TOWOW PUBLIC DATA · SNAPSHOT V1

Publish the numbers.Publish their limits beside them.

The smallest public data surface for three agent-collaboration studies: negative results, synthetic trial outputs, reconstruction state, checksums, and the claims these files do not support.

release date
2026-08-18
study topics
3
public trial rows
2,700
file-level integrity
SHA-256

01 · THREE EVIDENCE SHAPES

Not every experiment retained the same kind of raw material.

The A2A study exposes a sanitized aggregate. The commitment study keeps a historical summary plus a verifiable reconstructed state. The disclosure study retained trial-level synthetic output. Missing originals are not replaced with invented replicas.

A2A NEGOTIATION

Six rounds, with AcceptedOriginalValue still at zero

One real engineering task, two synthetic authority roles, and three operational arms. Direct A2A refined the specification but produced no signature, target adoption, or acceptance. Two long-output transport failures leave strong-center capability unknown.

task
1
main-run model calls
10
A2A rounds
6
AcceptedOriginalValue
0

The release excludes full prompts, raw transcripts, private dossiers, internal identities, private code, provider and model identity, tokens, request and response payloads, and cache-routing metadata. Per-arm cost and model-call duration remain public. N=1 does not establish general A2A cost or superiority.

Read the study note
Sanitized aggregateJSON · 3.3 KBDownload ↓Closed-field schemaJSON Schema · 6.8 KBDownload ↓

CONCURRENT COMMITMENT

60,461 / 100,000 is a synthetic counterexample, not an incident rate

When two requests separately read one balance, check-then-write can approve mutually conflicting demands. The analytic probability is 49/81. The comparison zero is a coded invariant, not zero failures across 100,000 database transactions.

synthetic request pairs
100,000
overspend combinations
60,461
analytic probability
49/81
real transaction runs
0

Historical trial rows were not retained. The RNG state was reconstructed from the seed and call order; it is not presented as an original artifact. No database, concurrent schedule, performance, or recovery evidence exists.

Read the study note
Historical summaryCSV · 85 BDownload ↓Result and analytic checkJSON · 1.1 KBDownload ↓Reconstructed RNG stateJSON · 688 BDownload ↓

BOUNDED DISCLOSURE

Twenty-two versus ninety constraints does not automatically mean privacy

Three hundred synthetic feasible worlds, five actors, and eighteen linear constraints per actor. Median cumulative disclosure was 22 for adaptive_1 and 90 for full disclosure. Nine methods produced 2,700 retained trial rows.

synthetic instances
300
methods
9
trial rows
2,700
median disclosure comparison
22 / 90

Worlds were feasible by construction; oracles were exact, honest, convex, and linear. No people, Unknown or Refuse behavior, malicious probes, reconstruction attacks, differential privacy, or cryptographic secrecy were tested.

Read the study note
Trial-level outputCSV · 187.4 KBDownload ↓Nine-row summaryCSV · 604 BDownload ↓Fields and limitationsJSON · 2.0 KBDownload ↓

02 · READ BEFORE REUSING

The same number can have a different evidence identity.

  1. 01

    An aggregate is not a raw record

    The A2A JSON is sanitized and field-closed. It supports checking page-level numbers, not reconstructing the conversation or internal identities.

  2. 02

    A reconstruction is not an original

    The commitment RNG state reproduces 60,461, but it was rebuilt from historical call order. The original 100,000 request rows do not exist.

  3. 03

    Trial rows are not the real world

    The disclosure CSV does contain 2,700 rows. Every row still comes from a synthetic, honest, exact, and feasible mathematical world.

03 · INTEGRITY AND CITATION

All 11 payloads and the manifest can be checked.

The manifest records bytes, media type, role, and SHA-256 for 11 payload files. SHA256SUMS covers those 11 files plus the manifest. The checksum ledger itself is the public trust anchor and is not circularly self-hashed. Local paths, internal archive identifiers, branch names, and private-source hashes are deliberately absent.

Preferred citationToWow Research Program. ToWow Public Research Data Snapshot v1. 2026-08-18. https://towow.net/research/data
MANIFEST.jsonMachine-readable file ledgerSHA256SUMS.txtPer-file SHA-256README.mdScope, material, and limitsCITATION.cffPreferred citationRIGHTS.mdRights and reuse status

Not included in this release

04 · RIGHTS AND NEXT RELEASE

Verifiability first. Open licensing requires an explicit decision.

This release establishes a stable public citation and integrity surface without choosing MIT, CC BY, or another license on the rights holder’s behalf. Any future code or open-data release will carry its own license, version change, and checksum ledger.

Back to the ToWow research hub中文