Ambient Field Lens

The improvement loop

The refusal is the input.

A refusal is not a dead end. It is the first step of a repair that ends in the operator's own documentation — and the model is never retrained.

Two ways a gap is flagged

Nothing here starts with a success. The loop begins when the system cannot answer — and there are exactly two ways that happens.

1 · The coverage gate

The system measures how far the question sits from anything in your manuals. If nothing is close enough it stops and says so, rather than reaching for the nearest related passage.

2 · The operator

Or a human says the answer was useless. Both routes carry the same payload: the question, the answer, and the exact passages that were retrieved.

3 · Propose

A candidate answer is drafted against your own documents. In our run it declined to invent a procedure and instead listed what an expert would need to confirm.

4 · A human approves

An expert edits the draft and approves it in their wiki. Nothing deploys unsigned. In our run the human discarded the draft and wrote four steps of their own.

5 · Re-index

The corpus is re-versioned — a new immutable collection, the original preserved byte-for-byte, the addition embedded locally on CPU.

6 · Ask again

The same question now cites the new passage. The documentation improved. Nothing was retrained, fine-tuned or re-certified.

What we built / to test the thesis

A robot, a phone, and a manual you can edit.

We needed a machine where the failure mode is observable, so we instrumented one we own.

The Robot Ross rig: the Huenit arm, a phone on a tripod, and the controller screen — Robot Ross — the rig

The Field Lens app showing a verdict card with a retrieved manual passage — Field Lens — the card

The review console showing a feedback item approved — The wiki — approved

The whole thing

Four pieces. A machine to set up, something that looks at it, somewhere the thinking happens, and a place a human fixes the documentation.

The machine

A drawing robot we own and instrumented, so a wrong answer costs a sheet of paper instead of a part. The arm moves only on a human command, over a local serial link the cloud never reaches.

What sees it

The phone runs against the same interface the glasses expose — the frame source is an implementation detail the rest of the pipeline does not see.

Where it thinks

Retrieval over your manuals runs on CPU. Segmentation and the Glimmer readout run on GPU, inside a window capped at four single-use call tokens.

Where it is fixed

An approved passage is re-indexed into a new immutable corpus version. The only thing that changes between a refusal and an answer is the documentation you own.

Calibrated, not asserted

Where the gate sits, and why.

The coverage gate is a distance threshold. It was measured, not chosen to make a demo pass — and the measurement has a wrinkle we are not hiding.

Eight the manual covers

Questions the Robot Ross manuals genuinely answer. Their top-1 retrieval distance tops out at 0.741.

The gate

So the threshold sits at 0.760 — just above the hardest covered question. There is no setting anyone can turn to make the system say yes.

Three near-misses

Questions about the right machine that the manual does not actually answer. Two of the three fall below the gate. We are showing you that rather than letting a reviewer find it.

And two that are nowhere near

"How do I make coffee with this machine" measures 0.787 — closer to the corpus than the legitimate near-miss at 0.820. Distance measures proximity, not relevance. Both are correctly refused, but near-miss and absurd negatives have to be evaluated separately.

The receipt

The same question, before and after.

On 30 October an operator wrote the missing passage in their own wiki and approved it. These are the two citations the identical query returned, either side of that approval.

	Cited chunk	Distance	Source
before	8135308218c2c50d…	0.8199	atf-wiki/Topics/Calibration.md
after	59bb1a272e2ed09d…	0.3997	sm-faq/robot-ross-atf/fb-1793359723991-x33ca

Query, verbatim and identical both times: what to do if the paper is in the wrong position. Corpus ross-manuals-2ca5f2f152267049 (38 records) → ross-repair-ad3c9065b45ac2d3a0562011 (39). The comparison tool reports cited_chunk_changed: true and cited_text_changed: true. The original collection and its vectors are preserved unchanged; activation requires a code-reviewed pin, and rollback is one command.

It is not the heading doing the work

The new passage opens with the question as a heading, which would flatter the distance. Measured against the same corpus: answer body only scores 0.476, a neutral heading 0.451, the published version 0.400. All three clear the gate by a wide margin — the match is the answer, not the echo.

The negative control held

Adding the passage changed nothing else. "How do I make coffee" still measures 0.7867 and is still refused; the two other calibrated questions returned bit-identical distances.

Negative result / the one we are not burying

We tried to read it from inside the model. It did not transfer.

We wanted the grounding signal to come from the model's own activations rather than from retrieval. Open weights are why the attempt was possible at all — the reading is taken on the residual stream at prefill, which a hosted API cannot return.

Probe	Result	Verdict
Held-out text pilot — prefill AUROC	1.000	separated, but n=24, two topic clusters, length confound unresolved
Live visual task — supported question	−0.018	no separation
Live visual task — unsupported question	−0.015	no separation
Contrastive probe	0.575	retired
Jacobian lens vs plain	0.861 / 0.889	no advantage, retired

So it does not gate

The score is pinned to zero on the wire and the verdict never reads it. What gates is measured retrieval distance. Groundedness here is a property of the composition — retrieved passage, hash-bound evidence, fail-closed on anything missing — not a model readout.

Why we still say open weights matter

This class of check simply is not available on a hosted API, which returns tokens and never internal state. The result is negative; the capability to run the experiment at all is the point.

Stated on camera

What it does not do.

No green steps, no red blocks, no instructions to touch the machine. The verdict is always verify and there is no threshold anyone can turn to change that. The arm moves only on a human command.
No capture-time proof. The evidence card binds which frame, question, passage and masks went into an answer, each by hash. It does not bind when the photograph was taken — a photograph from yesterday passes every check on the card.
Autonomous masks are candidates, never presence and never approval. SAM did not find a paper candidate in several legitimate framings during the shoot; the gate refused, which is the correct behaviour and not a claim that the scene was read correctly.
Distance is not relevance. Two of three labelled near-misses fall below the gate.
The gap report did not always file. On camera the phone's report returns sm_unreachable; the item in this write-up was filed through the service identity. The report genuinely happened — the transport failed.

Hackathon Rule 4

What pre-existed, and what we built.

Pre-existing, used as a test bench

The Robot Ross Huenit arm, its controller and the physical rig. The ATF wiki corpus used as the operator's source documents. The Sovereign Mind wiki and review console that the repair loop writes back into.

Built during the hackathon

The Field Lens iOS app on the Meta Wearables DAT interface. The default-deny /assist contract on Modal. SAM 3.1 scene gating with millimetre geometry. The Glimmer prefill readout, its four probe protocols and the negative results that retired them. The coverage gate and its calibration. Immutable corpus versioning with a code-reviewed pin registry. And the documentation repair loop, end to end.

Clean-room repository created 26 October 2026; nothing copied from earlier private prototypes. Calibration labels were reviewed by a human and scoring ran only on frozen labels. Every negative result is in the repository under docs/verification/.

Two minutes fifty-eight, shot live.