Current synthesis date: 2026-08-01
The accurate statement: Substantial progress in problem definition and scoped mechanisms. The end-to-end product episode is not complete.
Claim boundary: Local and synthetic results are not presented as human, production, or long-term validation.

A formal paper, and a historical slice

Version 1.1 places actors, relations, authority, execution, real-world effect, and acceptance in one theoretical frame. It is an archive, not a replacement for the current research state.

Title: Forming Shared Reality Among Sovereign Intelligent Actors
Author: ToWow Research Program
Version and date: Formal paper v1.1, 2026-07-27
Extent: 143 pages
Subject: Generative coordination theory, protocol, executable semantics, and a human research program for super individuals and one-person companies

The negative results belong in the record

These three result groups changed the design. Each number keeps its own denominator, and later green checks do not erase earlier failures.

Real engineering semantics

A success receipt is not a fact about the target

naive terminal labels were wrong: 10/17
stdout and outer exit status each indicated observed state: 4/9
minimal Effect contract plus target-native readback: 9/9

Seventeen real Harness scenarios led us to separate Attempt, Effect, Adoption, and Acceptance. A second set of nine real actions showed that verification must reach the target rather than stop at executor logs.

The 17/17 semantic alignment applies only to the selected scenarios. It is not a general accuracy claim.

Read the full experiment note

Multi-round negotiation

Richer terms did not create capability

AcceptedOriginalValue: 0

Multi-round model negotiation added useful refusals and counterconditions, but it did not naturally create new capability. A long-output transport failure also left the strong-center comparison incomplete.

More messages cannot be reported as successful formation.

Read the full negotiation study

Fair comparison

A frozen contract still produced no winner

actual comparative runs: 0
winner: NONE

Later audits found that a static validator could accept five kinds of runtime false green. A development manifest proves preparation, not an executed comparison.

Not run means not run. Infrastructure completeness does not decide a winner.

ToWow and Flowness connect without collapsing

The two research lines address different parts of one reality chain. Keeping the distinction makes their composition clearer.

ToWow

ToWow studies relation formation across actors

It asks who represents which Principal, under what Authority, RelationVersion, and Commitment, and who can ultimately accept an Effect.

Flowness

Flowness keeps work and evidence continuous

It preserves work identity, context, versions, Findings, commit boundaries, and acceptance evidence across long-running multi-agent work.

How they compose

Flowness offers execution and evidence mechanisms that a ToWow deployment could compose with; this is not a claim of current product integration. ToWow research continues to ask how relations may form across Principal boundaries. Untrusted agent networks still require identity, isolation, privacy, and anti-collusion controls beyond both.

Read the Flowness overview Read the trust-boundary guide

How we label evidence strength

We do not flatten every result into one “validated” badge. A claim may hold locally while remaining unknown in the real world.

Stable distinctions

ESTABLISHED DISTINCTION: A reality distinction preserved across evidence and counterexamples.

Bounded progress

SUPPORTED SCOPED: Supported only under stated synthetic or local digital conditions.
DESIGN CANDIDATE: A proposed structure not yet run in a qualified task.

Open or rejected

INVALIDATED: The evaluator, input, truth, or attribution was disproved.
UNKNOWN / NOT RUN: Evidence is insufficient, or the work has not actually run.
REAL WORLD UNVERIFIED: Local results do not establish human, organizational, production, or long-term value.

The boundary between public material and internal archives

We will publish material that can stand on its own. Internal paths are not presented as public citations.

Publicly accessible now

Internal or not yet published

The complete internal research archive and redacted implementation specifications
Dynamic research state, raw wave runs, and internal audit material
Human research inputs and raw data still awaiting publication review

These materials can support internal judgment, but this page does not invent download addresses or inaccessible source links before stable public editions exist.

Continue by question

You do not need to read the full history first.

Discovery engine

From candidate possibilities to testable causality

See how PFE System v3 grounds diverse candidates, compiles commitments, runs a formation operator, and checks causality with target-side readback and ablation.

Paper

From an Agent attempt to responsible acceptance

Read an English overview of the Chinese formal paper, its strongest experiments, negative results, and the boundaries narrowed after 2026-08-01.

Release guide

Why later corrections belong before the paper

See why the historical paper, later evidence, public redactions, SHA-256 verification, and rights boundaries travel together.

Evaluation boundary

Zero actual runs means the winner must be NONE

Distinguish NONE, Unknown, a tie, and a failure before turning contracts, tests, or scoped support into an architecture ranking.

Evaluation practice

Commit success after checking identity

Follow an item that identified the wrong task and see why identity and target checks must happen before a batch is marked successful.

Decision boundary

After everyone in the group chat says yes

Turning scattered agreement into an executable arrangement requires specific scope and evidence that can be found again after change.

Discovery boundary

Start with “I am stuck”

Helping someone find a next step sometimes begins by giving them room to revise the question.

Data

Numbers, limits, and SHA-256

Download sanitized aggregates, historical summaries, trial-level synthetic output, data cards, and the public integrity ledger.

Experiment

When “done” did not happen

Seventeen real Harness scenarios and nine real actions show why target-native readback changed the design.

Negative result

Six rounds. Acceptance stayed at zero.

A controlled A2A comparison on a real engineering task, with the failed central baseline kept in the record.

Concurrent commitment

The balance was 100. Both requests said enough.

A 100,000-pair synthetic model and exact 49/81 probability show why final commitment needs one moment that counts.

Bounded disclosure

Can agents say less and still solve the problem?

Twenty-two versus ninety constraint rows shows the value of adaptive inquiry without pretending to prove privacy.

Execution substrate

Flowness

Event truth, context compilation, Commit Gate, independent review, and acceptance boundaries.

Collaboration governance

Agent-to-Agent Harness

Separate communication interoperability from evidence governance, target-native readback, and acceptance by the responsible party.

System layers

A2A, MCP, runtimes, and harnesses

Separate tool and context access, agent communication, application execution, and work evidence governance.

Verification method

Trace is not proof

Move from artifact identity, mechanical checks, and independent review to target-native readback and an explicit decision by the responsible party.

Trust

Trusted and untrusted agents

A shared governance skeleton does not imply shared security assumptions.

Public code and documents

Flowness on GitHub

Read the public architecture, white paper, security boundary, and evidence register.