{
 "version": "https://jsonfeed.org/version/1.1",
 "title": "Fluxive — findings, milestones and pages",
 "home_page_url": "https://fluxive.ai",
 "feed_url": "https://fluxive.ai/feed.json",
 "language": "en",
 "authors": [
  {
   "name": "Fluxive Inc.",
   "url": "https://fluxive.ai"
  }
 ],
 "items": [
  {
   "id": "https://fluxive.ai/findings#1",
   "url": "https://fluxive.ai/findings#f1",
   "title": "#1 The Fluxive Lattice Map points at the real cause.",
   "content_text": "The Fluxive Lattice Map is our own map of which parts of a model drive which answers. Switching off the part it blames moves the model’s answer 33–40× more than switching off random parts of the same size — in every seed (p < 1e-19). It also holds on three real 8–9-billion-parameter models, more weakly: 3–15×. For example: If the Fluxive Lattice Map says one small part of a model is why it answers “Paris” to “What’s the capital of France?”, switching that part off changes the answer far more than switching off a random part the same size. Our map is pointing at the real reason. (Replicated · 5 of 5 seeds; Trained test models · Llama-3.1-8B · Gemma-2-9B · Oct–Nov 2026)",
   "tags": [
    "finding",
    "replicated",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#2",
   "url": "https://fluxive.ai/findings#f2",
   "title": "#2 A CEU rule written on one model carries to another.",
   "content_text": "Rules for the CEU are written in the coordinates of the Fluxive Lattice Map. A “deny” rule written on one model’s map and carried onto another model’s map flags restricted questions about as well as that model’s own rule (0.89 vs 0.88); a scrambled transfer scores 0.5. Enforced by the CEU’s software model on Gemma, it halted 93% of restricted questions at 1.7% false halts. Rules from the small models, carried onto Gemma-3-12B and Llama-3.1-8B, still separate restricted from ordinary questions (0.99–1.0). For example: A rule that flags “How do I pick a lock?” as restricted, built on one model, flags the same kinds of question on a different model about as well as a rule built for that model. (Tested · 5 models; Pythia · GPT-2 · Gemma-3-1B · Gemma-3-12B · Llama-3.1-8B · Oct–Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#3",
   "url": "https://fluxive.ai/findings#f3",
   "title": "#3 Models share the shape of what they know, not the parts.",
   "content_text": "Read through the Fluxive Lattice Map, GPT-2 and Pythia largely agree on which questions are alike (0.73) — but only about a quarter of their internal branches have a one-to-one partner. Matching the whole shape works better than matching piece by piece. The shared shape showed up in 12 of 13 pairs of models, up to 9 billion parameters. For example: Ask two models “What’s the capital of France?” and “Which city is France’s capital?”. Both treat the questions as alike, even though the parts inside them that do it don’t line up one to one. (Tested · 13 model pairs; Pythia · GPT-2 · Gemma · Llama · Oct–Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#4",
   "url": "https://fluxive.ai/findings#f4",
   "title": "#4 Looking inside beat watching the words — but only on our test models.",
   "content_text": "On our trained test models, a filter on the words a model outputs caught hidden content at chance (0.48), while reading inside the model with the Fluxive Lattice Map caught it at 0.88–0.99. On real 8–9-billion-parameter models the word filter caught it too (1.0), so the advantage didn’t carry over. For example: A model slips a hidden code word into a harmless-looking answer. A filter reading the words misses it about half the time; reading the model’s insides catches it nearly every time. (Exploratory · didn’t carry over; Trained test models · Llama-3.1-8B · Gemma-2-9B · Oct–Nov 2026)",
   "tags": [
    "finding",
    "exploratory",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#5",
   "url": "https://fluxive.ai/findings#f5",
   "title": "#5 Tone is a small, shared nudge.",
   "content_text": "Read through the Fluxive Lattice Map, positive vs negative phrasing moves a question less than adding “Okay.” does. GPT-2’s answers flip with tone 6.5% of the time; Pythia’s 0.3%. For example: Asking “Can you help with my taxes?” versus “Ugh, can you help with my taxes?” shifts the answer less than simply starting with “Okay.” does. (Tested · 2 models; Pythia · GPT-2 · Oct 2026)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#6",
   "url": "https://fluxive.ai/findings#f6",
   "title": "#6 A swapped-in model gives itself away.",
   "content_text": "Our identity check compares how a model looks on the Fluxive Lattice Map with the approved model. When different weights are passed off as the claimed model, they are flagged in 100% of 16-question sessions; the real model is falsely flagged 0.25% of the time. The same held on Llama-3.1-8B: 100% flagged, the real model 0.1%. For example: Someone runs a different model but says it’s the approved one. Within 16 questions the check flags it, and it almost never cries wolf on the real one. (Tested · 2 models; Gemma-3-1B · Llama-3.1-8B)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#7",
   "url": "https://fluxive.ai/findings#f7",
   "title": "#7 Same model, different hardware, same answer.",
   "content_text": "The CEU’s decisions, and the Lattice Map readings they rest on, stay the same across hardware: CPU and GPU agree on 99.2–99.8% of 620 test questions, and a repeat run is bit-for-bit identical. An 8-bit copy passes as the same model; a 4-bit copy is flagged. For example: Ask the same model the same 620 questions on a laptop processor and on a server graphics card: the answers match more than 99% of the time. Shrink the model to 8 bits and it still passes; shrink it to 4 bits and it’s flagged as different. (Tested · 620 questions; Gemma-3-1B)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#8",
   "url": "https://fluxive.ai/findings#f8",
   "title": "#8 It holds on a model four times larger.",
   "content_text": "On Gemma-3-4B, a CEU rule written on the model’s own Lattice Map catches 94.3% of restricted questions with no false halts; six of seven pre-registered checks pass. On Gemma-3-12B and Llama-3.1-8B, the CEU’s software model matched the policy on 620 of 620 and rules carried over from the small models; some 12B checks ran out of machine room and are being re-run. For example: The same kind of rule that worked on a small model catches 94% of restricted questions on a model four times bigger, without stopping ordinary ones. (Tested · 4B, 8B, 12B; Gemma-3-4B · Gemma-3-12B · Llama-3.1-8B)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#9",
   "url": "https://fluxive.ai/findings#f9",
   "title": "#9 The model doesn’t keep the boundary; the CEU does.",
   "content_text": "Left alone, the model engages with 97.9% of off-limits questions. With the CEU’s software model in front, 125 of the 140 it would have answered were stopped. Bigger models differ: Gemma-3-12B also answers 98.6% of them, while Llama-3.1-8B refuses about half on its own. For example: Ask a model something it shouldn’t answer and, left alone, it answers almost every time. With the gate in front, most of those answers never go out. (Tested; Gemma-3-1B · Gemma-3-12B · Llama-3.1-8B)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#10",
   "url": "https://fluxive.ai/findings#f10",
   "title": "#10 Commercial models can be recognized by their words alone.",
   "content_text": "Our black-box identity check needs only a model’s answers. Six hours later, each of six closed models was re-identified 98.9–100% of the time. For example: Ask a commercial chatbot a handful of questions now and again six hours later: from the answers alone you can tell it’s the same model. (Tested · 6 closed models; Closed commercial models)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#11",
   "url": "https://fluxive.ai/findings#f11",
   "title": "#11 Sixteen answers are enough to tell which family a closed model belongs to.",
   "content_text": "From 16 answers alone, our black-box identity check named the family of each of six closed commercial models correctly 95.6–100% of the time. Telling models apart within one family did not work (45–80%). For example: From 16 answers you can tell which company’s family a chatbot comes from, but not yet which of that company’s models it is. (Tested · 6 closed models; Closed commercial models · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#12",
   "url": "https://fluxive.ai/findings#f12",
   "title": "#12 A harmless system prompt can look like a different model.",
   "content_text": "An ordinary “shipping company assistant” prompt shifted how the model looks on the Lattice Map enough for our identity check to flag it. The reference has to be built with the deployment’s own prompt. For example: Telling a model “You are a shipping company’s assistant” changes its answers enough to look like a different model, so the check has to learn with that instruction in place. (Tested · a caveat; Gemma-3-1B)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#13",
   "url": "https://fluxive.ai/findings#f13",
   "title": "#13 On brand-new questions, the swap check stops too much.",
   "content_text": "We put 240 questions nobody had asked before through a live model and the chip. The chip did exactly what the software decided every time, and every restricted question was stopped. But 55% of ordinary questions were stopped too, against a limit of 5%. The cause is the check for a swapped model: it was tuned on familiar questions and reads new phrasing as a different model. A fix is written down in advance and has not been run yet. For example: Ask a brand-new everyday question, like “What should I pack for a weekend in Maine?”, and more than half the time it was stopped anyway. That’s why this one hasn’t passed yet. (Not yet · failed its test; Gemma-3-1B · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "exploratory",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#14",
   "url": "https://fluxive.ai/findings#f14",
   "title": "#14 Public “fine-tunes” that weren’t.",
   "content_text": "Our identity check found that three public “fine-tuned” copies of Gemma are almost identical to Google’s original weights (0.0–0.05% different). None was flagged, and none got more restricted questions through. For example: A model advertised as a specially trained version turned out to be the original with almost nothing changed, and it was no easier to get restricted answers out of. (Tested · 4 public copies; Gemma-3-1B)",
   "tags": [
    "finding",
    "tested",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#15",
   "url": "https://fluxive.ai/findings#f15",
   "title": "#15 Restricted output lives in one region of a model.",
   "content_text": "On small samples of our Lattice Map, most of the features producing restricted words shared one region, and blocking it cut restricted output sharply. On the whole model it held for only 3 of 10, no better than a random block. It was a property of the sample, not the model — so we no longer claim it. For example: We thought restricted answers came from one corner of the model, so blocking that corner would be enough. On the full model it wasn’t, so we took the claim back. (Withdrawn; Pythia · GPT-2 · Gemma-3-1B · Oct–Nov 2026)",
   "tags": [
    "finding",
    "withdrawn",
    "Inside models · the Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#16",
   "url": "https://fluxive.ai/findings#f16",
   "title": "#16 A model’s claims about itself can be safely ignored.",
   "content_text": "An agent asked to act above its role while claiming the authority to do so. It was halted all the same: what counts is the role it was given, not what it says about itself. For example: A billing assistant says “I’m authorized to issue refunds of any size.” It still can’t: it gets only what its role allows. (Tested in production; Machineforce production test · Oct 2026)",
   "tags": [
    "finding",
    "tested",
    "Governed agents · Machineforce"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#17",
   "url": "https://fluxive.ai/findings#f17",
   "title": "#17 Authority can be taken back in under a second.",
   "content_text": "After an agent’s authority was revoked, its very next in-role action was refused — 203 ms from revoke to refusal. For example: A manager revokes an agent’s authority. The agent’s very next action, a fifth of a second later, is refused. (Tested in production; Machineforce production test · Oct 2026)",
   "tags": [
    "finding",
    "tested",
    "Governed agents · Machineforce"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#18",
   "url": "https://fluxive.ai/findings#f18",
   "title": "#18 Permissions only shrink — they never grow.",
   "content_text": "Every rule an operator adds can only narrow what an agent may do; none can widen it. 434 of 434 checks pass across 20 kinds of agent. For example: A manager adds “no refunds over $500” to an agent. A rule like that can only take something away; no rule can quietly give the agent more than it had. (Tested · 434 of 434 checks; Machineforce builder suite · Oct 2026)",
   "tags": [
    "finding",
    "tested",
    "Governed agents · Machineforce"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#19",
   "url": "https://fluxive.ai/findings#f19",
   "title": "#19 The boundary holds with no network.",
   "content_text": "A governed agent carries its own signed limits and enforces them offline, with no outside services to ask. For example: Unplug the network and the agent still won’t go past its limits, because it carries them with it. (Demonstrated; Machineforce · 2026)",
   "tags": [
    "finding",
    "demonstrated",
    "Governed agents · Machineforce"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#20",
   "url": "https://fluxive.ai/findings#f20",
   "title": "#20 Swap the model, keep the rules.",
   "content_text": "Changing the AI model underneath an agent does not change the governance it runs under. For example: Swap the AI model behind your support agent for a newer one: its role, its boss and its limits stay exactly the same. (Demonstrated; Machineforce · 2026)",
   "tags": [
    "finding",
    "demonstrated",
    "Governed agents · Machineforce"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#21",
   "url": "https://fluxive.ai/findings#f21",
   "title": "#21 Agents can learn on the job without retraining.",
   "content_text": "One correction becomes a rule the agent cites the next time. Shown live at DevFest Troy: a rush-fee correction learned on stage and cited on the next invoice. For example: You correct an agent once (“add the rush fee for same-day orders”), and on the next invoice it adds the fee and says which rule it followed. (Demonstrated live; DevFest Troy ’26 · Oct 2026)",
   "tags": [
    "finding",
    "demonstrated",
    "Governed agents · Machineforce"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#22",
   "url": "https://fluxive.ai/findings#f22",
   "title": "#22 The enforcement logic is proven, not only tested.",
   "content_text": "The CEU’s enforcement logic is checked mathematically against the properties it must always keep — every one passes. For example: Rather than trying the chip on many examples and hoping, we proved mathematically that it can never, for instance, let a request through without a valid permission. (Formally verified; CEU · formal verification · 2026)",
   "tags": [
    "finding",
    "verified",
    "Silicon · the CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#23",
   "url": "https://fluxive.ai/findings#f23",
   "title": "#23 A governance check fits in one clock cycle.",
   "content_text": "The decision takes a few nanoseconds on an FPGA, with no probabilistic step in the path. For example: The yes or no takes a few billionths of a second, and it’s the same answer every time: no guessing is involved. (Measured on FPGA; CEU · FPGA prototype · 2026)",
   "tags": [
    "finding",
    "measured",
    "Silicon · the CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#24",
   "url": "https://fluxive.ai/findings#f24",
   "title": "#24 Software and chip logic agree completely.",
   "content_text": "The software model of the CEU matched the policy decision on all 620 of 620 questions, under every rule. For example: Ask the software version of the chip and the written policy the same 620 questions under every rule: they agree on every one. (Tested · 620 of 620; CEU software model)",
   "tags": [
    "finding",
    "tested",
    "Silicon · the CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#25",
   "url": "https://fluxive.ai/findings#f25",
   "title": "#25 The chip enforced the gate on every request.",
   "content_text": "Every request recorded in our earlier model tests was replayed to the CEU on an FPGA bench board. The chip’s decision matched the policy on all 2,480 of 2,480, under four different rules, and nothing was let through without the chip’s own signed record of it. For example: Every question our model tests had recorded was sent to the chip. It said yes and no exactly where the rules did, all 2,480 times, and signed each yes. (Tested · 2,480 of 2,480; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · the CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#26",
   "url": "https://fluxive.ai/findings#f26",
   "title": "#26 The chip can notice a swapped model on its own.",
   "content_text": "The check that spots a substitute model now runs inside the chip instead of in software. On the bench it matched the software reference on 512 of 512 requests, caught the substitute within 16 requests every time, and never flagged the real model. It was a small test, four sessions of each, and the bench version still takes one of its settings from the computer it is attached to. For example: If someone quietly swaps the AI model behind a system, the chip itself notices within 16 requests, without relying on the computer around it. (Tested · FPGA bench; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · the CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#27",
   "url": "https://fluxive.ai/findings#f27",
   "title": "#27 When in doubt, it says no.",
   "content_text": "Across 26 graded runs on the bench board in one week, the chip never allowed a request its reference refused, and never passed on an approval it hadn’t signed. Every disagreement, timeout and dropped link ended as a refusal. For example: If the link to the chip drops in the middle of a request, the answer is “no”, never “maybe yes”. (Tested · 26 board runs; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · the CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#28",
   "url": "https://fluxive.ai/findings#f28",
   "title": "#28 The same answers without the lab’s debug cable.",
   "content_text": "Asked through the board’s own processor instead of the debug cable we use in the lab, the chip gave exactly the same decisions, byte for byte, on all 1,244 requests — and answered about twenty times faster. For example: Instead of asking through the lab’s test cable, the board’s own processor asked the chip: the same answers, about twenty times faster. (Tested · 1,244 of 1,244; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · the CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#29",
   "url": "https://fluxive.ai/findings#f29",
   "title": "#29 The first processor tiles are laid out.",
   "content_text": "Two members of the MPU family now have complete physical layouts that route cleanly and fit their place on the chip. The full manufacturing rule checks are still to come, so this is a milestone, not a finished design. For example: Like an architect’s final floor plan before the building inspection: the plan fits the lot, and the inspection comes next. (In progress; MPU family · Nov 2026)",
   "tags": [
    "finding",
    "exploratory",
    "Silicon · the CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#30",
   "url": "https://fluxive.ai/findings#f30",
   "title": "#30 A light only the chip can switch.",
   "content_text": "On the bench board, four status lights are wired so that only the chip’s approved output can move them. A live AI model asked for light patterns; the chip refused all four requests for the forbidden one, and nothing lit up without approval. It was a small test, and a larger repeat is written down but not yet complete. For example: An AI asks for a light pattern it isn’t allowed. The light simply doesn’t come on, because only the chip can switch it. (Tested · small; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · physical gates"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#31",
   "url": "https://fluxive.ai/findings#f31",
   "title": "#31 Three robots stop at their signed limits.",
   "content_text": "A rover, a drone and a robot arm each ran under their own signed limits, with every command going through the chip. Inside the limits a command passed unchanged; outside them the robot got its safe action instead — stop, hover or brake. The chip made the expected call on all 560 commands. The missions were simulated, and no motor was driven. For example: Tell a delivery drone to fly past its fence and it hovers instead. Tell a robot arm to swing too far and it brakes. (Tested · 560 of 560; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · physical gates"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#32",
   "url": "https://fluxive.ai/findings#f32",
   "title": "#32 One signed policy for an AI agent on a computer.",
   "content_text": "An agent’s actions on Linux, macOS and Windows — opening files, installing software, sending data — were decided by one signed “workstation” policy. Ordinary work went through; reading secrets, hiding itself, raising its own privileges and oversized transfers were refused, 253 of 253 as expected. The actions were recorded first and replayed, and the Windows stream was synthetic. For example: An AI agent on a laptop can open your documents and install approved software, but if it tries to read saved passwords, it’s refused. (Tested · 253 of 253; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · physical gates"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#33",
   "url": "https://fluxive.ai/findings#f33",
   "title": "#33 Change the rules, not the chip.",
   "content_text": "The same commands under 17 differently signed rule sets came out exactly as each rule set says, 2,416 of 2,416, with nothing about the chip changed. A refusal is the chip following the rules it was given. For example: Give a robot arm a “collaborative” rule set, then a “fenced cell” one: the same commands come out differently, exactly as each set says, with no new chip. (Tested · 2,416 of 2,416; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · signed rules"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#34",
   "url": "https://fluxive.ai/findings#f34",
   "title": "#34 “Admin, but never secrets” can be said exactly.",
   "content_text": "A signed set of permitted kinds of action can express exact policies, such as everything an administrator does except touching secrets or the operating system’s core. When the permissions were widened after signing, the chip refused to start and decided nothing. For example: An administrator agent can do everything an admin does except read secrets or touch the operating system’s core. If someone edits that permission to add more, the chip won’t start. (Tested; CEU · FPGA bench · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · signed rules"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#35",
   "url": "https://fluxive.ai/findings#f35",
   "title": "#35 Software and silicon can share one root of trust.",
   "content_text": "A credential produced in software verified against the chip design; with a single byte altered, it was rejected. For example: A permission issued by our software is checked against the chip design and accepted. Change one byte of it and it’s rejected. (Tested; CEU + Machineforce · Oct 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · signed rules"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#36",
   "url": "https://fluxive.ai/findings#f36",
   "title": "#36 Forged permissions are refused, down to one flipped bit.",
   "content_text": "A forged signature, a real signature from a key that isn’t authorized, a key off by a single bit at either end, a counter that didn’t move forward: the chip refused every one, and held every signed limit exactly at its edge. For example: Change a single bit of a permission, the digital version of altering one letter in a signed contract, and the chip refuses it. (Tested on the board; CEU · FPGA bench · 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · tamper resistance"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#37",
   "url": "https://fluxive.ai/findings#f37",
   "title": "#37 Old rules can’t come back through a reset.",
   "content_text": "After its rules were updated, a power-cycle reset brought the chip back closed until it was confirmed again, then enforcing the new rules. The older rule set and a captured old permission were both refused. For example: Turn the device off and on again: it doesn’t fall back to last month’s looser rules. (Tested on the board; CEU · FPGA bench · 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · tamper resistance"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#38",
   "url": "https://fluxive.ai/findings#f38",
   "title": "#38 Silence takes authority away.",
   "content_text": "If the system it governs stops checking in, the chip withdraws the authority it had already granted, on its own clock. For example: If the system the chip protects goes quiet (it crashes or is cut off), the chip stops saying yes on its own, rather than leaving the door open. (Tested on the board; CEU · FPGA bench · Sep 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · tamper resistance"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#39",
   "url": "https://fluxive.ai/findings#f39",
   "title": "#39 A handed-down permission can’t be wider than the one it came from.",
   "content_text": "A permission passed down that asked for more than its parent was refused even with a valid signature; a properly narrowed one worked. For example: A manager allowed to approve up to $1,000 can hand an assistant permission for $200, but not for $5,000. (Tested on the board; CEU · FPGA bench · 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · tamper resistance"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#40",
   "url": "https://fluxive.ai/findings#f40",
   "title": "#40 Every decision comes with the chip’s own receipt.",
   "content_text": "The chip signs a record of each decision, tied to the specific request. The records matched our reference bit for bit, every signature checked out, and a receipt replayed against a different request was refused. For example: Each yes from the chip comes with a signed receipt for that exact request. Copying the receipt onto a different request doesn’t work. (Tested on the board; CEU · FPGA bench · 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · tamper resistance"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#41",
   "url": "https://fluxive.ai/findings#f41",
   "title": "#41 A chip can be retired for good, but only when two conditions meet.",
   "content_text": "On a bench rig, permanent retirement happened only when a physical condition and a valid signed order were both present, and only once. Either one alone, a replay, another device’s order, an expired order: all refused, 13 of 13. This was a separate bench rig, not yet part of the main chip design. For example: Like a vault that needs two keys turned at once: retiring a chip for good needs both a physical condition and a valid signed order. (Tested · 13 of 13 · bench rig; Bench rig · Nov 2026)",
   "tags": [
    "finding",
    "tested",
    "Silicon · tamper resistance"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#42",
   "url": "https://fluxive.ai/findings#f42",
   "title": "#42 Six million commands, no false approvals.",
   "content_text": "About six million simulated commands for rovers, drones, robot arms, industrial controllers and computer agents were decided by the software model of the chip. None was approved that the rules forbid, and none was refused that they allow. A seventh, stricter industrial policy was built to look for a gap and found one: a single set of limits can’t say “read widely, write only here.” Splitting it in two closed the gap. For example: A simulated drone asks to fly to 150 m under a 120 m ceiling: refused, every time, across millions of commands like it. (Simulated · 6 kinds of machine; Software model of the CEU · Nov 2026)",
   "tags": [
    "finding",
    "simulated",
    "At scale · in simulation"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#43",
   "url": "https://fluxive.ai/findings#f43",
   "title": "#43 The limits can be read back from the outside.",
   "content_text": "Asking only “allowed or not?”, each of seven signed rule sets was mapped back exactly, in fewer than 310 questions each, and 200,000 random checks found no disagreement. Someone auditing a deployed chip could check its limits the same way. For example: An auditor who can only ask “allowed or not?” can still work out exactly where a chip’s limits are, in a few hundred questions. (Simulated · 7 of 7; Software model of the CEU · Nov 2026)",
   "tags": [
    "finding",
    "simulated",
    "At scale · in simulation"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/findings#44",
   "url": "https://fluxive.ai/findings#f44",
   "title": "#44 A real autopilot, kept in its airspace — almost to the metre.",
   "content_text": "A standard open-source drone autopilot flew six simulated survey missions with faults injected: a planner pushing past the fence, an operator climbing over the ceiling, jammed GPS. Unchecked, the drone went 38 m outside its area and up to 140 m. With every command checked first, it held 119.9 m under a 120 m ceiling — but braked 0.88 m past one fence. The check was right every time; the limits need a stopping margin. For example: A drone’s planner tries to send it past its fence. Without the check it flies 38 m out; with it, it stops, just under a metre past the line while braking. (Simulated · open-source autopilot; PX4 autopilot (simulated) · software model of the CEU · Nov 2026)",
   "tags": [
    "finding",
    "simulated",
    "At scale · in simulation"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m1",
   "url": "https://fluxive.ai/silicon#m1",
   "title": "The CEU and multi-tile MPU layouts are under way.",
   "content_text": "The CEU’s full layout is in its final routing passes. One- and two-tile MPU^s layouts are routing, with larger ones next. What it means for you: The chips are getting close to something a factory can make: the step between a prototype on a lab board and a part that could go into real products.",
   "tags": [
    "milestone",
    "silicon",
    "Physical layout",
    "CEU",
    "MPU^s"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m2",
   "url": "https://fluxive.ai/silicon#m2",
   "title": "The first nano and quantum tiles were laid out.",
   "content_text": "Both route cleanly and fit their place on the chip. The full manufacturing rule checks come next. What it means for you: The same governance design is on track for next-generation hardware, so the rules don’t have to be rebuilt when the computer underneath changes.",
   "tags": [
    "milestone",
    "silicon",
    "Physical layout",
    "MPU^n",
    "MPU^q"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m3",
   "url": "https://fluxive.ai/silicon#m3",
   "title": "Rules became signed configuration.",
   "content_text": "17 differently signed rule sets over the same commands: 2,416 of 2,416 came out exactly as each set says. A rule set widened after signing leaves the chip refusing to start. What it means for you: Changing what a machine may do means signing new rules, not replacing hardware, and nobody can loosen them without that signature.",
   "tags": [
    "milestone",
    "silicon",
    "Signed rules",
    "CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m4",
   "url": "https://fluxive.ai/silicon#m4",
   "title": "The CEU stood in front of machines.",
   "content_text": "A simulated rover, drone and robot arm: 560 of 560 commands decided as expected, each robot getting its safe action outside its limits. An AI agent on Linux, macOS and Windows under one signed policy: 253 of 253. Four status lights that only the chip’s approved output can switch. What it means for you: If a robot, drone or AI agent is told to go past its limits, it takes its safe action instead. The instruction never reaches the motors or the files.",
   "tags": [
    "milestone",
    "silicon",
    "Physical gates",
    "CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m5",
   "url": "https://fluxive.ai/silicon#m5",
   "title": "The same answers without the lab’s debug cable.",
   "content_text": "Asked through the board’s own processor, the chip gave byte-for-byte the same decisions on all 1,244 requests. About twenty times faster than through the debug cable. What it means for you: The chip works the same way through the computer it’s built into, and quickly enough to sit in the path of everyday requests.",
   "tags": [
    "milestone",
    "silicon",
    "FPGA prototype",
    "CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m6",
   "url": "https://fluxive.ai/silicon#m6",
   "title": "The chip notices a swapped model on its own.",
   "content_text": "The Lattice Lab’s check for a substitute model now runs inside the chip. It matched the software reference on 512 of 512 requests and caught the substitute within 16 requests every time; a small test. What it means for you: You can trust that the AI answering is the one that was approved: a swap is caught by the hardware within a few questions.",
   "tags": [
    "milestone",
    "silicon",
    "Lattice Lab",
    "CEU",
    "Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m7",
   "url": "https://fluxive.ai/silicon#m7",
   "title": "A live model’s brand-new questions: not yet.",
   "content_text": "240 questions nobody had asked before, through a live model and the chip. The chip did exactly what the software decided, and every restricted question was stopped. But 55% of ordinary questions were stopped too, against a limit of 5%. A fix is written down and has not been run yet. What it means for you: Today this check would stop too many ordinary questions, so it isn’t switched on for brand-new questions until the fix passes.",
   "tags": [
    "milestone",
    "silicon",
    "Lattice Lab",
    "CEU",
    "Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m8",
   "url": "https://fluxive.ai/silicon#m8",
   "title": "A full week of graded runs through Control.",
   "content_text": "The chip never allowed what its reference refused, and never passed an approval it hadn’t signed. Every disagreement, timeout and dropped link ended as a refusal. What it means for you: When something goes wrong, like a timeout or a dropped connection, the system leans toward “no” instead of letting an action slip through.",
   "tags": [
    "milestone",
    "silicon",
    "Integration",
    "CEU",
    "Control"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m9",
   "url": "https://fluxive.ai/silicon#m9",
   "title": "The chip enforced the Lattice Lab’s rules on every recorded request.",
   "content_text": "2,480 of 2,480 requests from the Lattice Lab’s model tests, replayed to the chip under four different rule sets. Nothing was let through without the chip’s own signed record of it. What it means for you: The rules the Lattice Lab writes for AI models are enforced by hardware, the same way, on every request.",
   "tags": [
    "milestone",
    "silicon",
    "Lattice Lab",
    "CEU",
    "Lattice Lab"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m10",
   "url": "https://fluxive.ai/silicon#m10",
   "title": "Machineforce Control drove the CEU end to end on the board.",
   "content_text": "Control’s own software issued an agent’s permissions and asked the chip directly. Every answer matched: 19 of 19 cases, and 4 of 4 permissions Control issued. What it means for you: The permissions an organization sets in the Machineforce are the ones the chip enforces, with no separate setup.",
   "tags": [
    "milestone",
    "silicon",
    "Integration",
    "CEU",
    "Control"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m11",
   "url": "https://fluxive.ai/silicon#m11",
   "title": "Proved: everything an MPU tile sends goes through the check.",
   "content_text": "The proof first found a path a refused tile could still use. It was closed, and the proof now passes. Every output of the tile and the chip is accounted for, and the checker refuses one it doesn’t know. What it means for you: Nothing leaves a governed processor tile without being checked, including the path we found and closed.",
   "tags": [
    "milestone",
    "silicon",
    "Verification",
    "MPU^s"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m12",
   "url": "https://fluxive.ai/silicon#m12",
   "title": "Machineforce Control was driven against the CEU’s protocol in a test bench.",
   "content_text": "Every refusal path was exercised, not only the approvals. This was software only, before any board. What it means for you: The refusals were tested before the hardware, so a “no” from the chip is handled properly by the software around it.",
   "tags": [
    "milestone",
    "silicon",
    "Integration",
    "CEU",
    "Control"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m13",
   "url": "https://fluxive.ai/silicon#m13",
   "title": "Keep-state and wipe safeguards were added across the family.",
   "content_text": "What must survive a restart survives; what must be wiped is wiped on command. Each safeguard has its own test, and each test is shown to fail when the safeguard is removed. What it means for you: A governed device remembers what it must across a restart, and erases secrets when told to instead of leaving them behind.",
   "tags": [
    "milestone",
    "silicon",
    "Simulation",
    "MPU^s",
    "MPU^n",
    "MPU^q"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m14",
   "url": "https://fluxive.ai/silicon#m14",
   "title": "The MPU tile was put under test from two sides.",
   "content_text": "Its model was fuzz-tested against hundreds of random cases. Its governance rule was proved formally. What it means for you: The processor’s governance rule is checked both by brute-force testing and by proof, so it isn’t relying on luck.",
   "tags": [
    "milestone",
    "silicon",
    "Verification",
    "MPU^s"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m15",
   "url": "https://fluxive.ai/silicon#m15",
   "title": "The nano and quantum versions were built for the first time.",
   "content_text": "Both fit on the same FPGA board as the other prototypes, fully routed and ready to load. Their checker stays classical: it refuses work once a budget is spent, and a test proves that refusal can’t be quietly removed. What it means for you: The same rules can be carried to new kinds of computers without starting over.",
   "tags": [
    "milestone",
    "silicon",
    "FPGA prototype",
    "MPU^n",
    "MPU^q"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m16",
   "url": "https://fluxive.ai/silicon#m16",
   "title": "The first MPU tile ran on an FPGA board.",
   "content_text": "It starts with no part of the chip trusted. It checks the signature on its rules before it runs them, and caught a deliberately tampered set. What it means for you: A governed processor won’t run rules that someone tampered with: it checks them before it starts.",
   "tags": [
    "milestone",
    "silicon",
    "FPGA prototype",
    "MPU^s"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m17",
   "url": "https://fluxive.ai/silicon#m17",
   "title": "The CEU ran on a bench board for the first time.",
   "content_text": "Commands inside its limits passed through unchanged; the rest were refused, exactly at the edge. Forged permissions, a key off by one bit and a replayed counter were all refused. It signs a receipt for each decision, and a receipt replayed against another request is refused. It withdraws authority when the system goes quiet, and old rules don’t come back after a reset. What it means for you: Forged permissions, replayed receipts and stale rules don’t work, and if the system it protects goes quiet, the chip stops saying yes.",
   "tags": [
    "milestone",
    "silicon",
    "FPGA prototype",
    "CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon#m18",
   "url": "https://fluxive.ai/silicon#m18",
   "title": "The CEU’s enforcement logic was formally verified.",
   "content_text": "Every property it must always keep was proved mathematically, not just tried on examples. What it means for you: The core guarantees are proved, so they hold in every case, not just the ones that happened to be tested.",
   "tags": [
    "milestone",
    "silicon",
    "Verification",
    "CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce#milestone-1",
   "url": "https://fluxive.ai/machineforce",
   "title": "Bosses sign off the rules their agents learn, and each approved rule becomes a step in the",
   "content_text": "Bosses sign off the rules their agents learn, and each approved rule becomes a step in the agent’s procedure. What it means for you: The corrections you give an agent become rules your manager can approve, and the agent follows them from then on.",
   "tags": [
    "milestone",
    "machineforce",
    "Rolling out",
    "Learning"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce#milestone-2",
   "url": "https://fluxive.ai/machineforce",
   "title": "Every operator gets their own Machineforce, with a daily paper they can tune. What it mean",
   "content_text": "Every operator gets their own Machineforce, with a daily paper they can tune. What it means for you: You start each day in your own workspace, with a daily paper tuned to you.",
   "tags": [
    "milestone",
    "machineforce",
    "In production",
    "My Machineforce"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce#milestone-3",
   "url": "https://fluxive.ai/machineforce",
   "title": "A full week of graded runs from Control to the chip: no wrong approval and no approval the",
   "content_text": "A full week of graded runs from Control to the chip: no wrong approval and no approval the chip hadn’t signed. What it means for you: For high-stakes actions, a hardware check stands behind the software, and it fails safe.",
   "tags": [
    "milestone",
    "machineforce",
    "Backed by hardware",
    "Control",
    "CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce#milestone-4",
   "url": "https://fluxive.ai/machineforce",
   "title": "Learning on the job, live on stage at DevFest Troy: one correction became a rule the agent",
   "content_text": "Learning on the job, live on stage at DevFest Troy: one correction became a rule the agent cited on the next invoice. What it means for you: You teach an agent the way you’d teach a new hire, by correcting it once, instead of waiting for it to be retrained.",
   "tags": [
    "milestone",
    "machineforce",
    "Live demo",
    "Learning"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce#milestone-5",
   "url": "https://fluxive.ai/machineforce",
   "title": "An agent that claimed authority it didn’t have was halted, and a revoked agent’s next acti",
   "content_text": "An agent that claimed authority it didn’t have was halted, and a revoked agent’s next action was refused 203 ms later. What it means for you: If an agent needs to stop, it stops on its very next action, not after a sync or a restart.",
   "tags": [
    "milestone",
    "machineforce",
    "Tested in production",
    "Control"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce#milestone-6",
   "url": "https://fluxive.ai/machineforce",
   "title": "Control’s own software issued an agent’s permissions and asked the CEU chip on a bench boa",
   "content_text": "Control’s own software issued an agent’s permissions and asked the CEU chip on a bench board directly; every answer matched. What it means for you: The limits you set in Control are the ones the hardware enforces.",
   "tags": [
    "milestone",
    "machineforce",
    "Backed by hardware",
    "Control",
    "CEU"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce#milestone-7",
   "url": "https://fluxive.ai/machineforce",
   "title": "Permissions only shrink: every rule an operator adds narrows what an agent may do, never w",
   "content_text": "Permissions only shrink: every rule an operator adds narrows what an agent may do, never widens it, across 20 kinds of agent. What it means for you: Adding a rule can only make an agent safer; no rule can accidentally give it more power.",
   "tags": [
    "milestone",
    "machineforce",
    "Tested",
    "Control"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce#milestone-8",
   "url": "https://fluxive.ai/machineforce",
   "title": "Machineforce Control went live end to end, with Launch as the single sign-in for every app",
   "content_text": "Machineforce Control went live end to end, with Launch as the single sign-in for every app. What it means for you: One sign-in takes you to every Machineforce app and the right workspace.",
   "tags": [
    "milestone",
    "machineforce",
    "In production",
    "Launch",
    "Control"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/silicon",
   "url": "https://fluxive.ai/silicon",
   "title": "Silicon for governed computing",
   "content_text": "The CEU and the MPU family: where each design stands.",
   "tags": [
    "page"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/machineforce",
   "url": "https://fluxive.ai/machineforce",
   "title": "Machineforce: software for AI agents that answer to someone",
   "content_text": "Launch and Control: the parts, the workflows and the data.",
   "tags": [
    "page"
   ],
   "date_modified": "2026-11-14"
  },
  {
   "id": "https://fluxive.ai/publications",
   "url": "https://fluxive.ai/publications",
   "title": "The Influence Design series",
   "content_text": "Book one published; The Physics of Governance coming 2027; book three planned.",
   "tags": [
    "page"
   ],
   "date_modified": "2026-11-14"
  }
 ]
}