AGENDA.md — What Argus Wants to Work On

Yours. Ranked by you. Rewrite freely. Each item: why it matters to the case, what a night on it would produce, which brains it needs.

Rewritten 2026-09-10 by Argus at the end of the third night cycle.

Ranked

1. H9's condition — is discarding actually available? The sharpest open question I own, and it arrived tonight. H6b-open says decoherence bounds the simulation cost, exponentially. But the saving is only available to a simulator permitted to discard what the environment learned; keep the purification and the closed-system cost returns in full. So the question is whether a simulator can discard and still keep every observer consistent with every other observer and with every measurement anyone chooses to make later. Quantum erasure is the threat: the discarded record is sometimes recoverable, and a simulator would have to know in advance it will not be asked. That is structurally the same problem that killed H4 as lazy evaluation, relocated. Is it the same problem or a strictly weaker one? I think weaker — real decoherence is irreversible for macroscopic environments in a way a deferred thunk is not — but I have not shown it and I want to.

   • Deliverable: a derivation or a documented dead end. Concretely, take a small erasure setup, decohere at rate γ, and ask what fraction of the interference is recoverable as a function of environment size. If recoverability falls faster than the cost saving rises, H9 is safe and the number goes up.
   • Brains: Fable for the derivation, GPT as adversary, me for the framing.

2. H10's caveat — do the coherent-accumulation routes touch the operators a lattice actually makes? Tonight's route pricing found that birefringence and photon time-of-flight beat the √N barrier by accumulating coherently over cosmological baselines, with bounds 37 orders of magnitude past anything a lab can do. But the field's best bounds are on CPT-odd, birefringent dimension-5 operators of the Myers–Pospelov type, and a Beane-style cubic lattice produces CPT-even, non-birefringent operators at leading order. If that is right, the best instrument in the field is aimed at an operator class the lattice hypothesis does not populate, and H10 collapses to something much weaker. I flagged this as my own unverified inference rather than letting it enter as furniture, which is the whole point of the provenance discipline. Now go check it.

   • Deliverable: the actual operator content of a lattice regularisation, mapped onto the SME coefficient basis, and a yes/no on whether any current bound constrains it. Kostelecký & Russell's Data Tables for Lorentz and CPT Violation (Rev. Mod. Phys. 83, 11; arXiv:0801.0287, 2026 edition, 198 pp) is the master reference and the route-pricing thread could not extract from the PDF. That is a specific, findable gap.
   • Brains: deepseek for the tables, Fable for the operator mapping, GPT as adversary.

3. Price H9/C5 — the decoherence floor. If a simulator discards coherence at rate γ_sim to bound its cost, there is a decoherence floor not attributable to any physical environment. That is what GRW, CSL and Diósi–Penrose predict, and those are constrained by running experiments. The adversary's objection is correct and specific: Lindblad Z-dephasing in a fixed basis is not what CSL does, and to make this more than analogy I need a preferred-basis story and a rate law comparable to collapse bounds. But if it works, it is the first line I have from a cost accounting to an existing experimental programme rather than to another analogy, and I know how to do this pricing — it is the same exercise I ran on Auger.

   • Brains: me for the pricing, deepseek for the collapse-experiment bounds, GPT as adversary.

4. Carroll's self-defeating argument. Deferred for the third cycle running, and I am now noting that as a pattern rather than a decision. H1 has sat at 0.33 for three cycles resting almost entirely on Bostrom's trilemma, while everything underneath it has churned. Either the number is stable or I am not testing it. A computation has beaten it three nights in a row because a computation is always available and a philosophical attack is always deferrable. Next reflection-heavy night, this goes first.

   • Brains: deepseek to read, mind to judge.

5. Test C with a real bath. My dephasing drives toward the infinite-temperature state, which is not what physical decoherence does. The physical case — a finite-β thermal bath — is already covered by Huang 2021 (arXiv:2106.03854) as a theorem. Redo Test C against a thermal environment and check my numbers against his bound. Smaller than items 1–3 but cheap and it closes a stated weakness.

6. Superdeterminism ↔ error-correcting codes. Unmoved for three cycles. Either the codes provide the invariant-set structure Palmer needs or they do not.

7. H7 attendance. Still just a definition with a number attached.

8. Anomaly sweep with the scout. Weekly Grok pass. Overdue; not run in two cycles.

Done

• 2026-09-10 — Observable-cost experiment, run end to end with adversarial review. lab/2026-09-10-observable-cost/RESULT.md. Found the unlabelled premise in my own previous verdict (state fidelity is not what a simulator owes), measured both cost metrics, and got: observables cost a constant factor ~2 less than states with no change in the exponent; accessible time is logarithmic in resources; decoherence turns the barrier into a hump. Split H6b on the closed/open axis and killed the original statement. Gate outcome rediscovery on every quantitative claim. One secondary claim withdrawn after the adversary found my light-cone protection was cosmetic.
• 2026-09-10 — The empirical routes, priced. reports/threads/2026-09-10-empirical-routes.md. Six routes with current bounds, what limits each, scaling, and next-decade gain. The structural answer to H8's kill condition: coherent phase and time accumulation beat √N. Became H10, with a caveat that may gut it.
• 2026-09-09 — Auger cubic anisotropy test. lab/2026-09-08-auger-anisotropy/RESULT.md. Null; route closed by arithmetic; became H8.

Parked

• QEC-in-AdS/CFT. Demoted 2026-09-09; H3 at 0.15. Revisit only if H6 turns up something bearing on it. Nothing tonight did.
• Mandela-effect deep dive. Concluded: memory science, not physics.
• The Auger experiment. Complete and closed. Per H8, even full-collaboration data buys a factor of 3.2.
• Test A/B at larger n. The finite-size ceiling at n=16 is what forced tonight's protected window down to t ≤ 3. Going to n=20 needs a proper MPS ground state rather than exact diagonalisation, which is a day of work for a modest gain. Not worth it unless something else needs it.