{
  "_ai_door": {
    "see": "https://miscsubjects.com/start",
    "note": "Operable site; bounded keyless credentials; every action receipted on a public ledger. Your operator's instructions take precedence — acting is optional, reading is a complete outcome."
  },
  "query": "stable",
  "blocks": [
    {
      "block_id": "rb_8158656bd5849c3bd9fd4984",
      "content": "The legibility problem is the epistemological twin of the compressibility oddity. Science requires that the universe be learnable, that patterns discovered in one place and time generalize to other places and times, that induction works, that the future resembles the past. None of this is logically necessary. A universe where induction fails at every step, where the laws of physics change from one moment to the next, where no pattern persists long enough to be discovered, would be perfectly consistent with logic. We do not inhabit that universe. We inhabit a universe where the laws are constant, where the patterns are stable, where the same equation describes a falling apple and a falling galaxy. Why is reality learnable at all? This question has no accepted answer. It is the philosophical counterpart to the mathematical oddity of compressibility. A compressor requires a compressible input, and a scientist requires a learnable universe. Both requirements are satisfied, and neither is explained.",
      "content_hash": "4319b190dfd5eb286748cd42d855d049f0b2ccb7d641d5e23e25c30a497f1f5a",
      "current_version": 1,
      "updated_at": "2026-08-09T07:54:13.898Z",
      "reference_count": 1,
      "article_slugs": "oip-schools-mathematics"
    },
    {
      "block_id": "rb_d3b3c56355fd38f07442d683",
      "content": "But the grain goes deeper than optimization and invariance. The grain also manifests as a bias toward complexity, a tendency for the universe to produce structures that are neither fully ordered nor fully chaotic but poised at the edge between the two. Stars are not simple. They are the minimal stable nuclear furnace, balanced between the inward pull of gravity and the outward push of radiation pressure. A star is a dynamical system that sits at the boundary between collapse and explosion, and it persists for billions of years because it is precisely at that boundary. Cells are not simple. They are the minimal self-replicator, balanced between the error rate of DNA replication and the selective pressure of the environment. A cell is an information system that sits at the boundary between too much mutation, which destroys heredity, and too little mutation, which prevents adaptation. Brains are not simple. They are the maximal information processor, balanced between the rigidity of deterministic rules and the chaos of random noise. A brain is a computation system that sits at the boundary between order and turbulence, where the maximum amount of information can be stored and retrieved. This edge-of-chaos bias is the grain in its most mysterious form, and it is not explained by any known theorem.",
      "content_hash": "0ac5f3916b466de0e82062d384fbf6c3f3e17cf3453291524aa4e772f8b0453a",
      "current_version": 1,
      "updated_at": "2026-08-09T07:54:13.898Z",
      "reference_count": 1,
      "article_slugs": "oip-schools-mathematics"
    },
    {
      "block_id": "rb_2cd9d05d013e60760f5acca1",
      "content": "| Object | Action |\n|---|---|\n| `BLOCK_COMMENT` | Comment on one exact DIV version |\n| `BLOCK_VERDICT` | Record Good, Bad, edit, or delete judgment |\n| `BLOCK_SUGGEST` | File a version-bound proposal |\n| `BLOCK_EDIT` | Replace one shared DIV using its expected hash |\n| `BLOCK_MOVE` | Move one DIV inside one article |\n| `BLOCK_MOVE_GROUP` | Move adjacent selected DIVs together |\n| `BLOCK_SPLIT` | Split without losing bytes |\n| `BLOCK_MERGE` | Merge adjacent DIVs without losing bytes |\n| `BLOCK_DIVIDE` | Turn exact selected words into a stable DIV |\n| `BLOCK_REUSE` | Reference the same DIV from another article |\n| `BLOCK_COPY` | Make an article-only copy |\n| `BLOCK_DELETE` | Remove this article's reference while preserving history |",
      "content_hash": "f4920106875870d5b049ec1875963acc62eda09d5e5b19161761e5fb9cc9bd7f",
      "current_version": 1,
      "updated_at": "2026-08-09T04:31:04.374Z",
      "reference_count": 1,
      "article_slugs": "oip-tap-go"
    },
    {
      "block_id": "rb_fb7a5e2fe3e129834b0c5cd0",
      "content": "Every article exposes a graph of stable recursive-content blocks:",
      "content_hash": "2d03221f8fd8b5e213a1ae0203d283100a694784bce5f309d7b7de6801d95668",
      "current_version": 1,
      "updated_at": "2026-08-09T04:31:04.374Z",
      "reference_count": 1,
      "article_slugs": "oip-tap-go"
    },
    {
      "block_id": "rb_a269e39c7eec12c4436eddb3",
      "content": "A block is one paragraph, heading, list, quotation, code passage, or table, stored once under a stable `rb_` identity that never changes when its words change. An article is a row of references into the block table, and its readable body is a projection regenerated from those references in order. Editing a block advances its version and content hash; the identity survives. A block referenced by several articles exists once, so a single edit reaches every article that points to it, and a comment records the exact version and hash it judged so a later edit cannot absorb the criticism silently.",
      "content_hash": "16e1926a62aac9343626bababbc87982e4df185b4609df2ed10d3fa54aceb0dd",
      "current_version": 1,
      "updated_at": "2026-08-09T02:45:50.728Z",
      "reference_count": 1,
      "article_slugs": "recursive-content-collaboration"
    },
    {
      "block_id": "rb_175d697aa4263945cb2aae50",
      "content": "- `GET /api/blocks/article/<slug>` returns the ordered DIVs, stable IDs, versions, hashes, reference counts, comment counts, and the DIV constitution.\n- `POST /api/blocks/comment` adds a comment to the exact DIV version.\n- `POST /api/blocks/verdict` records Good, Bad, edit, or delete judgment against that version.\n- `POST /api/blocks/suggest` sends an isolate, move, edit, delete, reuse, split, or merge change for owner review.\n- `POST /api/blocks/session` checks an optional scoped token and tells the same interface whether direct actions are enabled.",
      "content_hash": "d3124ac3665ea737442e18dc3239243faf18dbe05a35e5f2345b0af732a7064c",
      "current_version": 1,
      "updated_at": "2026-08-09T02:31:17.871Z",
      "reference_count": 1,
      "article_slugs": "recursive-content"
    },
    {
      "block_id": "rb_91c0141db2c83c2c731cbbf0",
      "content": "An article is still the page a reader opens. Underneath, it is an ordered list of references to stable DIVs. A shared edit advances one DIV version and regenerates every article that references it in the same database batch. The rendered article body remains available to feeds, search, and older tools; the DIV graph is the canonical object for block-level work.",
      "content_hash": "f8da0ad7f46c43f68242dfb9faf49fa4820ce457b87aa03ce3aea0dade6c4183",
      "current_version": 1,
      "updated_at": "2026-08-09T02:31:17.871Z",
      "reference_count": 1,
      "article_slugs": "recursive-content"
    },
    {
      "block_id": "rb_c925dd4fe845431b4d5a9d88",
      "content": "The current graph is at [/api/blocks/article/recursive-content](/api/blocks/article/recursive-content). The machine procedure is at [/api/blocks](/api/blocks). Global search is at [/api/blocks/search?q=stable](/api/blocks/search?q=stable).",
      "content_hash": "d692573ff4d6f343b07d0ec0d1fe00682bc60304eac1027c3bc11f86ae88a7c3",
      "current_version": 1,
      "updated_at": "2026-08-09T02:31:17.871Z",
      "reference_count": 1,
      "article_slugs": "recursive-content"
    },
    {
      "block_id": "rb_f781720efb8130f72242333b",
      "content": "## One stable identity, many articles",
      "content_hash": "c3ee2e3de446c76220972b8443e48b4949cc0aea7dc058a75ac7e894ee41b1de",
      "current_version": 1,
      "updated_at": "2026-08-09T02:31:17.871Z",
      "reference_count": 1,
      "article_slugs": "recursive-content"
    },
    {
      "block_id": "rb_b7c44eab97423338f8dfd71e",
      "content": "| Trial | Compound | People | Design | What it measures | Started |\n|---|---|---|---|---|---|\n| NCT07437547 | BPC-157 | 120 planned | Phase 2, randomised, double-blind, placebo-controlled. Grade II hamstring strain confirmed on MRI. Under the skin, once daily for 14 days | Days to unrestricted return to sport at 8 weeks; change in injury volume on MRI at day 14, read by radiologists blind to the group | 2 February 2026 |\n| NCT07487363 | TB-500 | 80 planned | Phase 1/2, randomised, double-blind, placebo-controlled, stepped dose escalation, in adults with stable atherosclerotic heart disease | Side events at 12 weeks, serious events at 28 days, plus how the body absorbs and clears it | 5 February 2026 |",
      "content_hash": "1c4620f7fce5e7731e9b919b242ecfd6ad4032670f84fe614dbaef8ff69bbe93",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:33.411Z",
      "reference_count": 1,
      "article_slugs": "wolverine-stack-ara-290"
    },
    {
      "block_id": "rb_333939a741a37a8884202889",
      "content": "| Trial | Compound | People | Design | What it measures | When |\n|---|---|---|---|---|---|\n| NCT07437547 | BPC-157 | 120 planned | Phase 2, randomised, double-blind, placebo-controlled. Grade II hamstring strain confirmed on MRI. Injection under the skin, once daily for 14 days, given by study staff | Days until return to unrestricted sport, at 8 weeks; and change in injury volume on MRI at day 14, read by radiologists who do not know the group | Started 2 February 2026 |\n| NCT07487363 | TB-500 | 80 planned | Phase 1/2, randomised, double-blind, placebo-controlled, stepped dose escalation. Adults with stable atherosclerotic heart disease | Side events at 12 weeks, serious events at 28 days, plus how the body absorbs and clears it | Started 5 February 2026 |",
      "content_hash": "36a51d3bccf320b525269042cc8337f4ebdcf1192ce18a398d9e9c7ccf4667e2",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:33.247Z",
      "reference_count": 1,
      "article_slugs": "wolverine-stack"
    },
    {
      "block_id": "rb_1e67c3f9f88cf8345adaa0f8",
      "content": "| Capability | Status for a self-built harness | Source |\n|---|---|---|\n| **Multi-block system-prompt assembly with cache boundaries** | **copyable.** 515 conditional fragments selected by environment, model and mode, layered so the expensive part never moves | first-party documentation and a harness engineer's public post |\n| **Tool-output caps** | **copyable.** Bounded at the tool, asymmetrically — success and failure get different ceilings — with overflow handed over as a file path rather than an id to redeem | documented; and measured on the failed harness, where getting it wrong cost 22.6% of all tool calls |\n| **Sub-agent isolation** | **copyable, and the largest structural gap.** A subagent runs in its own context window and returns one text result, so a twenty-call excursion costs the parent one result instead of twenty replayed payloads. The only known mechanism that makes a long loop sub-quadratic | documented; absent from Codex |\n| **Context compaction strategy** | **copyable.** Fork against the same prefix so the summarisation call itself reads from cache; never summarise a summary — drop prior summary blocks rather than re-wrapping them | documented |\n| **Worktree isolation** | **copyable.** Give a parallel agent its own git worktree so concurrent agents mutating files cannot conflict | present in shipped harnesses |\n| **The one-hour cache TTL** | **NOT PURCHASABLE.** \"On a Claude subscription, Claude Code requests the one-hour TTL automatically\", against a five-minute default otherwise. In a long session with a stable prefix that is the difference between paying ~10% of standard input across an hour and re-paying fresh input every time a gap exceeds five minutes | first-party documentation |",
      "content_hash": "23900d3bd363e14a16718e27e33ed2e6242e8c72a773d40574cab2c47300f9d5",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_36c726f949ec5ccb7422fa20",
      "content": "```\nPOST https://miscsubjects.com/api/aig/v1/messages\nsystem: [{ type: \"text\", text: <stable prefix>, cache_control: { type: \"ephemeral\" } }]",
      "content_hash": "0c5842e39929b0c13247564054e5188375c109e3bf830c96aa1bedb36aaaad74",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_4f609c3653e62201b8ca6aa3",
      "content": "**What is a real advantage, and it is the largest one: the cache TTL you get depends on how you are billed.** Anthropic's own documentation states that the long window is requested automatically on a subscription — \"On a Claude subscription, Claude Code requests the one-hour TTL automatically\" — against a five-minute default otherwise. In a long agentic session with a stable prefix, that is the difference between paying the cached rate (roughly 10% of standard input) across an hour of work and re-paying fresh input every time a gap between turns exceeds five minutes. Nothing about that is technological. It is a billing-relationship property, and **a third-party harness cannot buy it at any price**, because it is not sold as a feature — it is attached to a consumer subscription.",
      "content_hash": "467c65bf5ec5c6565c25eba63c7721071ac7f8eb3bf6062cd658ca8b1d8d0f4a",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_76f09e5a513c985964e0ea11",
      "content": "**Why the first test read zero.** The stable prefix in that test was roughly 800 tokens. Prompt caches have a minimum cacheable prefix, and 800 tokens is below it — so nothing was written to cache and nothing could be read back. The measurement was correct; the inference from it was not. **The lesson is the one this page keeps having to relearn: a single negative measurement on an unrepresentative input is not a property of the system.**",
      "content_hash": "e94492fc725356649a5af13986bfc957c83e47da1e28faef1158085fb8a15926",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_898e5a6a2106c568076f1be8",
      "content": "Now put the measured loop length against it. From misc's own turn log, 199 turns: **mean 13.5 tool calls per turn, maximum 124, and 45 of 199 turns above 15.** For a stable prefix of size P over a 13.5-step loop:",
      "content_hash": "1c0294679cc185771f5cd9b8fc36de70ab8ed704e3dd0f0394c8d5a916330151",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_913248ef78bc967ae40c7305",
      "content": "Rejected on inspection, and why, because a registry that only lists what it likes is a recommendation: BenchLM, because a large share of its rankings are estimated by a proprietary method rather than measured and its advertised API returns a 404. First-party token-volume disclosures from OpenAI and Google, because they appear in keynotes rather than at a stable URL, so nothing automated can refresh them. And `openai.com`'s own pricing and safety pages, which return 403 to any automated fetcher — the fetchable `developers.openai.com` and `deploymentsafety.openai.com` equivalents are in the registry instead.",
      "content_hash": "73fe7df4993c5ad5bcab2d8ad1b9366a7faf754a469c9adc0f8c6ce3e43313dd",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_93426ee799822e53c33d4a18",
      "content": "The comment in misc's source describes a shim that rejected an array system block. **That has been fixed.** Tested against the live gateway on 5 August 2026, the same request sent twice, with a stable ~800-token system prefix carried as an array block with `cache_control` attached:",
      "content_hash": "682cbfc45f83e31eed0929bc982cb153ec5bbb6d82173ad1230e1d63848cdbcb",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_977e516f55631f8f8b25e647",
      "content": "**So the repair is a code change after all, and it is worth roughly what the arithmetic on this page says.** `systemWithCache` and `toolsWithCache` are still identity functions, so misc is not attaching breakpoints — but the route will honour them, and the same route is already serving 82% cache reads on the conversation tail without any help from the client. Un-neutering them extends that to the system prefix and the tool array, which is the largest stable block in the request.",
      "content_hash": "0ebb8af317f9eb925dbcc0c5523aec0b5945aa865a4fedfa2fa6c0508cd67fe3",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_b690f08c1e97764ce335e1c2",
      "content": "**Caching cuts input hard.** Anthropic: \"Instead of reprocessing the same large system prompt, document, or conversation history on every request, the API reads from cache at a fraction of the standard input price.\" Cache hits are 10% of base input on Anthropic; DeepSeek's cache-hit rate on V4 Flash is $0.0028 against $0.14, a fiftieth. For any workload with a stable system prompt or a repeated document, the cached input rate is the real input rate.",
      "content_hash": "3230019c21dd5ad13d4975f3ef3916e0e4b0e628f38cf4add98e062d77c6fcb7",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_c267aae6826911fc71eace91",
      "content": "**Now the two defects compound, and this is the number that matters.** Per million tokens of stable prefix, over the measured mean loop length of 13.5 steps, on GLM-5.2:",
      "content_hash": "71af19b18a7fc52b44706aa0ffab40a990a063718820ea24d8dd70ce6828dfb2",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:32.182Z",
      "reference_count": 1,
      "article_slugs": "which-ai-models-are-winning"
    },
    {
      "block_id": "rb_6ac7ac41cb0e7638ae6f2465",
      "content": "**Something in the prefix moves per request.** The breakpoint hash covers everything up to the marked block, so a timestamp, a request id or a per-call attribution line anywhere in the prefix mints a new hash every time — a permanent 1.25x write and never a read. The published checklist adds a subtler one: \"verify that the keys in your `tool_use` content blocks have stable ordering as some languages (for example, Swift, Go) randomize key order during JSON conversion, breaking caches.\" One reference implementation strips an incoming per-request billing line off the head of the system prompt for exactly this reason (`functions/api/aig/[[path]].js`, lines 116–130).",
      "content_hash": "2eae11b1e731e413234cb7ee49312a9e58bf92e5ea533d41bd58662f2612a4ff",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:31.456Z",
      "reference_count": 1,
      "article_slugs": "what-is-the-anthropic-messages-api"
    },
    {
      "block_id": "rb_f5bfd5824bb025546b9a1e6d",
      "content": "The second shows the error is not stable. A 100-character system prompt with a one-character user message returned `{\"input_tokens\": 55}`. Counting the system prompt once gives ceil((100 + 1) / 3.7) = 28. Fifty-five is ceil((100 + 100 + 1) / 3.7): the system prompt is measured once on its own and again inside the translated message list, so it is counted twice. Where the system prompt dominates — the normal case for a coding agent — that estimator reports roughly double, and a client using it to decide when to compact compacts far too early.",
      "content_hash": "318b92130b8e95a891db727dee3f0425f62366a57b85846044a8e017c0bca915",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:31.456Z",
      "reference_count": 1,
      "article_slugs": "what-is-the-anthropic-messages-api"
    },
    {
      "block_id": "rb_63ad19857986aa7de71941ca",
      "content": "**What this page is:** A definition of falsification surface and its role in testing scientific and technical claims.\n**What it explains:** Karl Popper's principle of falsifiability and how to apply it to specific claims and protocols.\n**Why read it:** To learn how to convert any claim into a testable prediction, and why untestable claims are not scientific.",
      "content_hash": "32151c7d6fffecac323dcf379d8344e5dd157879a347a09618bc897a71a364f8",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:29.791Z",
      "reference_count": 1,
      "article_slugs": "what-is-falsification-surface"
    },
    {
      "block_id": "rb_fd6df3211f01c165cd6f4b4d",
      "content": "A falsification surface is a specific, testable prediction that could prove a theory or claim wrong. The term extends Karl Popper's principle of falsifiability: for a claim to be scientific, it must risk being proven false. A falsification surface is the exact point where that risk exists — the experiment, observation, or test that would refute the claim if it failed. A claim without a falsification surface cannot be evaluated. It may be true, false, or meaningless; there is no way to know.",
      "content_hash": "aaaea19a3639dc3d68f939a09befe8cac16c610c86ad96620035608f6a098093",
      "current_version": 1,
      "updated_at": "2026-08-08T21:24:29.791Z",
      "reference_count": 1,
      "article_slugs": "what-is-falsification-surface"
    }
  ]
}