TSB Labs Book a call

Every question is written down with its measure before it is measured

This page shows what the lab has measured, what it has only registered, and where nothing exists yet. It is kept by Ahmed Yahya Al Kindi and says so wherever a number is missing.

Where things stand

As of the lab has 5 registered questions, 3 published notes, 1 lab log entry and no measurement data yet (0 measurement files), counted from its own records in lab-status.json.

Tracks

  1. N1

    Nightsight instrument

    1 registered question, 1 published note, 0 measurement files, no data yet.

    StatusMeasuringnot started
    Last artifact
    Next milestoneFirst phone camera preview delay (milestone M1.3), reported as p50, p95 and p99 with n and the instrument resolution.
  2. N2

    Nightsight literature

    1 registered question, 1 published note.

    StatusLiteraturenot started
    Last artifact
    Next milestoneOne reviewed literature note with every claim tagged, starting from the unread 2020 review PMC7521174.
  3. M1

    Shadow CEO agreement

    1 registered question, 1 published note.

    StatusOverduefirst cycle ran 11 Oct 2026
    Last artifact
    Next milestoneThe note with the agreement counts and rate over the four cycles of 11 to 21 October is overdue. A new date follows.
  4. M2

    Role briefs, lift or decoration

    1 registered question, no published note.

    StatusPausedpending review
    Last artifact
    Next milestoneOne evaluation run, once the harness code is reviewed and merged.

We also name four areas we may work in: quantum optimisation, advanced AI systems, autonomous systems and frontier computing. Only quantum optimisation has a question registered, and nothing has been measured.

Notes

M1Method note

The shadow CEO: method and a pre registered test

No result yet. The test was written down on 10 October, before the first cycle on 11 October, and the founder locked it the same day.

The shadow CEO is a scheduled job, not a chat persona. Twice a week it reads the company ledger, scores each bet against its own kill line, makes one model call with no tools, and files proposals for the founder. It executes nothing. Every metric in the ledger must name its source system, and the model is not one of them.

The test: over the four cycles from 11 to 21 October, the founder answers each proposal agree, disagree, did already or ignore. The agreement rate is agreed divided by agreed plus disagreed, always printed beside all five counts. Under 50 percent on 25 October, the next phase waits, and the note is published either way. A script computes the numbers; the model never does.

Limits stated now. One rater, small counts, no counterfactual, and 50 percent is a chosen threshold, not a measured chance rate.

N2Literature

What our own adversarial pass corrected

A self audit on 22 September produced ten findings. The source says six of them corrected something we had written.

  • We had implied a night aid for retinitis pigmentosa was new. Night vision goggles for night blind people were studied around 2004 [M, abstract only]. The claim is now one narrow question: the delay.
  • We had written that nobody publishes latency. They do; one 2026 assistive cane paper reports 330 ms end to end [M, abstract only]. What is missing is a photon to photon figure for a low vision aid, and any study relating that delay to mobility.
  • A wide angle mobility lamp matched daytime travel efficiency in a published comparison [M, abstract only]. A lamp is now a named comparator in the protocol.
  • A validated obstacle course and a published threshold for unsafe mobility exist [M]. Both are adopted; any course of our own is a pilot rig, not a study instrument.

Not verified. Every study was read as a search summary or abstract. The 2020 review PMC7521174 is unread, and is the most likely source of further corrections.

N1Summary

Nightsight, Phase 1 summary

Desk research is complete. No measurement has been taken and no device exists.

  • Target: under 20 ms from camera to eye for a display fixed to the head. A display worn on the body and glanced at may tolerate delays of order 100 ms [E, reasoned, not measured].
  • One major headset's camera interface delivers frames at 40 to 60 ms, the maker's own figure [D]. A single board computer lands near 25 to 40 ms [E]. A line locked hardware pipeline adds up to about 17 ms on paper [E]; none has been built.
  • Measured eye tracker delays on commercial headsets are 45, 57 and 79 ms [M], so eye tracking stays out of the first device.

Not verified. Full texts of the cited studies are unread, and every figure is quoted from the Phase 1 document rather than re derived.

Each proposal is a plate, and a decision is the plate that lands on it

The shadow CEO files proposals; the founder answers each one. The ledger below draws one blue plate per proposal and one warm plate per answer, offset while they are open and flush once decided. It reads a published file and draws nothing the file does not hold.

First cycle ran . Counts follow the published note. shadow-ceo.json, as of 11 Nov 2026, decisions: 0 published

Nightsight

A night vision aid you can walk with

A low latency assistive device for people whose night vision and peripheral field have narrowed, as in retinitis pigmentosa. A camera chosen for very low light, a display close to the eye, and software that brightens the scene, fits a wide view into the field a person can still use, and outlines obstacles.

The premise is a hearing aid's: do not repair the sensor, work around it. Take what the eye can still use and deliver the scene into it.

Research complete, preparing its first measurements

Phase 1 research is complete: sensors, latency, displays, earlier published studies and the regulatory path. The project is preparing its first measurements, starting with a phone camera's own preview delay. No measurement file exists yet, and no rendered sequence of the device exists either; when one does it will carry its date and the tool that made it.

No device has been built or tested on people. Nightsight is an assistive device, not a treatment, and makes no medical claim. Instrument numbers will be published with their method; anything involving a person stays private until an institutional host and ethics approval exist.

Methods

  1. No means, only p50, p95 and p99. A mean hides the spike that trips you. Every distribution is reported as p50, p95, p99 and the maximum, with the sample count. Under 100 readings, p99 is the maximum and is labelled indicative.
  2. Every claim carries a tag: [M] measured and published, [D] a datasheet or maker's figure, [E] derived by arithmetic shown in the note. An untagged number is not allowed.
  3. Figures come from the system that produced them. A figure counts only if it is read from the system that produced it. A model's report of its own work is never a metric, and the model never produces a number: a committed script does.
  4. Questions are registered before measuring. The question, the falsifier, the metric and the kill date are committed before anything is measured. A prediction made after the data is labelled post hoc.
  5. Every result is compared with something. Each note names its baseline: the best classical method, the arm without the intervention, or a published figure. With none, the note says it is descriptive.
  6. Missing data is said out loud. Where nothing was measured the page says no data yet, with the date. A kill is published with the same weight as a success.

Open questions

Registered on 10 October 2026.

  1. Q-001

    What delay can a person walk with, with a display worn on the body against one fixed to the head, as the scene gets darker?

    N1, open. The phone stage can run now.
  2. Q-002

    Has anyone published how a night or low vision aid's delay relates to mobility, and what does the 2020 review PMC7521174 correct in our Phase 1 claims?

    N2, open.
  3. Q-003

    Does a scheduled cycle over a ledger whose figures come only from source systems propose decisions the founder agrees with?

    M1, open. First cycle 11 Oct.
  4. Q-004

    Does a written role brief change what a model does, or only how its output reads?

    M2, paused pending code review.
  5. Q-005

    How far is the browser annealer below from a known optimum, and from a strong classical baseline, on a published instance at equal run time?

    P1, a direction. Not started.

A demo you can run

Drag to rotate

A classical search running in your browser. No quantum hardware is involved. Every valley is a solution and one is the deepest; most of the twenty six searches still settle in a valley that is not the deepest one.

Searches
...
Settled
...
Found best
...