Theory · working draft
Theory of Collaborative Intelligence
The research loop
- Research question→
- Hypothesis→
- Experiment→
- Finding→
- Architecture change→
- Noetis release→
- Next question
Status rule
A proposition is marked Supported only where a published round's own text reports a supporting finding, and that round is cited. Everything else is Open. Supported means supported in the bounded setting the round describes, not proven in general.
Working propositions (v0)
- v0.1Open
Reliability is a property of the process
Reliability is a property of the system's process, not only of the underlying model.
Working position from Theory v0, not yet tested as stated.
- v0.2Open
Finding evidence and judging it are separate steps
Locating a quotation and judging whether it supports a claim are separate operations.
Working position from Theory v0, not yet tested as stated.
- v0.3Open
Confidence belongs to the final answer
The object that should carry confidence is the final revised answer, not the first draft.
Working position from Theory v0, not yet tested as stated.
- v0.4Open
Improvement counts only against a matched baseline
An architecture change counts as an improvement only when measured against a matched baseline with the same number of calls on an adjudicated endpoint.
Working position from Theory v0, not yet tested as stated.
158 open hypotheses; none yet tested. These propositions are positions to be tested, not results.
Round 40 (NCT-40) propositions
- P1Supported
Structural abstention
A composition can hold only if some component can produce HOLD or a disagreement can produce review.
Round 40 reports H40.1 as Supported (E40-A: a forced majority releases 108/243 states unsafely, the same as one witness, and holds 0/16 HOLD-truth states without abstaining witnesses).
- P2Supported
Read and quote, then decide
Models should locate and quote; a deterministic, span checked resolver should decide.
Round 40 reports H40.2 as Supported (E40-A: 4/12 unsafe with the span check versus 8/15 without).
- P3Supported
Disagreement buys a hold, not a vote
The second independent read should only be able to stop a release.
Round 40 reports H40.3 as Supported (E40-A: disagreement routing raises the fewest faults for an unsafe release to 2). The false hold cost was not observed.
- P4Open
Decorrelation is the safety budget
A routed topology is as safe as its readers are decorrelated.
Round 40 notes all 4 unsafe routed states are identical misreadings; joint misreading rates are not yet measured.
- P5Supported
Evidence-rewriting stages add risk
Stages that can rewrite evidence add risk.
Round 40 reports H40.4 as Supported (E40-C: on X24 a validation stage that rewrote a correct ledger produced the only composed error).
Round 40 names each proposition as a candidate for the Round 41 to 50 block, with a removal test that could reject it.
Roadmap to Round 50 · currently Round 40
- Round 41
Hard pool: items on which witnesses disagree, so topologies separate on accuracy
- Round 42
Measured decorrelation: joint misreading rates across witness pairs and prompts (P4)
- Round 43
Flag only validators versus rewriting validators (P5)
- Round 44
Abstention at scale with externally reviewed HOLD labels (P1, Round 39 handoff item 1)
- Round 45
Resolver failure: where fixed rules are wrong and witnesses are right (P2)
- Round 46
Review routing load: false hold rates and what a reviewer must see (P3)
- Round 47
Working memory and procedure reuse inside the routed topology
- Round 48
Transfer of the structure to a second domain outside contracts
- Round 49
Matched call baselines across all surviving propositions
- Round 50
Theory v1 draft: keep only propositions that survived Rounds 40 to 49
Proposed research path from Round 40, not a published commitment to particular results.
See it on your contracts
This research powers Lexplio and Noetis pilots. The same evidence checks and release rules studied here run on real agreements, with every finding tied to a quoted passage.