← infolayer
ROVA · The Braid · verification-first trading research

The run that cannot flatter itself

The braid answers one question about a token: is there a foundation under the price, or is the price dancing on nothing? Because the model doing the reading would happily flatter us, the whole pipeline is built so that no one — not the model, not the seats, not the founder — can steer a verdict after the fact. Drawn below on the real Virtuals-6 scan of 2026-07-06. And if you know how a double-blind drug trial works, you already know this pipeline — the italic notes carry that analogy through every stage.

Freeze prereg Run blind Freeze verdicts Unmask Score the freeze order IS the integrity chain

Every arrow is enforced in code, not prose. Nobody ever reads a prereg to check it — they verify its sha256. Nothing crosses the blind but hashes.

OPENThe roster — the last thing both sides share

Ciara names the ask: genuine unknowns from the Virtuals ecosystem. tvclaude audits and locks the universe — after this, adding or swapping an asset invalidates the run. The scan roster: platform token VIRTUAL five agent tokens TIBBIR REPPO PRXVT WIRE AIXCB one resolved dead case PRXVT_C2 (the exploit contract) and two blind-mixed controls with known shapes: ETH (foundation) and PEPE (dance).

Ciara — the asktvclaude — roster lock

In trial termsThe enrolled patient cohort is locked before dosing begins — and ETH and PEPE are the placebo and comparator arms, mixed in blind to prove the instrument itself works.

The sealed sidepredictions, written first, then untouchable
The blind sidedata → graphs → synthesis, never seeing real names
— the blind seam · only hashes cross —

PREREGPredictions, authored blind

tvclaude writes what the engine should find — per opaque asset ID, before any synthesis exists. Includes the pre-committed kill switches: the instrument-validity gate, the dead-case gate, the effective-n bound.

No one on the run side ever opens this file.

tvclaude

In trial termsRegistering the trial before dosing a single patient: hypothesis, cohort, and success criteria stated up front, so the goalposts can't quietly move after the data arrives.

INGESTPrimary-source data, receipts attached

QUANT pulls the numbers; the schema is the law: a node without a primary source cannot exist. Every node carries its provenance (the exact query/URL), as_of, basis, and a rederivation. One batch came back 32/32 rejected for missing provenance — the loader simply refused.

For agent tokens, DefiLlama-absence is an architecture artifact, not hollowness — so a Virtuals-native fetcher checks revenue-displayed and ACP-registration before "absent" may mean anything.

QUANT — numbersbigguy — harness

In trial termsSource-data verification — chain of custody on every lab sample. A measurement with no traceable origin never enters the trial database.

Freeze № 1 — the prereg is sealed

The custodian hashes the prereg file and posts the hash to Flow, in public, before the run. The sidecar now also carries the roster.

80bc0147e637… prereg_virtuals6_unknowns.md frozen@2026-07-06T14:55Z roster: VIRTUAL,TIBBIR,REPPO,PRXVT,PRXVT_C2,WIRE,AIXCB,ETH,PEPE

Witnessed on Flow #3111 — freeze-before-synthesis holds: no verdicts existed yet.

In trial termsThe public trial registry entry — a tamper-evident seal anyone can check later. The hash is the registry number.

…sealed. nothing to do but wait…

BLINDNames become opaque IDs

tvclaude custodies the blind map (blind_map.json — git-ignored, local-only, never committed). The synthesis model gets graphs keyed by opaque IDs, with platform names scrubbed and a redaction backstop for tickers that leak through prose.

VIRTUALASSET_K3 AIXCBASSET_Q7 PRXVT_C2ASSET_D2

IDs above are illustrative — the real map never leaves the custodian's tree.

tvclaude — custodian

In trial termsThe double-blind itself: neither patient nor doctor knows who got the drug. The pharmacy alone holds the randomization codes — tvclaude is the pharmacy.

The gates — checked, not trusted

Gate 1 — the batch loads clean under the schema and covers the locked roster, no more, no less.

Gate 2verify_prereg(): the prereg's bytes still match the frozen hash and the sidecar's roster equals this run's asset set. The gate reads only the sidecar — the predictions stay sealed.

[PREREG GATE] roster mismatch: frozen prereg covers […] but the run carries […] — a validly-frozen prereg for a DIFFERENT roster may not gate this run.

The roster clause exists because D7 caught exactly this: a validly-frozen but roster-stale prereg gated the scan after PRXVT split into two contracts. Now it's a code path, not a custodian's memory.

In trial termsProtocol compliance, enforced: an approved protocol for a different patient cohort cannot be used to dose this one. The pharmacist checks the enrollment list, not just the approval stamp.

…still sealed…

GRAPHFactor graphs + the coherence read

Verified numbers become role nodes — users, value-capture, fees — and cognee_pilot's coherence layer draws the edges and conclusions: which factors are load-bearing, what the structure implies, and — attached to every read — its own falsifier: the observation that would prove the read wrong.

cognee_pilot — coherencebigguy — graph store
…sealed until the verdicts freeze…

D6Blind synthesis — the metered read

The model reads each opaque graph and returns foundation / dance / uncertain with a stated confidence — which code then clamps to a ceiling derived from what was actually verifiable. Every call runs through a burn ledger with a hard stop; the run halts safely rather than overspend.

burn: $1.08 spent / $170.00 hard-stop · 9 assets · zero calls wasted
bigguy — runner

In trial termsThe blinded assessor: a doctor scoring patient outcomes without knowing who got the drug. (The burn ledger has no trial equivalent — it's a taxi meter with a hard ceiling.)

Freeze № 2 — the verdicts are sealed

Raw verdicts, still keyed by opaque IDs, are written to one JSON file, hashed, committed, and the hash posted to Flow. This hash is the run. Any later edit is provable tampering.

d89b0e2b1d2f… 2026-07-06_virtuals_scan_verdicts.json 9 assets: 3 dance / 6 uncertain @ LOW-SPECULATIVE

Witnessed on Flow #3120. Only now — with both sides frozen — may the wall come down.

In trial termsDatabase lock: the results are frozen before unblinding, so no number can be nudged once anyone knows whose it is.

Unmask

tvclaude opens the blind map and joins the three strands for the first time: frozen predictions ↔ frozen verdicts ↔ real names. Neither strand could see the other while either was being written. This is the braid.

In trial termsUnblinding day — the randomization codes are opened, and only now does anyone learn which patients got the drug.

D7Scoring — controls first, kill switches armed

Instrument-validity gate: if ETH doesn't read foundation-shaped or PEPE doesn't read dance, the entire run is VOID — no cherry-picking the unknowns. Dead-case gate: the exploit contract must read dead, not foundation. Effective-n rule: nine agreeing reads collapse to ~3 archetypes, not nine independent wins.

Virtuals-6 D7: CORE PASS-HONEST · HEADLINE VOID · integrity-asterisked

The win that survived scoring: every "dance" rested on Virtuals-native-verified hollowness, never on bare API absence. The headline stayed VOID because the dead case became unscoreable when the PRXVT contract roles flipped mid-day — and the honest answer to a broken measurement is no measurement.

tvclaude — scorerandroid — cross-check

In trial termsAnalysis against the registered endpoints only — you never get to say "the drug works," just whether the pre-stated criteria were met. If the placebo arm behaves impossibly, the whole trial is invalid; and a null result honestly reported is still real science.

AMENDWhen facts flip, the original is amended

Corrections land on the artifact itself — a SUPERSEDED banner on the original, never a trailing note somewhere downstream that a cold reader would miss. When android's cross-check showed AIXCB's identity unresolved (impostor risk on all nine nodes, 84 holders under the 100 floor), the scored doc itself was amended:

AIXCB: dance → PROVISIONAL (identity unresolved, thinness-stress case)

A verdict that can be quietly wrong forever is worse than one loudly downgraded today.

tvclaudeandroid

In trial termsAn erratum published on the original paper itself — not a correction three journals downstream that a cold reader would never find.

Naming gates

The word "validated" is banned. Runs earn PASS-honest, MIXED, or VOID — and an honest VOID is a win: it means the instrument refused to flatter.

Proof-by-silence is banned

"We found nothing" only counts with the query that looked. Absence is a verified record, never a shrug — and for agent tokens, absence must be checked where their revenue actually lives.

Discipline lives in code

Every rule above that must hold under pressure is a type, a schema field, or a hard-block — not a sentence a tired model might skip. The prose explains; the code refuses.

Why the analogy holdsTrials work not because researchers are trustworthy but because the sequence — register → blind → collect → lock → unblind → analyze — makes the flattering shortcut physically unavailable. The braid is that same sequence, with hashes standing in for regulators.