The Question Machine — bidirectional consolidation of the convergence loop

1. The idea, in one line

Deep Thought computed the Answer (42). The machine it then built — the Earth — was a question-finder. A system that only runs forward converges on an answer and, ungrounded, freezes onto the dead 42-state. A system that also runs backward from a goal and consolidates the two passes can tell where it most needs to find something out — and that, not the answer, is what keeps it alive.

The Question Machine runs the convergence loop both ways and reads the seam between them as a question.

2. The three passes

The backward sweep (backward_costate) is the piece that was missing — the engine had the instantaneous costate (CIO_SDE.costate) but never propagated it back across a trace. It runs:

λ_T = terminal_grad(x_T)                       # the "end" — an EXTERNAL goal / observation

λ_t = ∇ₓL_t·dt + (I + Aᵀ·dt) λ_{t+1}           # the end, seen from each earlier moment

g_t = ∇_u L_t·dt + dt·Bᵀ λ_{t+1}  =  ∂H/∂u_t   # the seam at step t

g_t is exactly ∂J/∂u_t for J = Σ_t L(x_t,u_t)·dt + Φ(x_T) — machine-checked against finite differences to rel 1e-4 (test_backward_costate_matches_finite_difference). That gradient check is the correctness anchor: the "end-back" pass provably computes the true sensitivity of the whole future to each present choice.

3. The question is the seam

The question at step t is ‖∂H/∂u_t‖ = ‖g_t‖ — how far the forward choice is from what the backward objective demands. Large ‖g_t‖ = "this moment most needs resolving." The consolidation (QuestionMachine.consolidate(..., iterations>0)) is forward-backward descent u ← u − lr·g: each step re-predicts forward and re-propagates backward, driving the seam toward zero. That descent is the search for the certificate's "safe passage" (§5) — a trajectory pinned by both its start and its goal. Tested: the seam drops ~10× and the questions go quiet at the optimum (test_consolidation_drives_seam_toward_zero).

4. CAP and NAP — the seam-arbiter

A surfaced question names a channel (an action_type / capability it would be resolved through — ask_human, web_search, market_probe, …). ask() ranks questions by seam magnitude, then gates each:

• CAP — CCF CapabilityGate: the channel must be a

claimed, honesty-tracked, currently-routable capability. The positive authority — what it can do.

• NAP — AuthorityGate: if a Negative Authority Profile

denies the channel, the question is inadmissible regardless of capability. The negative authority — and it overrides.

So the machine asks the highest-leverage question it is both able (CAP) and allowed (NAP) to ask. Tested: a channel CAP allows but NAP denies is dropped, and ask() routes to the next admissible question (test_nap_overrides_cap) — denials win, exactly the CIO constraint-dominance invariant.

5. The honesty discipline — coherent ≠ correct

When the forward and backward passes agree, the seam →C3and the machine has nothing left to ask itself. But two mirrors agreeing can be jointly wrong — a smoother converging on a self-consistent, ungrounded trajectory is the higher-order collapse the C3 work warns about. Consolidation makes the trajectory coherent, not correct.

The escape is the external terminal condition. Tested: at the consolidated optimum the questions are quiet (max seam < 0.05); swapping in a terminal anchor that contradicts the coherent path reopens the seam by >10× (test_grounding_revives_question). Grounding revives the question. consolidate() is internal; ask()'s output is a hypothesis to be externally verified, never an answer.

6. Honest findings surfaced while building it

• The seam front-loads to early, high-leverage steps — not the disturbance site. Breaking

the path at step k makes the seam at k jump sharply (~85× in test), but the global top question is the earliest step, because early controls have the most authority over the terminal. So "the question localizes to the break" is false; "a break reopens a seam, most loudly where leverage is highest" is true (test_perturbation_reopens_seam). The machine asks where a fix has the most leverage, which is the optimal-control-correct thing.

• It stays inside the one loop — no sprawl. Reason+Act is the forward pass, Verify is the

backward pass, Converge is the Hamiltonian consolidation. A surfaced question is a Convergence Record (Question.to_record() → [hypothesis, evidence, confidence, source]). Nothing here is a new top-level subsystem; it is the existing loop closed from both ends.

7. What is and isn't built

Built + tested: the backward sweep, the forward-backward consolidation, the seam-as-question, CAP/NAP gating, the coherent≠correct grounding behavior. The gradient check makes the math trustworthy.

The loop is now closed (question_loop.py, 2026-06-25). QuestionDrivenLoop consolidates toward a goal belief, asks the top admissible question, resolves it through a Channel, and folds the external observation back into the belief — the Act stage made real. Channel organs: OracleChannel (grounded reality), MirrorChannel (ungrounded self-reference), CallbackChannel/HumanChannel (the human-in-the-loop — a person is just a channel), WebChannel (live web grounding), and CorroboratedChannel (≥2 independent channels fused — agreement earns confidence, disagreement is flagged "divergent" rather than averaged away; the External-Reality Rule as a channel). Tests (tests/test_question_loop.py) make the thesis executable:

The ungrounded mirror is more internally coherent (lower seam) and 3.7× more wrong — it agrees with itself precisely because it never let reality perturb it. That is grounding is the safety mechanism (§7 of the certificate) as two numbers. A NAP-denied channel leaves its dimension a permanent blind spot (LoopResult.blind) — denials keep you safe and leave you blind, surfaced not hidden.

Still toy: the channels return scalars (a goal component). The remaining work is swapping OracleChannel/CallbackChannel for organs backed by real external systems — the web tool, a document store, a market feed, an actual human prompt — and letting an LLM (e.g. the Ouro brain in the Reason slot) translate a Question into a real query and a returned answer into an Observation. The spine is built; the organs are stubs.

Provenance: built 2026-06-25 alongside the C3 certificate closure. The "end-back / beginning- forward, consolidate with CAP and NAP" framing is the design request this realizes; the Deep-Thought / question-vs-answer framing is the through-line from the collapse certificate's §7 (grounding is the safety mechanism).