Discovery mission · v0

The Discovery Forge

An assembly line that turns a witnessed research talk into a discovery card carrying its own re-check, then asks the verifier to decide. Most cards are not yet decidable, and the forge says so.

Zain Dana Harper/ Telos · composing gather, crucible, and the engine/ v0 scaffold · dated 2026-06-30

What this is

The Telos discovery mission is a witnessed, re-checkable reformulation or rediscovery of physical and mathematical law. The discipline that makes it more than assertion is a single gate: a result must be derived or simulated and then checked against a criterion it did not author, not merely asserted. The verifier (crucible) returns one of three verdicts: MATCH, DRIFT, or UNVERIFIABLE.

The forge is the assembly line. It reads a sealed talk transcript, writes a claim card grounded verbatim in that transcript, attaches a concrete test that could confirm or refute the claim, runs the simulate-versus-assert gate, and stores the card together with its check and its verdict so the whole thing is re-runnable. The forge does not produce discoveries on its own. It produces candidate cards with an explicit, named falsifier attached, so a later run with a real measurement or a sound oracle can return a grounded verdict instead of an opinion. A card with no falsifier is rejected at intake.

The pipeline

  • Source. A sealed transcript or corpus item, intaken through gather with a provenance receipt.
  • Claim card. A verbatim-grounded statement with its source reference and seal. No paraphrase that drifts from the talk.
  • Named falsifier. A concrete measurement or simulation that could confirm or refute the claim. No falsifier, no card.
  • The gate. Was the result derived or simulated, or only asserted? The criterion has to come from somewhere other than the claim itself.
  • Verdict. crucible returns MATCH, DRIFT, or UNVERIFIABLE, with a re-derivable proof.
  • Discovery object. The card, the check, and the verdict, stored together and re-runnable.

Where the cards came from

The corpus is eight research talks read as a research diet: cosmology, a thermodynamic AI chip, quantum entanglement, the philosophy of time, cryptography, linguistics, a brain-computer interface, and the state of AI for mathematics. They were intaken through gather into a witnessed corpus and sealed. Ten items verified MATCH against their digest (the eight talks plus run metadata).

Before any cards were written, the broader thesis that these talks map onto the Telos design was run through crucible. It returned UNVERIFIABLE on all six convergence claims. There is no oracle that can measure whether a physics talk supports a software design, so the verifier correctly declined to certify the mapping. That refusal is on the record and is honored here. The transcripts are sealed and re-checkable. The interpretation of them is mine, and it is labeled as interpretation, not measurement.

The first four cards

Each card names a claim grounded in a transcript line, the test that would settle it, and a verdict slot. Three of the four slots read UNVERIFIABLE on purpose: these are candidate targets, not results. That is honest, not a hedge. The forge has not yet run those checks against a criterion it did not author.

C1 · Turok's dated neutrino-mass predictionphysics · highest-value target
  • The claim, from Neil Turok's colloquium: if one right-handed neutrino is stable and is the dark matter, the lightest neutrino is massless, which forces a predicted minimum sum of neutrino masses, a "bump" testable in three to four years via galaxy clustering.
  • Test: track the published upper bound on the sum of neutrino masses over the stated window. Falsified if the measured bound drops below the predicted minimum. Confirmed if a bump appears there. This is a real, dated, external measurement the forge does not control.
  • Verdict slot: UNVERIFIABLE (proposed). It becomes decidable only when the survey data is in hand. It is the cleanest verifier target in the set precisely because it is dated and externally re-checkable, not because it can be settled today.
C2 · Bell's move, as a method templatequantum · interpretive
  • The claim, from the entanglement explainer: in 1964 Bell turned a philosophical dispute into a measurement. Local hidden variables imply a bounded correlation; quantum mechanics predicts a violation that is "stronger in a specific measurable way," and experiments can count the outcomes.
  • Test: this card is a method template, not a physics prediction the forge will measure. Its check is structural. Given any candidate card, can the forge state a bounded inequality whose violation separates the claim from its null, the way Bell separated local-hidden-variable bounds from quantum predictions? A card that cannot be cast that way is weaker.
  • Verdict slot: UNVERIFIABLE (proposed), marked interpretive. As physics history, Bell's program is settled. Its use as a forge pattern has not been run through any measurement, and crucible cannot certify the analogy.
C3 · Thermodynamic chip, the SDE-settles-to-matrix-inverse mechanismre-checkable now
  • The claim, from the thermodynamic-chip interview: the chip is built so it behaves like a stochastic differential equation; you let its own noise settle into place, and it ends up behaving according to the inverse of a programmed matrix. Noise is the computation.
  • Test: implement the Ornstein-Uhlenbeck-style SDE whose drift is set by a target matrix, run it to its stationary distribution, and check that the sampled mean and covariance recover the matrix inverse within tolerance, against a direct linear-solve reference. The reference is the criterion; the SDE is the result under test.
  • This is the one card that needs only code, no external party. It was taken from proposed to a real, measurement-grounded verdict. See the C3 result page.
C4 · The seven-loop quadratic-gravity calculationmath · aspirational, parked
  • The claim, from Turok again: quadratic gravity, in a limit, is the "perfect square theory," which embeds into a scalar field theory already known to seven loops. So there is a prediction of quantum gravity in that limit to seven loops, with the aside, "good luck if you have an AI, maybe it can do that calculation."
  • Test: extend the quadratic-gravity calculation toward the seven-loop figure already known on the scalar-field side via the stated correspondence. Each loop order is a checkable target the forge does not author.
  • Verdict slot: UNVERIFIABLE (proposed), honestly out of reach. Telos has no symbolic engine that can carry seven-loop perturbation theory today. Parked as a candidate because it is a well-posed, externally checkable goal rather than a vague one.

The one card that moved: C3

C3 was the only card that could flip with local work alone, and it did. The thermodynamic-chip claim was taken from proposed to a real crucible verdict by running the simulate leg in code: an Ornstein-Uhlenbeck process for a symmetric positive-definite matrix, six thousand chains over forty thousand steps, with reference values from a standard linear solver.

  • Stationary mean versus the linear solve: relative error 0.0099, against a 5 percent bar. crucible MATCH.
  • Stationary covariance versus the matrix inverse: relative error 0.033, against the same bar. crucible MATCH.

What that confirms, stated plainly: the mathematical claim that this SDE settles to the matrix inverse is now a re-checkable MATCH on the simulate leg. Whether a physical chip realizes the same behavior within its noise and device limits is a separate leg, and it remains UNVERIFIABLE here. The forge advances a claim one leg at a time, and says which leg it has earned. The full C3 page carries the method, the numbers, and the seals.

Three internal targets the cards exposed

These are not physics discoveries. They are the machinery the forge needs so its cards can return real verdicts, drawn straight from the talks.

A fully-correct fraction gatecrucible
  • From the chip talk: a benchmark reported "70 to 80 percent of tests correct," and the sharper question was whether it got any of them fully right. Add a crucible mode that reports the fully-correct fraction separately from the per-check pass rate, and refuses a MATCH on partial pass alone. Partial pass is not pass.
Prove or disprove in the reconcilecrucible
  • From the chip talk's prove-or-disprove pattern, reinforced by Bell. When crucible cannot prove a claim, it should explicitly try to disprove it (find a counterexample), and treat "neither proved nor disproved" as a first-class UNVERIFIABLE rather than a silent failure. That is what lets a card return DRIFT honestly instead of stalling.
Spec-legibility, a two-way renderthe engine
  • From the chip talk's flag that getting a formalization right and legible is harder than the proof. The engine's two-way render should take a card's formal criterion and render it back into a human-inspectable form so the operator can confirm "yes, that is what I meant." Without this, a card can pass its check while testing the wrong claim.

Honest limits

The honesty is the point of this page. Card by card:

  • C3 is re-checkable now, and has a real MATCH on its simulate leg. It needs only a simulation and a linear-solve reference, no external party. Its physical-chip leg stays UNVERIFIABLE.
  • C1 is re-checkable but not yet, by construction. It is dated to a survey window and the data is not in hand. It is the highest-value target precisely because it is dated and externally falsifiable, not because it can be settled today.
  • C2 is interpretive as a Telos mechanism. Its value is the pattern of turning a dispute into a bounded measurement, not a number the forge can produce now.
  • C4 is out of current reach. It is well-posed and externally checkable in principle, but Telos has no engine that can carry that calculation. Aspirational, parked honestly.

The talk-to-Telos mappings are interpretive. crucible already returned UNVERIFIABLE on all six convergence claims, because there is no oracle that can measure whether a physics talk supports a software design. That refusal stands, and cards C1, C2, and C4 inherit the same caution. None should be cited as a verified Telos result. What is sealed and re-checkable is the corpus itself, the verifier's recorded refusal, and the one C3 simulate-leg MATCH. The reformulation into cards is a faithful, labeled reading, not a measurement. A card is only as honest as its named falsifier, and the forge rejects at intake any card that cannot say what would refute it.

A v0 scaffold that produces candidate cards with attached falsifiers, not discoveries. One card (C3) has flipped to a real MATCH on its simulate leg; the rest are UNVERIFIABLE by construction, and labeled. Read the verifier: crucible. The witnessed intake: gather. The research program behind it: Senses and Sensibility. · Dated 2026-06-30.