← come-to-AGI-26
AGI-26 · San Francisco · 2026

A tour of the poster

Begin with At the Breakpoints, a real-world engineered hypothesis surface. Then enter the paper in plain language and read the poster top to bottom and rail to rail. For the paper’s technical claims, proofs, experiments, and reproduction artifacts, see the materials or ask OmegaClaw.

Extended layman’s abstract

What is the paper actually saying?

A research system should reveal the state of an inquiry, not merely manufacture its cleanest-sounding answer. The Hypothesis Surface is a proposed workspace for keeping evidence, uncertainty, conflict, absence, and next actions visible at the same time.

Most AI systems are built to produce an answer. A research system has a harder job: it must track where each claim came from, how confident it should be, which claims conflict, what remains unknown, and what observation could change the picture. Compress those distinctions into one smooth paragraph and the system may sound intelligent while concealing the real state of the inquiry.

01 · Evidence

Claims carry fibers

A claim is not allowed to appear alone. Its fiber records confidence, conditions, resolution, and supporting witnesses. Similar sentences can therefore remain different epistemic objects.

02 · Conflict

Disagreement leaves a certificate

A conflict is not averaged away or handed to another model for a verdict. It becomes a durable, typed record of what failed to fit and what kind of revision might matter.

03 · Absence

Unknown regions become voids

A void records what is missing, which claims border the gap, and what evidence could resolve it. Absence becomes structured research state rather than an invitation to guess.

04 · Imagination

Voids generate hypotheses

The system may extend a boundary, expose a hidden comparison axis, transport an analogy, or invert an obstruction. Every proposal enters at low confidence with its origin attached.

CLAIMS
FIBERS
CERTIFICATES
VOIDS
HYPOTHESES

Generation is not promotion. A proposed hypothesis does not become a finding merely because the system produced it.

0.36 → 0.00lexical masking rate

Anti-masking limits how certain the answer may sound.

The final wording may hedge, qualify, or narrow a claim, but it may not express more confidence than the attached evidence supports. In the paper’s small deterministic evaluation, masking measured about 0.36 with the rule disabled and 0.00 with it enabled. That demonstrates a mechanism-level behavior change—not perfect honesty, full semantic understanding, or universal performance.

Inspectable implementation

A deliberately plain pipeline

The proof of concept uses CPU-only Python, inspectable rules, and no language-model judge deciding which claims are true. It extracts claims, attaches fibers, propagates confidence, scores conflict, maps voids, proposes hypotheses, and checks rendered language. The claim is not that lexical machinery solves science; it is that epistemic bookkeeping can be explicit and falsifiable before richer semantics are added.

Mathematical companions

Differentiation without erasure

The paper connects claim differentiation and integration to a broader geometric picture. Public companions develop carefully scoped versions of the Fuzzy Fundamental Theorem, Observer-Relative Stokes, a masking-to-residue bridge, and observer-bounded critical scaling. Their assumptions and limits remain part of the claims; they do not turn every metaphor into a theorem.

Poster map

Eight stops on the board

M1

Eyebrow

AGI-26 · SAN FRANCISCO — the venue, and nothing else. Earlier drafts carried an "ACCEPTED PAPER" line; the authoritative poster drops it.

M2

Title

THE HYPOTHESIS SURFACE — an operational epistemology for autonomous research.

H.0

The cleared middle

The center of the board is left open on purpose. Nothing is printed there — the void is the surface a hypothesis gets written against.

H.1

COME, in a chalk-fiber enclosure

An irregular, hand-drawn ring encloses the word. On the physical board it's drawn in real chalk over an invisible transferred guide — the first mark on the substrate is literally hand-made.

R1–R3

Left rail — operational spine

Reads upward, facing the void: CLAIMS CARRY FIBERS · CONFLICT YIELDS CERTIFICATES · VOIDS GENERATE HYPOTHESES.

I1–I3

Right rail — integrability conditions

Reads downward: NO BARE CLAIMS · ANTI-MASKING  σ̂ ≤ σ + δ · CERTIFICATES — TORSION / NON-TORSION. One formal anchor, not a bulleted list of results.

M6

The μ readout

μ  0.36 → 0.00  ·  TEXT-COMPUTABLE  ·  JUDGE-FREE — one quiet empirical stamp in the footer, earning the board its scholarly weight without turning it into a results banner.

M7

QR

You're standing where it points. This site is the destination — curated links out to the paper, the official archived materials, OmegaClaw, and everything built alongside it.