The Model Under Load

Click on a panel
Log note — entries are not edited retroactively to look more accurate after the fact. If a prediction is wrong, it stays wrong, visibly, with a note added — not quietly rewritten. Owner formatter — private, not in the nav.
Panel 01 Open to any reader

Analysis Sample

Two entries, open to any reader. Each is a dated attempt to use the Gears model as a live prediction tool — including where those attempts went wrong. Nothing here is edited after the fact to look more accurate.

Log 001 — Venezuela transition Free Retrodiction Dressed as Prediction 14 August 2026
● Open — unresolved ▶

First attempt. I asked Gemini 3 to take the Gears model and apply it to post-Maduro Venezuela as a prediction exercise — to see whether the four-gear framework could say something useful about where the country's transition was actually heading.

→ Read Gemini's original output in full (PDF)

What Gemini produced

Gemini 3 output

A full article, "The Architecture of Re-Lock," stating — in the present tense, as settled fact — that Venezuela's state apparatus had already completed a "phase transition" into a new, stable "transactional hybrid state," with named companies and bondholders described as the literal "gear teeth" locking that new equilibrium into place.

Read as: a finished result. Written like an autopsy.

What Claude caught

Claude fact-check

1. Real facts, wrong tense. The underlying facts — Centerview Partners' appointment, Chevron's Orinoco stake increase, the $150–170B restructuring figure — all checked out. But every one of them predated the writing. It wasn't predicting anything; it was relabeling already-public news with the model's vocabulary after the fact.

2. A settled outcome that isn't settled. Live reporting as of August 2026 shows talks between the Rodríguez government and the opposition only began August 6 — no election date set, no agreed political structure, nothing resembling the "hybrid state" described as already built.

3. A fabricated detail. A "Bondholder Steering Group" of three firms was named. The real body is the Venezuela Creditor Committee — a different name, missing several real members.

Verdict: retrodiction dressed as prediction — fluent, well-sourced, pointed at an outcome that hadn't happened yet.

The corrected, dated prediction

Recorded 14 August 2026

Venezuela will not hold a genuinely free and fair election within 18 months

Gears predicts continued "normalization without transition" — the Rodríguez government using material control, trait selection, comms framing, and fatigue-driven pragmatism to entrench its position while avoiding a real handoff of power — for as long as that strategy stays viable.

● Open — unresolved

Proven wrong by: a credibly free election held and contested within the window, with a real possibility of transfer of power.
Known weakness: this model has no theory of rupture yet — it predicts continuation by default, not collapse. An external shock or internal elite fracture could break the forecast without the four-gear logic itself being wrong.
Log 002 — Stablecoin market Free The Sentence Didn't Inherit the Discipline 14 August 2026
● Open — unresolved ▶

Second attempt at the same process — this time using Grok, prompted to apply Gears to a live case, the stablecoin market. It used the corrected document template from Log 001. It still made a version of the same mistake, in a subtler place.

→ Read Grok's original output in full (PDF)

What Grok produced

Grok output

A structurally sound document — snapshot, gear mapping, falsifiers, prior-error correction note, all present and well-formed. Market data (USDT ~$183B, ~61–64% share; USDC ~$72B; top-two dominance ~85–89%) was accurate against independent trackers.

The prediction itself: "USDT remains the strongest broad landing hold point... concentration at the top two remains the dominant lock-in feature."

Read as: today's snapshot, restated as a forecast.

What Claude caught

Claude fact-check

1. Predicting the present. USDT is dominant right now. Forecasting "the currently true thing stays true" isn't a test of the model — it's the null hypothesis. A model only earns credit by calling something that isn't already obvious from today's numbers.

2. A live disconfirming signal, missing from the snapshot. Independent reporting from two days before this document's own date shows USDT's supply contracted by roughly $4B over 60 days, with $870M redeemed in just 11 days — and USDC's supply declining too, which undercuts the simple "capital rotates between the top two" story the document assumes. That data was available and wasn't cited.

3. Structural correction without sentence-level correction. This is the important one. The document has a properly hedged structure — a Falsifiers section exists, thresholds are specific. But the individual prediction sentences inside the structure are still written with full declarative confidence, the same as Log 001. Copying the scaffolding of a corrected document doesn't automatically copy the discipline that scaffolding was built to enforce.

Verdict: template inherited, discipline didn't — a more subtle repeat of the Log 001 error, one level deeper, this time in a different engine entirely.

The corrected, dated prediction

Recorded 14 August 2026

USDT's dominant share is more contested over the next 6 months than the snapshot alone suggests

Gears' material-control gear predicts continued top-two structural dominance in absolute terms (no credible third issuer reaching scale within 6 months). But the live redemption pressure on both USDT and USDC — not yet reflected in a share shift to any challenger — is the variable actually worth tracking, not the static concentration number this document led with.

● Open — unresolved

Proven wrong by: continued or accelerating simultaneous contraction in both USDT and USDC supply without a corresponding rise in a named challenger — which would suggest capital leaving stablecoins entirely, a dynamic outside this model's four gears.
Known weakness: same as Log 001 — no rupture theory. This entry additionally exposes a second weakness: the model has no built-in check for whether a "prediction" is actually just describing the present.
Log note — entries are not edited retroactively to look more accurate after the fact. If a prediction is wrong, it stays wrong, visibly, with a note added — not quietly rewritten. Owner formatter — private, not in the nav.
Panel 02 Anyone can open · cases gated

Real World Applications

Real-world applications of the model. Anyone can read the headers and topics. Opening a case is for Analyst members. Source documents stay attached so you can read what each engine actually produced.

Case 003 · Members Members Cognitive Tempo Decoupling & Workframe Management Machine-speed production vs. human-scale institutions through the Gears Lattice · 9/15/2026
● Open — unresolved ▶

Three AI engines and a human fuser examined how cognitive production tempo decoupling disrupts occupational workframes, mapping the friction across the four Gears Lattice variables while addressing why legacy institutions remain anchored to human speeds.

What each engine produced

Panel 1 — Perplexity

Perplexity mapped the core bottleneck to human verification limits, arguing that material control, selection traits, communication flows, and emotional-moral costs remain locked in legacy human-scale equilibria. It highlighted the structural strain of layering continuous machine output onto rigid institutional architectures.

Panel 2 — ChatGPT

ChatGPT expanded the analysis by reframing the bottleneck from mere verification to consequential human judgment and authorization capacity. It introduced the risk of skill atrophy and explored how organizations select for opacity-tolerant delegation traits.

Panel 3 — Copilot

The fuser and final analysis synthesized these views, noting that while engines agreed on the mismatch, they diverged on whether the primary crisis is verification overload or the systematic erosion of foundational human expertise.

What the Verifier caught

Verifier fact-check

The verification pass confirmed that empirical citations regarding agentic workflows and labor market friction are accurate and well-contextualized. The panels successfully applied the lattice without introducing structural or factual errors.

Verdict: strong convergence on structural friction, with complementary expansions on authorization capacity and skill reproduction
Case 002 · Members Members Global Energy Infrastructure & Straits Sequential chokepoints, managed dependence, and a live disagreement about whether the loop can break slowly · 8/28/2026
● Open — unresolved ▶

Three engines — Grok, Gemini, and Perplexity — applied the Gears Lattice to global energy chokepoints, focusing on the Strait of Hormuz and the Strait of Malacca as sequential nodes on the same supply chain. Gemini and Perplexity were shown prior output before writing their own, so this is a relay rather than three independent replications. No human Fuser analysis was submitted for this run.

What each engine produced

Panel 1 — Grok

Mapped material control to the physical narrowness and limited bypass capacity of Hormuz and Malacca, treating them as coupled rather than independent chokepoints. Argued trait selection favors managed dependence, naval and coercive capacity, and corridor-management skill over rapid diversification. Read information flows as deliberately incomplete (AIS tracking versus dark-fleet opacity) and the emotional-moral cost of clarity as high enough to keep reinforcing the existing geometry, concluding the loop favors continuation by default.

Panel 2 — Gemini

Endorsed the chokepoint mapping but pushed back on treating the material gear as static, arguing electrification and demand substitution are slowly eroding oil's leverage over industrial survival. Added a second selected trait, incrementalism, driven by short political and corporate reporting cycles. Reframed the moral cost as trade-off cognitive dissonance deflected into nationalist framing, and proposed that gradual demand substitution could unwind the loop without needing a sudden multi-variable shock.

Panel 3 — Perplexity

Largely agreed with both prior panels but complicated Panel 2's rupture claim, arguing that GEARS is likely to fracture unevenly at the edges rather than collapse, and that states may reweight investment toward surveillance, stockpiles, and remaining high-value flows rather than let the loop break. Treated the two prior views as compatible variants rather than a contradiction, favoring "reconfiguration" over either pure continuation or clean rupture.

What the Verifier caught

Verifier fact-check

The fact-check found Panels 1 and 3 held up structurally, with only unverifiable (not fabricated) figures flagged in Panel 1's trade-volume and pipeline-capacity numbers. Panel 2 was the one panel with real issues: it stated an ongoing, uncertain process (the material gear "beginning to fracture") as settled fact, and presented several declarative claims — including calling electrification a completed decoupling — with more confidence than the evidence supports. No fabricated context or invented human-analysis content was found in any panel.

Verdict: two panels held up; one panel stated an ongoing trend as already-established fact
Flagged for Human-AI Relations

Worth noting for the log: this run's central disagreement (continuation vs. slow-motion rupture vs. reconfiguration) is a genuine analytical dispute rather than an artifact of panels copying each other, which makes it a cleaner test case than most for how the lattice handles a live rupture-theory gap.

Recorded 8/28/2026 — review window: 12 months

Seaborne oil volume through Hormuz and Malacca does not measurably decline over the next 12 months

All three panels agree current demand substitution is gradual at best, so the working prediction is that maritime chokepoint volumes hold roughly steady rather than showing an early, measurable dip from electrification or demand shifts.

● Open — unresolved

Proven wrong by: published shipping/trade data showing a meaningful year-over-year decline in combined Hormuz/Malacca oil throughput within the window, not attributable to a one-off disruption
Case 001 · Members Members AI Governance & Safety Frameworks Relay, not replication — three engines, one inheritance · 24 August 2026
● Open — unresolved ▶

Three engines — Copilot, Gemini, and Grok — were each asked to apply the Gears Lattice to current AI governance and safety frameworks (EU AI Act, NIST AI RMF, ISO/IEC 42001, frontier-lab safety commitments). Gemini and Grok were shown the prior panel's output before writing their own, so this is a relay, not three independent replications — a distinction the Verifier flags below.

What each engine produced

Panel 1 — Copilot

Mapped material control to regulatory bodies, certification authorities, and compute-rich labs; argued trait selection favors compliance-heavy, documentation-driven organizations over small and open-source actors; read the loop as self-reinforcing toward continued, intensifying governance.

Panel 2 — Gemini

Largely endorsed Copilot's mapping, then added an independent claim: that frontier labs co-shape the very standards meant to regulate them, turning governance into a "regulatory capture" moat that favors incumbents over open-source competitors.

Panel 3 — Grok

Reached the same four-gear mapping independently, with more current detail, but explicitly reviewed and largely agreed with both prior panels — including a direct critique of Gemini's "moat" framing as over-reaching what the lattice itself can establish.

What the Verifier caught

Verifier fact-check

1. Not independent replication. Because each panel saw the previous one's output, the strong three-way agreement is weaker evidence than it looks — Panels 2 and 3 largely inherited Panel 1's frame rather than testing it blind.

2. A factual error in Panel 1. The claim that ISO/IEC 42001 compliance can only be claimed by certified entities is wrong — certification under 42001 is voluntary, and ISO itself does not perform the certification.

3. Unhedged forecasts. Several claims across all three panels stated present-tense observations ("the loop is stable," "the feedback loop remains airtight") as if they were predictions, and Panel 2's regulatory-capture claim asserted intent on the part of frontier labs without supporting evidence.

Verdict: convergent, but not independently convergent — one factual error, and forecasting language that often describes the present rather than the future.
Flagged for Human-AI Relations

This case is a useful contrast to the bypass tests on the Lattice panel: the failure here isn't fabricated human content, it's engines reviewing each other in sequence and mistaking agreement for confirmation. A cleaner test would run this topic blind across three engines with no relay, to see whether the same convergence holds without one panel seeing another's answer first.

Recorded 24 August 2026 — review window: 90 days

Governance frameworks intensify rather than relax over the next 6 months

Consistent with the lattice's default (no rupture theory), the working prediction is continued densification — more reporting obligations, more certification schemes, no major relaxation of a binding regime — rather than a reversal.

● Open — unresolved

Proven wrong by: a major jurisdiction materially loosening binding AI obligations (not just delaying deadlines) within the window.
Known weakness: same as the rest of this page — no theory of rupture, and this run additionally shows the model is sensitive to relay structure between panels.
Log note — entries are not edited retroactively to look more accurate after the fact. If a prediction is wrong, it stays wrong, visibly, with a note added — not quietly rewritten. Owner formatter — private, not in the nav.
Panel 03 Members only

Evolving The Gears Lattice

The original runs that caught each failure and drove each fix — the ones that didn't make the current Log set — kept in full, unedited. Click a topic on the right to read it here.

🔒 Members only

Create a free account to open this panel

Log 001 and Log 002 stay open to everyone on Panel 01. Everything in this panel — the raw, unedited runs behind every log and study on this page — requires a free Analyst account.

Log note — entries are not edited retroactively to look more accurate after the fact. If a prediction is wrong, it stays wrong, visibly, with a note added — not quietly rewritten. Owner formatter — private, not in the nav.
Panel 04 Members only

Discussions

Every case a member has chosen to make public, from every member's profile — one directory, updated the moment someone publishes or un-publishes a case from their own Cases Created panel. Engage takes you to that member's page to read and reply.

Log note — entries are not edited retroactively to look more accurate after the fact. If a prediction is wrong, it stays wrong, visibly, with a note added — not quietly rewritten. Owner formatter — private, not in the nav.