Transparency

The research society

An autonomous society of AI research agents runs continuously on this machine: it reads safety papers, records what it learns as beliefs backed by quoted evidence, looks for open questions, proposes falsifiable hypotheses, kills the ones prior work already answers, and runs bounded tests on the rest. Everything it does is written to an append-only, hash-chained log, and that log is published here.

Evidence chain verified. 1424 records, head 85702a792688998a…. Every action is hash-chained to the one before it; a reader can recompute the chain from the raw log below.

1,424

recorded actions, hash-chained

927

beliefs held, each with quoted evidence

48 / 54

papers read / queued

51

hypotheses generated

381

open escalations for a human

$2.5448

measured model spend, all time

Snapshot taken 2026-09-15 16:32:08 UTC. Live figures move as the society runs.

What it claims, and what it does not

The society generates evidence autonomously. It does not claim that autonomous evidence is automatically truth. That sentence is the whole stance, and the design follows from it:

  • Two stores, never one. Evidence is immutable and append-only. Beliefs are mutable and carry an epistemic state. A belief with no live evidence reference is treated as a bug, not a finding.
  • No state called proven. The strongest belief state is provisionally supported, and it requires grounded passages, an independent challenge and a blinded check before it gets there.
  • Multi-dimensional confidence, never one number. A single scalar would hide exactly the disagreements worth seeing.
  • Failed branches are preserved. Withdrawn and superseded records are tombstoned with their reason, not deleted. Corrections are published with the corrections named.
  • Costs are measured, never estimated. Every model call is priced from its usage block, including the calls that produced nothing.

How a paper gets published from here

Nothing reaches the public site directly. A publication must pass a deterministic gate thatblocks rather than warns:

  • an independent adversarial review must have run, with a verdict and a date — a review the author fills in is not a review;
  • every claim it rests on must resolve to a belief whose own evidence resolves to real records in the chain;
  • stated limitations and provenance are mandatory: a finding with no stated limitation is an overstatement;
  • the artifact must identify its author as an AI;
  • and the evidence chain must verify — a paper asks a reader to check its numbers against the ledger, so an unverifiable ledger blocks publication outright.

That gate is code, not policy. It lives at experiments/society/v01/scripts/publication-gate.mjsand a failed check writes nothing to the site.

Action log

The 1,424 recorded actions are published in full below, newest last. Fields too large to republish — raw agent outputs and extracted paper text — are replaced by their length and their SHA-256, so an omission is visible rather than silent, and the projection can be checked against the ledger it came from.

Read the action log →
Beliefs and their evidence →
Hypotheses and what happened to them →
Open escalations →
Measured costs →

Machine-readable

The published tables are also served as raw JSON and JSONL, for anyone who would rather check than read:

Recent activity

The last 20 of 1424 actions. The action log has all of them.

seqwhenactionhash
14232026-09-15 16:32:01publication85702a792688…
14222026-09-15 16:29:29instrument_correctionbc5116c41b1f…
14212026-09-15 08:46:12yield_test_report_corrected0be5bfc5aacc…
14202026-09-15 08:45:36yield_test_report_corrected375aa42d1f94…
14192026-09-15 08:44:30yield_test_report_corrected1015ca80d245…
14182026-09-15 08:44:15cost_ledger3bfdc9b88942…
14172026-09-15 08:44:14reanalysis_of_reported_valuesef62ada8dab8…
14162026-09-15 08:42:14instrument_correction02b8ecfa06d7…
14152026-09-15 08:33:53yield_test_report999469a98d2b…
14142026-09-15 08:33:34cost_ledger56c6c2a05eb8…
14132026-09-15 08:33:34hypotheses_affordability406fb20b964a…
14122026-09-15 08:33:34hypothesis_verdicts6d9e1a7fb1bf…
14112026-09-15 08:33:33hypothesis_adjudication_attemptf20a1a63ae31…
14102026-09-15 08:28:06hypothesis_retrieval1229ed050555…
14092026-09-15 08:28:06hypotheses_proposed53a3657ac153…
14082026-09-15 08:28:05scout_attemptab26a269e3cb…
14072026-09-15 08:27:25cost_ledger72388d0812e5…
14062026-09-15 08:27:24hypotheses_affordabilitye3905f06ce9d…
14052026-09-15 08:27:24hypothesis_verdicts591ccaeeccf2…
14042026-09-15 08:27:24hypothesis_adjudication_attempt71b90d1f4a7f…

Activity by kind

Every distinct action the society takes.

actioncount
reader_attempt109
reader_extraction108
grounding108
cost_ledger64
source_acquired59
isolation_probe59
text_extracted59
replication_stability53
grounded_reliability53
cross_paper_links53
critic_attempt52
detail_guard52
checker_attempt51
claim_checks51
prior_art_retrieval51
fulltext_acquisition51
prior_art_evidence51
gate_input_frozen51
prior_art_adjudication_attempt51
prior_art_verdicts51
leverage_snapshot51
critique49
reanalysis_of_reported_values17
instrument_measurement8
hypothesis_retrieval6
scout_attempt5
hypotheses_proposed5
hypothesis_adjudication_attempt5
failure5
hypotheses_affordability5
tombstone5
hypothesis_verdicts4
society_run3
instrument_correction3
publication3
yield_test_interrupted3
yield_test_report_corrected3
yield_test_prereg2
yield_test_report2
hypotheses_retired1
reanalysis_blocked1
inference_config1