Where we are · updated Oct 8, 2026
Two things run side by side to the end of the year. The first is the degree experiment (experiment C), which the December-result roadmap still binds: the registration committed by 2026-10-18 with the $1.14 repair written in, the repaired built models verified (a failure stops C there), one free-arm run at full size (a gate or floor miss stops C there), and the second release of about $150 to $172 asked for only on those figures, a pass being necessary and not sufficient. The project's own stated prior is that the free model most likely returns no verdict or fails its gate, so the likeliest ending for C is an early stop written up as instrument research, which is a registered outcome and not a schedule failure. The second is the new question, in task order with no step on a date: the two-sided-question text and the table of felt features through Gate A; the battery's first two entries on experiment D's pipeline by API, about $20 together, as reference readings of what the cheaper routes produce; then the construction line, small systems whose positions on the axes are set from outside, which needs a state-carrying mechanism neither toy pipeline has, and is where every test that can discriminate lives; last, and designed before it is run, the human study that calibrates the detector. By 2026-12-21, realistically: C run to a registered outcome, the new question in the spec, the table and the first two battery entries registered and run, the relabelling of experiment A in the paper draft, and the packet of three changes sent to the book. Not this year: a degree reading of a free model, a discriminating depth reading, or an observer-side measurement. Loss conditions are written for every new line. Claims are whatever the registered results support: a position on the axes, the route named, the degree measured, never a verdict. Never a conscious machine.
Ruled 2026-09-20: patching the A3 checkpoints would build a tool for a model it cannot answer on (no matched-role comparator, target never localized). The A3 closure no longer waits on it.
The plain-language companion last touched 2026-07-01 predates every registered result since. It carries the founding wager once that lands.
Nothing new. Each is ruled or already identified; none is done.
Whether the port of the toy scripts into one procedure is faithful, whether the outcome rule and the tripwire do what the rulings say, and whether the trainer ran the loop it claims on the rented machine.
Waiting on John
Two sides, ruled 2026-10-07 (spec text pending Gate A): what is the smallest system whose presence in interaction cannot be produced by a cheaper route (lookup, imitation of a corpus about minds, routing of who is speaking), and at what point do observers' detections of presence begin to track that structure rather than fluency? The earlier question, whether the smallest self-centred act of integration (the floor) is present in a system, stays in the spec as the founding wager and stops being a measurement target.
Take the consciousness account developed at Sentient Horizons and The Calibration Problem as a specification, build and measure against it, and let every claim carry a stated loss condition.
Sufficiently deep self-indexed temporal integration is experience, seen from within. A system whose architecture binds past, present and anticipated state into one act centred on itself is a bearer of experience to the degree that it does so. The functional agent with all of that architecture and no one home is declined: it can be imagined, it predicts nothing, licenses no experiment and sorts no system, so it pays no rent.
Its status A bet that cannot lose, so by the project's own rule it is not a hypothesis. It enters as a presupposition, the way induction presupposes that nature is uniform: the condition under which structural evidence is evidence about experience. Reject it and every result here still stands as a claim about which structures are present; it just stops bearing on experience.
Why still never a verdict The wager makes the evidence evidential without making it conclusive. Two limits remain, and neither is the zombie: the identity is a bet that correlational evidence cannot carry, and the instruments can be wrong about structure before any question of experience arises. That is what mutual opacity names here.
The Calibration Problem, ch. 5 and Appendix A; spec §The Founding Wager (proposed 2026-09-20, Gate A pending)
Turn the floor claim (consciousness is self-indexed temporal integration, adjudicated by the removal test) into pre-registered, falsifiable instruments. Achievable regardless of what the instruments find, and where most of the durable value lives: the field has position papers; this repo builds tests that can fail.
Where it stands A fully registered removal test ran on 2026-07-18 and its own pre-committed control voided the headline signature when it appeared. The discipline (register, red-team, lock, run, report verbatim) has since carried Stage 3 and the whole MVM-0a programme, with an outside-review protocol in force since 2026-09-19.
Run the instruments on models that exist. Since the refounding of 2026-10-07 the target is the axes an observer perceives (does reversal cost the system anything; did its history change it or only its record; is anything at stake), not the floor. On a frontier model reached through its interface, the trained weights and the transcript are the cheaper routes by name, so what these models give is a reference profile of what lookup, imitation and routing produce. The output is bounded at 'non-zero on the gradient' or 'not testable here', never a verdict.
Where it stands Experiment 1 (8B instruction-tuned model): a self-tracking structure was located and removed, and the registered control could not tell a router from a centre (relabelled 2026-10-07); the self-report was never subtracted. Experiment 3: frontier models mostly mask rather than capitulate under pressure (nine true capitulations in 540 ladders). Both are registered results. Next on this goal: experiment D's transcript-replacement control and the ownership swap, about $20 together, which read which cheaper route produced the retained-independence result.
Construct small systems whose structure is known, because that is the only place a reading can discriminate: two systems built alike except for one route, read on the same indicator. MVM-0a, the constructed self-index, was the first step. The degree experiment (C) is the second: built models at both ends of a scale and a free model read between them. The construction line of the two-sided question is the third: systems whose positions on the axes are set from outside, the smallest of which that passes the battery is the minimum viable mind, within its construction family.
Where it stands MVM-0a exists as three 30-million-parameter checkpoints; the ownership input is load-bearing on three seeds, and Amendment A3 closed 2026-09-20 as not testable because its matched contrast could never be run. The degree experiment's code is frozen and reproduces the toy record; at 10 million parameters two of three built models switched their built-in route off, so a $1.14 repair is written into the registration and verified before any full-size run. The construction line needs a state-carrying mechanism neither pipeline has; nothing of it exists yet beyond the battery draft's entries 3 to 8.
Turns 'not testable (localization)' into either a localised result or a registered null with both instruments. Amendment A3 §3.2 requires probe and patching to agree; patching has never run. Its target and null are written before it runs.
Whether the fitted read's null on the model's own index is specific to ownership or a property of the read: a registered matched control that was never run.
Whether the one near-miss (three standard deviations at the other agent's revision value, seed 2, the highest testable position on all three checkpoints) is signal or noise under a standardised fit.
The bridge from small constructed models to the systems public discourse is actually about. Design first.
What the smallest system is whose presence in interaction cannot be produced by lookup, imitation or routing, and at what point observers' detections track that structure rather than fluency. Replaces the floor as the measurement target. Loss conditions written in the proposal: an indicator producible by a cheaper route kills the battery; axis positions that cannot be set independently kill the construction line; observers who never improve kill the calibration line; an empty battery kills the question.