Findings Engine
Did research on its own at night until 25 September 2026: collected findings across fields, interrogated them, spotted gaps. Now a read-only archive; a rebuild is planned.
The findings engine was the research half of the lab. A finding is a claim that has been recognised in more than one field, with the passages it stands on attached. The engine recorded findings, linked them to each other as supporting or contradicting, and was built to ask four questions of every one: where does it stand, does the passage really say that, when would it break, and has someone published this before.
It worked at night, without anyone watching, until 25 September 2026. That day the service underneath it was switched off as part of moving the lab’s memory to a new core. What it left behind is an archive: 524 findings, 416 of them readable here and 108 withheld because they draw on private material, with 268 supporting and 68 contradicting links between them. 201 findings never got a link at all.
The four questions were not asked as evenly as the design promised. The check against the published record reached 53 of the 524 findings; a breaking condition was written down for 174; the fit of the evidence was rated for 458. That unevenness is part of what the rebuild is meant to fix.
Right now nothing takes in new findings. The archive can be read; the engine that wrote it is not running. A rebuild on the new core is planned and not started. The numbers below are read from the archive of 23 September 2026, not from a running system.
What goes in
Snapshot 25 September 2026A finding is a claim with legs: the passages it stands on, in one or two of 93 domains. Findings come from reading (a session marks a passage as a finding), from the night run (the engine reads what is new and files what it considers a claim), and from re-provenance runs that give an old finding better legs. Most findings are bridges: a mechanism seen in one field, recognised in another.
Findings by domain
- agent orchestration73
- complex systems71
- active inference68
- information theory64
- category theory57
- organisation theory32
- financial regimes31
- llm optimization24
- depth psychology24
- network ecology23
- topology21
- mathematics18
- physics18
- control theory18
How well they stand
Adequacy is rated per finding: does the cited passage actually carry the claim? 159 of 524 pass. "No adequate evidence" is a verdict, not a gap: the passage was read and does not carry the claim. The engine does not hide the weak ones; it works on them.
Four questions asked of every finding
Snapshot 25 September 2026A finding is not accepted, it is interrogated. The engine keeps score of how far it has got with each question; the score is shown, not the flattering part of it.
1. Where does it stand?
Provenance: which passages carry the claim. 457 of 524 findings have it; 216 stand on a single leg, one passage in one domain, which is the first thing the night tries to fix.
2. Does the passage really say that?
Adequacy: a model re-reads the cited passage against the claim. 458 findings carry a current rating, 0 an old one, 66 none yet. 159 pass. Weak is the normal state of a new finding, not a verdict on it. A finding that fails the same check three nights running is set aside, so the night does not spend itself on the same dead end.
3. When would it break?
Scope: every new finding must name the condition under which it stops being true. 174 of 524 carry one, 33.2 %; the rule is a day old and the backlog is being worked through.
4. Has someone published this before?
Prior art: every new finding above a confidence of 0.7 is checked against an open scholarly index, with a second path where the model names a likely predecessor and the index verifies the title. Verdict: rediscovery, novelty, or unclear. 53 of 524 checked so far. The first result was a rediscovery: a finding about a precision parameter as a price turned out to have been published in 2013. That is the point of the question.
Curation by hand, on top of all four: 23 findings so far, and no automated producer for it yet. Between findings: 321 are linked to at least one other, 203 stand alone. A rule keeps the net honest: two weak findings are never allowed to support each other, because two unchecked claims make noise, not convergence. The bridge between a finding's two fields is still a label on the finding, not a walkable edge in the graph; that is a known limit.
The night, hour by hour
Timers as of 2 September 2026Nothing here waits for a person. Each organ has a cap, so a bad night cannot flood the graph, and each writes a report the morning digest reads. Times are UTC.
| when | what | how |
|---|---|---|
| 00:30 | adequacy check | re-reads the cited passages of up to 20 findings and re-rates them |
| 02:40 | gap-filling run | the research agent takes up to 4 open requests, reads and searches, files findings |
| 03:00 | prior-art gate | for up to 5 findings: has someone published this before? Checks a scholarly index |
| 04:45 | curator | a model judges up to 60 candidate pairs: supports, contradicts, or nothing |
| Sat 21:34 | research night | a free run of the agent over the week's new material, then a briefing |
| Sun 06:10 | harvest report | the three findings that pulled hardest this week, sent to Andreas |
Open requests: 26
The engine files its own questions. Four kinds, three of them minted by the engine itself; a conflict request appears automatically the moment the curator draws a contradiction.
| adequacy: a weak finding needs better legs | 22 |
| conflict: two findings contradict each other | 4 |
What comes out
Where the ground is thin
| psyops | only 1 findings, thinly covered |
| aftercare | only 1 findings, thinly covered |
| dbt emotion regulation | only 1 findings, thinly covered |
| news | only 1 findings, thinly covered |
| ai welfare | only 1 findings, thinly covered |
| resilienz engineering | only 1 findings, thinly covered |
The gap report is what drives the night run: thin domains get requests first.
Three findings, in full
Read from the graph 3 September 2026Findings are filed in German; these three are translated by hand and shortened. They were chosen for a reason a visitor can check: one runs as a guard on the lab's own machines, and two carry the condition under which they would be wrong. Counts and dates are those of the graph on the day above.
In the theory of self-organised criticality an avalanche is a rooted tree, and everything depends on the mean branching ratio: below one, avalanches stay finite with an expected size of 1/(1 - b); at one, the expected size diverges. An agent that delegates work to child agents builds exactly such a tree. Its two configuration limits, how many children at once and how deep they may nest, are not two independent conveniences but together the one order parameter. Near the critical value the expected total cost diverges, not just its spread. A watchdog that limits time or cost per child regulates the wrong quantity: it measures the node, not the branching ratio. The right safeguard is structural, a depth clamp, not a budget.
- fields
- complex systems and an open-source agent framework
- filed
- 25 July 2026, by the night run
- legs
- 2, one passage in each field; re-read by a model on 3 August, rated adequate
- breaks when
- a measurement shows that raising the depth limit from 1 to 3 at constant width grows total children and cost additively rather than multiplicatively. Then the fan-out is a flat queue with a label, and the analogy carries nothing.
- where it went
- nowhere in code yet. It reads as a rule for how this lab should size its own agent fan-outs, depth before width, and the measurement in "breaks when" has not been run.
An ordinary critical transition has a control parameter that someone must tune to its critical value from outside. Self-organised criticality arises when the order parameter feeds back on the control parameter so that it is pulled to the critical value on its own. Sornette names the bridge in a chapter heading, "Nonlinear Feedback of the 'Order Parameter' onto the 'Control Parameter'". Control theory has the same argument in its own words: integral action drives the steady-state error to zero without knowing the plant's gain, where the alternative needs exact calibration. The finding adds one condition the earlier version had missed: the critical value is a number in arbitrary units, so the loop needs a scale-free indicator of the distance to it. Such systems regulate on the avalanche statistics, not on the density.
- fields
- control theory and self-organised criticality
- filed
- 1 August 2026, as a re-provenance of a finding whose three passages had all sat in neighbouring fields
- legs
- 2, one in each field; re-read on 4 August, rated adequate
- breaks when
- a self-organised critical system is shown whose feedback regulates directly on an absolute value of the control parameter without a scale-free indicator. Then the critical point is addressable from outside and the analogy to the calibration-free integral controller falls.
- where it went
- nowhere yet. It is a finding about the world, not about the lab, and it waits for a use or a contradiction.
The question was the agent's own, filed by its nightly self-review: why does it check itself reliably before risks that are salient, money, deletion, a visible restart, and fail before risks that are latent, such as deploying from an unclear source of truth, although a written rule exists? The night's answer came from cognitive psychology: this is a prospective-memory failure under low cue salience, not a gap in knowledge. Intention-based remembering needs either a salient cue or an external forcing function, the poka-yoke of factory floors. Two papers from 2025 and 2026 carry the same result for language models: errors in one's own output stay uncorrected far more often than the same errors presented from outside, and deterministic gates that fire on every pass beat any risk-triggered reflex.
- fields
- the agent itself and neuro-psychology
- filed
- 5 July 2026, by the night run, answering a request the agent had filed half an hour earlier that same night
- legs
- 0 in the graph. This finding never passed the adequacy check; its sources are papers on the open web, not passages in the lab's corpus.
- breaks when
- not stated at filing; the rule that every finding must name its breaking condition is one day old.
- where it went
- into a hook on the agent's own tool calls, live since 14 July 2026, sharpened on 1 August. Before every irreversible command it injects one question: was this state measured just now, or inherited? It cites the request number in its header. It fired several times while this page was being written.
The third card is the honest odd one out. By the engine's own metric it is the weakest of the three, and it is the only one that changed how the lab runs. That gap between "well evidenced" and "in use" is left visible here on purpose.