Does 56% zero-shot success mean household work now clears T1?

No. T1 is a frequency target — roughly one human decision per hundred executions — and frequency is the wrong axis once a miss cannot be undone. Most of Figure's evaluated actions can be undone. That is why they are a valid generalization test and an invalid reason to collapse the Physical Intervention Threshold into the standard Intervention Threshold. Figure's own rule — a safety intervention fails the rollout — keeps the judgment gate outside the score, which is where it belongs.

Is Index pretraining the same compounding asset as an Operational Ledger?

No. Same shape, different owner. Index makes a behaviour likelier to hold in an unseen room. An Operational Ledger makes this business's exceptions, thresholds, and validated patterns available at execution time. One is a foundation-model substrate. The other is the knowledge layer a Steward governs.

Why isn't this a numbered How We Think memo?

Because it does not extend the model. MHS named the friction/judgment split. Memo #119 priced irreversibility. Helix 2.5 is external evidence that execution can transfer across homes without per-site teaching. That is an observation. Treating it as a new threshold rule would be the inflation those two pieces were written to prevent.

Is this a claim that Arco is entering home robotics?

No. Household locomanipulation sits in the same selection band Operational Arbitrage already uses. That is a rhyme with the MHS lab deployments, not a market entry, and Arco has not built or operated anything in this domain.

Helix 2.5: Zero-Shot Is Not Unattended Judgment

MHS showed that reducing hardware friction does not retire judgment. Helix 2.5 shows that reducing environment-specific teaching does not retire it either.

On 17 September 2026, Figure introduced Helix 2.5: a single Index-pretrained policy adapted into three whole-body behaviours — tidying a living room, folding towels, making a bed — then evaluated in 30 Bay Area homes with zero data collected in any of them. No fine-tuning on the evaluation rooms. No evaluation objects in the task-specification set. One checkpoint for all thirty houses. Success required the whole task: every toy in the basket, every towel folded and placed, the bed finished. No partial credit.

The number that matters is not the demo. It is the ablation. Holding task data, architecture, training, and evaluation fixed, a policy trained from scratch succeeded on 9% of zero-shot trials. The Index-pretrained policy succeeded on 56%. Figure also reports that Helix 2.5 matched a prior in-environment Helix 02 success rate with half the task-specification data, while the behaviour now had to hold across thirty unseen homes rather than the room it was taught in. A human-to-humanoid transfer scaling law sat underneath: doubling Index pretraining data improved downstream robot-action prediction smoothly enough that the largest run's loss was forecast to four decimal places before training.

Those results are real, and they bear on the property the Autonomy Spectrum Framework's first axis scores — who moves the work — under distribution shift, in rooms the system had not been taught. They also bear on the property Process Continuity Score scores, because locomotion, perception, and manipulation had to remain coordinated across the full task horizon. The axes themselves score a business's Revenue Loop, not a research evaluation; Figure did not score either, and the reading is Arco's. Figure is explicit about the limit. General humanoid robotics is not solved. Helix 2.5 is, in Figure's own scope claim, the first demonstration of zero-shot whole-body generalization at this scope on a humanoid — evidence that whole-body behaviour can be learned from human experience and transferred, rather than rebuilt for each address.

That is the sentence Overhead Is a Design Choice already used on software, and that MHS Collapses Hardware Friction, Not Judgment used on device interfaces. Specify once. Stop paying the per-site tax.

What the 56% does not buy

Rollback Cost is still the variable that determines how much unattended execution the architecture can safely tolerate. A misplaced toy is a re-pick. A messy towel fold is a retry. A comforter left short of the headboard is a walk around the bed — which Helix 2.5 now does, and which Figure correctly treats as a qualitative gain from pretraining rather than as a new form of governance.

Figure's own evaluation rule marks the boundary. The appendix states that if a human intervention is necessary for safety, the rollout is aborted and counted as a failure. The 56% is therefore a measure of execution inside a gate, not a measure of the gate's absence: safety judgment is not part of what the score contains, it is the condition that bounds it. A Steward reading the number should read the rule beside it. The evaluation was designed on the split this Perspective is about, and the designers put judgment on the failure side of the line, where the Stewardship Model puts it too.

None of the evaluated actions share a floor with the class The Threshold Was Never Priced for Irreversibility was written to isolate. A mis-pipetted reagent does not re-sort. A laser left off-lock does not get a second pass that undoes the first. The Physical Intervention Threshold calibrates by the cost of being wrong once, not by how often the task occurs. Escalation Rate can improve as Helix self-corrects more often without that changing the reversibility of the underlying action. Reversibility asks a prior question, and 56% in unseen rooms does not answer it: a miss that cannot be undone leaves the gate where it was.

MHS already showed the two halves inside one deployment — a compiled recovery script running 700 trials unattended, and the same stack pausing for confirmation on new physically risky writes. Helix 2.5 is the first half written at home scale. It is not the second half disappearing.

Unattended execution of a behaviour that transferred is not unattended judgment at first contact with a house, an object, or a failure mode the policy has not priced. Confusing the two is how a zero-shot result gets misread as architectural autonomy.

Index is not an Operational Ledger

Figure's transfer law is the same shape as the asset The Operational Ledger describes, and it is not the same object.

Index is a vendor-scale store of human experience. It makes a new behaviour cheaper to specify and likelier to hold in a room the robot has not seen. An Operational Ledger is the structured record of one business's executions, exceptions, interventions, and validated patterns. It compounds inside the firm. Index compounds underneath many firms, or underneath none, depending on who owns the weights.

That distinction matters for anyone tempted to treat Helix 2.5 as a reconstructed market rather than as a substrate result. The selection screen in Operational Arbitrage still holds: high human cost, high coordination friction, work that can be specified once the interface exists. Household tidying and bed-making sit in that band. So did the labs MHS shipped into. A selection rhyme is not a claim that Arco is operating robots in Bay Area homes, and it is not a claim that a foundation-model pretraining set substitutes for the ledger a Steward governs.

The Operator's Verdict

An operator reading Helix 2.5 against the Stewardship Model should score two things separately before changing a threshold.

First: did execution transfer? On three whole-body tasks, across thirty unseen homes, with half the specification data of the previous generation — yes, far enough to matter. Who moves the work is no longer an in-room property by default.

Second: did reversibility change? No. The tasks Figure chose are the right tests for generalization and the wrong tests for retiring a Physical Intervention Threshold. Score them on both axes. Importing the digital 1:100 frequency target because the robot now generalizes across homes — treating "home" as a distribution rather than as an address — is how a frequency metric approves something a single miss cannot unwind.

The Steward-interface chapter did not need this result to stand, and this result does not specify that interface. Helix is the body completing the path. It does not say how the human who owns the exception sees the room when the path breaks.

Technology changes what is possible. The cost of one miss determines what runs without a gate.

KEY TAKEAWAY

What does Figure's Helix 2.5 demonstrate about physical autonomy under the Stewardship Model?

Helix 2.5, announced 17 September 2026, is — in Figure's own wording — the first demonstration of zero-shot whole-body generalization at this scope on a humanoid: three long-horizon tasks across 30 unseen homes, with no data collection or adaptation in those environments. Index pretraining accounted for most of the result: 9% success from scratch versus 56% from the pretrained policy, on full-task criteria with no partial credit. Task-specification data fell by half relative to a prior in-environment Helix 02 behaviour while deployment scope rose from one taught room to thirty untaught homes. Read through Arco's instrument, that is evidence about the property Task Execution Autonomy scores — who moves the work — and about process continuity under distribution shift, the physical analogue of 'specify once, generalize at deployment'; the axes themselves score a business, and the reading is Arco's. It is not unattended judgment. Figure's evaluation rule counts any rollout requiring a human safety intervention as a failure, so the 56% measures execution inside a gate rather than the gate's absence. The evaluated tasks carry low-to-moderate Rollback Cost; they do not move the Physical Intervention Threshold introduced after MHS. Index is a vendor-scale human-experience store, not a firm's Operational Ledger. Figure's own limit stands: general physical intelligence is not solved. Arco is not operating robots and is not entering home robotics.