{"path":"research/corpus-graph/haiku-5-5-for-the-filling.md","content":"# Haiku 5.5 for filling the corpus graph: what it is, how it can be reached, and a test against last night's answers\n\n**Date**: 2026-10-11 · **Type**: an evaluation, with the predictions written before any result ·\n**Status**: Claude's test and reading. Nothing here is ruled.\n\n**His words**, 2026-10-11: *\"Also, Haiku 5.5 is out. Evaluate it for use in the corpus-graph filling.\"*\n\nThis note belongs to the corpus graph, the index of this project's notes and ledger rows\n([README.md](README.md)). The roles a model has in filling it are described in\n[reading-for-every-kind-of-insight.md](reading-for-every-kind-of-insight.md) § 2.\n\n---\n\n## 1. What it is\n\nRead on 2026-10-11 from the highlights a search returned of Anthropic's announcement page and its\nmodel page, and from the project's own price lookup. Neither page was fetched whole.\n\n| | Haiku 5.5 | Sonnet 5.5, which did the filling so far |\n|---|---|---|\n| Released | 2026-10-07 | |\n| Name for a call | `claude-haiku-5-5` | `claude-sonnet-5-5` |\n| How much it holds at once | 1 million tokens | 1 million |\n| Price for each million tokens, in and out | $0.10 and $0.50 while the prompt is under 100,000 tokens; $0.50 and $2.50 above | $2 and $10 |\n| Price to re-read a cached million | $0.01 under 100,000; $0.05 above | $0.10 |\n| How hard it thinks | adjustable, five steps, *medium* unless told | adjustable |\n\nAnthropic places it for narrowly scoped work in volume (summaries, classification, work as a\nhelper to a larger model) and says the larger models stay the better choice for long, complex\ntasks. On every benchmark in its own table Sonnet 5.5 scores higher: 1840 against 1620 on the\nknowledge-work test, 56.9% against 45.9% on the hard reasoning exam. These are the company's own\nruns.\n\n**What the price means here.** A reader's prompt grew to between 320,000 and 430,000 tokens last\nnight, so a reader on Haiku pays the higher of its two rates: a quarter of Sonnet's price for new\ntext and half for re-reading. A judge or a card-writer that is handed its passages in one message\nstays under 100,000 tokens and pays a twentieth and a tenth. On a subscription no money changes\nhands for either; what is saved is the plan's allowance.\n\n## 2. How it can be reached today\n\n- **Not through this app's own helpers.** A helper asked for *haiku* here still runs Haiku 4.5\n  (read from a probe's transcript: `claude-haiku-4-5-20251001`). The app is held from updating\n  while the founder travels, on purpose. *Changed 2026-10-15: the app was updated, and a helper asked\n  for* haiku *now runs `claude-haiku-5-5` (probe transcript), so a Haiku pass can be a helper of the\n  app as well as a script.*\n- **Through the command-line tool on his login.** The project's runner for the Claude engine\n  (`deliberus/extraction/claude_backend.py`) starts the tool in an empty folder with no settings\n  loaded, so no instruction file and no reminder reaches the call. Called that way with\n  `claude-haiku-5-5`, it answered in under four seconds. The dollar figure the tool prints for this\n  model is wrong (it has no price for it), so cost below is worked out from the token counts and\n  the published rates. It is a list price. No money is paid on a subscription.\n\n## 3. The test, with the predictions written first\n\nThree roles, each against something last night's Sonnet helpers produced, with the answers known.\nThe judge and the card-writer are given exactly the lines those helpers read (from the repository\nas it stood at the end of that night), in one message, with no tools. That is also the repaired\nform for a blind helper: it opens nothing, so nothing can be shown to it.\n\n**The judge.** The 26 pairs of the first judge: 16 a reader had tied and 10 decoys. Last night's\njudge called 11 of the 16 the same move and none of the decoys. Three runs on the same message:\nSonnet 5.5 at medium effort (does the new form reproduce last night's judge at all), Haiku 5.5 at\nmedium, Haiku 5.5 at high.\n\n**The card-writer.** The 67 passages one card-writer had. Its cards shared a name for the move with\nthe partner's card (written by another helper who never saw this passage) for 23 of 67 pairs, and\nstood among the partner's first three for 28. One run: Haiku 5.5 at medium.\n\n**The reader.** A kinship reading of four public pages that have had none\n(`docs/convergence.md`, `docs/analysis-and-attunement.md`, `docs/the-missing-layer.md`,\n`docs/open-problems.md`), with the brief of last night's kinship reader. Haiku 5.5 at medium, with\nthe tools to read and to search files and no shell. Last night's reader on four other pages\nreturned 14 findings, every quote exact, 11 of which a judge called the same move.\n\n**Written before any run.**\n\n1. Sonnet with the passages in one message agrees with last night's judge, yes or no, on at least\n   22 of the 26 pairs, and accepts at most one decoy.\n2. Haiku at medium agrees with last night's judge on at least 20 of 26 and accepts at most two\n   decoys.\n3. Haiku at high differs from Haiku at medium on at most three pairs.\n4. Haiku's cards share a name with the partner's card for at least 15 of 67 pairs and stand among\n   the partner's first three for at least 20.\n5. The Haiku reader returns at least 10 findings, and the quoted words stand in the files for at\n   least nine in ten of them.\n6. Of the reader's findings, a judge calls at least five the same move.\n\n**Against Haiku**: four or more decoys accepted, or fewer than six of the sixteen real pairs\ncalled the same move, and it cannot judge. Quotes that fail in more than a quarter of findings,\nand it cannot be a reader, since every finding would have to be read again by hand. Fewer than\nten shared names, and its cards do not meet another writer's.\n\n## 4. What came back\n\n*The judge and the card-writer, 2026-10-11 at midday. The reader was still running when this was\nwritten, and its result follows below when it is in.*\n\n**The judge** (26 pairs in one message of about 107,000 tokens; nobody has graded these pairs by\nhand, so the only hard key is the ten decoys):\n\n| Run | Called the same move, of 16 real | Of 10 decoys | Agrees with last night's judge, of 26 | Time | List price |\n|---|---|---|---|---|---|\n| last night's judge (Sonnet 5.5, reading the files itself) | 11 | 0 | | | |\n| Sonnet 5.5, passages in one message, medium effort | 9 | 0 | 22 | 34 s | $0.31 |\n| Haiku 5.5, medium effort | 11 | 0 | 24 | 134 s | $0.14 |\n| Haiku 5.5, high effort | 11 | 0 | 22 | 139 s | $0.14 |\n\n- *Predicted: Sonnet agrees on at least 22 with at most one decoy. Held (22, none). Haiku at medium\n  agrees on at least 20 with at most two decoys. Held (24, none). Haiku at high differs from medium\n  on at most three. Held (two).*\n- **No run accepted a decoy**: thirty refusals of thirty today, fifty of fifty with last night's.\n- **A verdict on one real pair is not steady.** The same model given the same pairs a second time\n  changed its answer on 4 of 26, all of them real pairs. Haiku agrees with last night's judge more\n  often than Sonnet's own second run does.\n- **Haiku splits the second question differently.** Of the eleven pairs it calls the same move, it\n  says *for the same reason* of five, where Sonnet said it of all nine it accepted. Nobody knows\n  which is right.\n- **Haiku was the slower of the two here**, four times, and wrote seven times as many tokens\n  (28,000 against 4,000), most of them thinking.\n- **The price was half of Sonnet's and could be a tenth.** The message crossed 100,000 tokens, so\n  Haiku's higher rate applied. Thirteen pairs to a message stays under it.\n\n**The card-writer** (67 passages in one message):\n\n| | Shares a name for the move with the partner's card, of 67 | With a stranger's card |\n|---|---|---|\n| last night's card-writer (Sonnet 5.5) | 23 | 3.3% |\n| Haiku 5.5, medium effort | 21 | 3.5% |\n\n- *Predicted: at least 15. Held.* Whether Haiku's cards stand as near the partner's by meaning is\n  not yet scored.\n- Haiku's card and last night's card for the same passage share a name in 51 of 67. So three in\n  four passages get the same name from two different models.\n- 124 seconds, list price $0.13, again just over the 100,000-token line.\n\n**The reader, first run** (medium effort; the tools to read and search and no shell; a one-line\nsystem prompt of Claude's in place of the tool's own; and the command-line tool, which is older\nthan the model, treating it as holding 200,000 tokens).\n\n- It stopped after 51 steps and under three minutes, and returned 5 findings. *Predicted: at least\n  10. Missed.*\n- It did not keep the form: no line numbers, a count of words where the quoted words belong, and\n  labels of its own for the kind. It wrote that it had made the subject-free sentence for 0\n  passages, which is the method the brief asks for.\n- Every phrase it quoted stands in the file it names, nine of nine, checked by code.\n- Two of the five were excluded by the brief's own rules: one pair is in the index, and one page\n  already names the other note. Three pairs are new.\n\n**A second run, with what was unfair to it removed**, and the prediction written before it: the\ntool's own system prompt for an agent, high effort, and the tool told that the model holds a\nmillion tokens (the model's name with `[1m]` after it, which a probe showed the tool accepts). The\ntools stay the same. Predicted: at least 10 findings, in the block form with line numbers, quotes\nexact in at least nine of ten, and at most one pair excluded by the brief's own rules. If it again\nreturns under 8 or breaks the form, the cause is the model on this task and no setting.\n\n**The reader, second run** (60 steps, six minutes):\n\n- 8 findings. *Predicted: at least 10. Missed, and not under 8.*\n- The form was kept: every block has its line numbers, and it reports the subject-free sentence\n  written for 44 passages. *Held.*\n- 15 of the 16 quoted passages stand character for character. The sixteenth dropped one quotation\n  mark that stands inside the words, so the checking script refused that finding and it was kept by\n  hand. *Predicted: nine in ten. Held.*\n- None of the eight pairs was in the index, and no page names the other note. *Predicted: at most\n  one excluded. Held.*\n- **All eight stand when both ends are read**, and they are kept in the index\n  ([the report](readers/2026-10-11-four-top-level-pages-kinship-as-a-haiku-reader-returned-it.md)).\n  They reach into notes the pages do not name: record linkage, the intelligence analyst's four\n  reasons a thing can be absent, an election test's order of runs, the note on starting values.\n- **Put to two judges with eight decoys.** A Sonnet judge called 5 of the 8 the same move\n  (*predicted: at least five; held*) and a Haiku judge 3 of 8. The three that both refused are\n  findings of another kind: the older note states a test that the page asks for and does not have.\n  That is the kind [likeness-seen-from-outside.md](likeness-seen-from-outside.md) describes, and a\n  question about *the same move* sorts it out by design.\n- **The first decoy any judge has accepted turned out not to be a decoy.** The Sonnet judge said\n  yes to a pair made by giving one page's passage another finding's partner. On reading, the two do\n  share a condition (independence decides both when strands of support add and when agreement\n  between systems is evidence), and the pair is kept as found by the session. Decoys made by\n  swapping partners among findings on one theme are not clean. A decoy has to come from a note\n  that has nothing to do with the target, or be read before it is counted.\n- At list price the run is between $0.09 and $0.46, most likely near the lower figure, since the\n  average prompt of its sixty calls was about 39,000 tokens.\n\nNot like for like: last night's Sonnet reader of the same brief had four other pages, a shell with\nthe search by meaning, and the corpus graph's reminders on opening a note. It returned 14 findings.\n\n## 5. What this says, by role\n\n| Role | Reading | Why |\n|---|---|---|\n| **A judge of pairs** | usable now, as a second judge beside Sonnet | 28 decoys in three runs and none accepted; it agrees with last night's judge as often as Sonnet's own second run does; it is stricter than Sonnet about *the same move* (3 of 8 against 5 of 8) and about *the same reason* |\n| **A writer of cards** | usable now | its name for a move meets another writer's as often as Sonnet's did (21 and 23 of 67), and it gives the same name as Sonnet for three passages in four. By nearness of meaning its card is among the partner's first three less often (21 against 28) and first slightly more often (16 against 14) |\n| **A reader** | usable for a further pass over notes already read; not shown to replace Sonnet as the first reader | with the tool's own instructions and high effort it kept the form, quoted exactly and returned eight pairs that all stand. With a one-line instruction and medium effort it returned five and broke the form |\n\n- **The saving is real only under 100,000 tokens a message.** A judge handed 13 pairs, or a\n  card-writer handed 60 passages, stays under it and costs about a twentieth of Sonnet. A reader's\n  calls stayed under it too in the second run, because the runner loads no instruction file.\n- **Two judges are worth more than one.** The same Sonnet judge given the same 26 pairs twice\n  changed its answer on four. So a single yes or no on one pair is weak evidence whatever the\n  model, and a pair on which a Sonnet judge and a Haiku judge differ can be marked unsteady for\n  the price of the cheaper one. Both are of one family of models, so this does not answer the\n  question of a judge from another family\n  ([reading-for-every-kind-of-insight.md](reading-for-every-kind-of-insight.md) § 6.3).\n- **More passes, to learn how much is unfound.** An estimate of what readers have not found needs\n  four or more independent passes of one kind. At Haiku's price that becomes affordable for the\n  first time.\n- **It can only be reached through the command-line tool** until the app is updated, so a Haiku\n  pass is a script and no helper of this app. The scripts are in [model-tests/](model-tests/).\n\n**Limits.** Nobody has graded these pairs by hand, so every count of *the same move* is one\nmodel's answer set against another's. One message for each role, on one day. The reader was tried\non four pages. And the comparison of readers is between different pages and different tools.\n\n## 6. The pilot's two briefs on Haiku as well\n\nThe pilot of [reading-for-every-kind-of-insight.md](reading-for-every-kind-of-insight.md) § 7 sends\ntwo Sonnet readers through four notes with new briefs: one for what does not fit\n([dissonance-reading-brief.md](dissonance-reading-brief.md)) and one asked an open question\n([open-reading-brief.md](open-reading-brief.md)). The same two briefs go to Haiku 5.5 on the same\nfour notes, at high effort with the tool's own instructions, with the tools to read and to search\nand no shell. This is the first comparison of the two models as readers on the same notes and the\nsame brief. It is not quite like for like: the Sonnet readers are helpers of the app with a shell\nand the search by meaning, and the corpus graph's reminders reach them.\n\n**Written before any of the four runs.**\n\n1. On the brief for what does not fit, Haiku returns at least half as many findings that stand on\n   reading as the Sonnet reader does.\n2. Its quoted words stand in the files for at least nine in ten passages.\n3. On the open question, Haiku returns at least three remarks that stand and that the Sonnet\n   reader did not make.\n\n**Against Haiku as a reader**: fewer than a third as many standing findings as Sonnet, or more\nthan a quarter of its findings failing when both ends are read.\n\n**What came back.**\n\n- *Quotes, predicted nine in ten.* **Held**: 20 of 20 passages in the report on what does not fit,\n  and 23 of 23 in the report on the open question, stand at the lines given.\n- *The open question, predicted at least three remarks that stand and that the Sonnet reader did\n  not make.* **Held, with a loss**: the first five of its thirteen remarks never arrived, because\n  the tool returns only a reader's last message and this one wrote two. Of the eight that came,\n  three are also in the Sonnet report and five are not.\n- *What does not fit, predicted at least half as many standing findings as Sonnet.* **Held**: Haiku\n  returned ten and Sonnet fourteen, and all twenty-four stand when both ends are read. They share\n  one sentence and one fault seen from two sides, so the two reports are nearly disjoint. Haiku took\n  six minutes and 43 steps where Sonnet took twenty-four minutes and 180.\n- On the open question Sonnet returned sixteen remarks to Haiku's thirteen, and three of Sonnet's\n  are faults found by reading the code beside the page. Haiku read code too: it reports the feed's\n  two-vote floor, and that the pass for unwritten premises has no check for a sacred commitment. It\n  did not see that the feed reads the links leaving a claim.\n\nThe reading of § 5 moves one step: with the briefs of the pilot, **Haiku is a usable reader\nfor a pass of its own**, and the two models' reports overlap so little that a second reading by the\nother model is worth more than its price.\n\n**Two things to change before the next run through the tool**: keep every message of a reader and\nnot only the last, and give it the search by meaning, which the helpers of the app have and it\ndid not.\n\n"}