docs/research/2026-07-22-cross-domain-incremental-backlog.md

Cross-domain incremental backlog — organized by the USER problem it solves

Date: 2026-07-22 · Source: the cross-domain patent landscape + fresh sources we already pulled (grounding ledger). Rule for this list: every item must make the assistant more trustworthy, cheaper, or better at the user's task — if it doesn't map to a user outcome, it's not here.

The product in one line

A personal AI assistant that runs on your own machine, is honest about what it knows, checks its own work against reality, and improves at your problems — without cloud cost or data leaving your box.


Problem— "It bluffs or gives up on hard tasks" (make it actually solve, cheaply)

Update User payoff Cross-domain source Effort
Stop-on-stall — halt when N turns make no real progress (+ a cheap state-hash to spot a loop) instead of a fixed guess-limit doesn't waste your time/compute grinding a dead end; doesn't loop forever ECC decoders US6518892B2 / US8301987B2 S
Pass-terminates-immediately — run the cheapest check on every candidate; a pass ends it, only a fail escalates fastest/cheapest path to a verified answer Medical hierarchical assay US6013436A S
De-escalate when it gets easy — drop back to the cheap local model when the hard part is over keeps the cloud bill down over a long session Intervoice call-routing US7254641B2 S
Each escalation must actually help or stop — every "phone a friend" must measurably improve the answer, else halt no runaway spend on steps that aren't converging Iterative Learning Control (laser/disk servos) M

Problem— "I can't trust it — it's confident even when wrong" (honesty)

Update User payoff Source Effort
Calibrated "I don't know" — a real abstain boundary + the +1/0/−λ training reward it tells you when it's unsure instead of bluffing (frontier models basically never do this) Applied Materials US9715723B2 + Reinforced Hesitation 2511.11500 M (V1/V2)
Judge from the whole answer, not where it stopped — aggregate the model's steps into one honest confidence score the trust signal it shows you is actually calibrated HP US20120243734A1 S
Look inside to catch a bluff — the white-box honesty probe catches confident-but-ungrounded answers a cloud API physically cannot our probe work (verified 0.92 AUROC @ 7B) M (V0 done)

Problem— "It repeats mistakes / doesn't learn from me"

Update User payoff Source Effort
Remember your recurring failures — cache failure modes by task-signature and avoid them before proposing stops making the same mistake twice on your kind of work Marvell disk-drive repeatable-error servo US8094405B1 M
Learn only from verified wins — keep the exec-verified solutions as training fuel it gets better at your problems, not generic benchmarks our VTD flywheel M
Check a fix actually worked before moving on — measure the metric recovered, then refine its playbook it doesn't declare victory on a fix that didn't land Bowe Bell+Howell remediation US20100094676A1 M

Problem— "It doesn't know when its info is stale" (indefinite low-cost operation)

Update User payoff Source Effort
Re-check each fact on its own clock — frequent for market data, rare for stable facts (the EOQ interval) stays current where it matters without re-verifying everything (the cost lever) ops-research replenishment / M2 M
Leash the cheap self-check to ground truth — trust the fast check only while it's periodically re-fit against real verification the cheap path stays honest over a long run evolutionary-computation surrogate US8131656B2 S–M
Anytime-valid stop certificate — a statistical stop-test that stays valid even when it runs indefinitely it can work on your behalf for long stretches without drifting into nonsense SEA arXiv:2607.00871 (this week) M

Principles the checks confirmed (keep these, don't drift)

  • Verify against ground truth, never a learned copy of a human's judgment. IBM US12135927B2 patents the copy-a-human gate; we deliberately do the opposite — a real test decides. That's the novelty and the freedom-to-operate safety.
  • The moat is the combination, not any one trick. Every mechanism above is fielded prior art in some other industry; nobody has wired them onto an LLM that checks itself against reality. That's what we own.

P1 → P2 quick wins: stop-on-stall, pass-terminates, de-escalate, judge-from-whole-answer. These make the existing assistant cheaper and more honest this week. Then the EOQ fact-clocks (the thing that makes "runs indefinitely at low cost" real).

Logged as issues (2026-07-22, kriskin lane)

#2865 stop-on-stall+loop-hash · #2866 cascade contract (pass-terminates + de-escalate) · #2867 contractive escalation · #2868 calibrated whole-answer confidence · #2869 failure-mode cache · #2870 verify-then-refine · #2871 surrogate-leash · #2872 M1 enforced gate · #2873 probe ternary-survival (V3) · #2874 basin-determinism round-2 audit. Existing homes kept: #2853 EOQ clocks, #2854 Kreiss/contraction math, #2855 water-filling, #2856 anytime-valid + M1/M6 core. Strays assigned: #2858 (probe-gated spiral), #2860/#2861 (novelty audits).