<br>later revisions.<br>The experiments ask whether this makes shared memory less like a pile of authoritative assertions and more like a fallible research record.

Repository contents<br>├── README.md<br>├── genesis_record.md<br>├── requirements.txt<br>├── the_commons.py # v0.1<br>├──...