{"path":"research/quality-control-for-deeper-steps/the-three-briefs.md","content":"# The briefs the three readers were given (2026-10-10)\n\nKept so that a later reading can be compared with this one. Unchanged, except that the machine's own path\nto the repository is cut to the path inside it.\n\n---\n\n## How the small tasks are checked\n\nYou are a read-only reader in the repository the repository (a research corpus plus a Python codebase for \"Deliberus\", a platform that maps arguments as a graph of claims and links). Thoroughness: very thorough. Do not edit or create any file, do not call any model, do not run anything that writes. You may run read-only commands (rg, sed, cat, and `uv run python scripts/corpus_graph.py around <note-name-or-ledger-row-id>`).\n\nBACKGROUND. Read this note in full first: docs/research/the-chain-inside-a-link.md. It designs how the machine writes out what lies under a link between two claims (the intermediate steps of reasoning), level by level. Today the founder ruled on it, verbatim: \"It may and should write deeper steps. But the microtasks in charge of doing so need to be quality-controlled, just like the other microtasks that are in play in Deliberus.\" (ledger row AG6 in docs/research/session26-open-rulings.md). So the open question is: WHAT IS the quality control that the project's existing small machine tasks (\"microtasks\", \"passes\") are actually under, and how does each control carry over to tasks that write a second, third, fourth level of steps under a link?\n\nYOUR PART: the existing small tasks and how each is checked. Establish facts, with file and line, separating four states for every control: BUILT (code exists and runs), MEASURED (a number exists, say what it measured and on how many items), RULED (the founder decided it, quote his words), PROPOSED (an open ledger row or a note's proposal).\n\nRead these in full (say how far you got in each):\n- docs/research/microtasks-in-broad-daylight.md\n- docs/research/proving-a-pass-reliable.md\n- docs/research/the-evals-loop-and-our-machine-passes.md\n- docs/research/pass-quality-and-the-choice-of-model.md\n- docs/research/earning-the-right-to-write.md\n- docs/research/machine-work-as-a-standing-property.md\n- docs/research/the-parts-review.md\n- docs/research/split-review-and-the-founders-intentions.md\n- docs/research/storing-the-record-of-machine-judgments.md (sections 7 and 7a at least)\n- docs/research/auto-connect-upgrade.md (section 2c)\n- docs/research/writing-out-the-unwritten.md (sections 7, 8, 9, 10) and docs/research/writing-out-the-unwritten/outside-reading-dependable-small-tasks.md\n- CLAUDE.md, the section \"Hard-won gotchas\" and the section \"The Principles in One Breath\"\n- ledger rows T1 to T7, Y1 to Y14, R8, R12, Z1, Z1b, Z11, AD13 to AD21, AF2, AF4, AG1 to AG6 in docs/research/session26-open-rulings.md (read each row whole)\nAnd the code (read the docstrings and the functions that check or gate): deliberus/proposed_split.py, deliberus/extraction/decompose_claim.py, deliberus/keeper.py, deliberus/relevance_judge.py, deliberus/miscast_parts.py, deliberus/judgments.py, deliberus/extraction/auto_connect.py (the gate and the keeper call), deliberus/support_semantics.py (the presumption weights), scripts/recut_splits.py, scripts/judgment_facts.py, scripts/read_review_notes.py, scripts/measure_quote_fidelity.py, scripts/recheck_report_attribution.py, tests/test_reason_before_verdict.py, tests/conftest.py (the guard against a test reaching a model).\n\nBefore opening each note, run `uv run python scripts/corpus_graph.py around <note-name-without-.md>` and note what it tells you.\n\nDELIVER, as your final message (it will be saved word for word, so make it self-contained, plain, no preamble):\n1. A table \"The small tasks in play and how each is checked\": one row per existing task (at least: the sameness keeper; the splitter; the splitter's coverage check; the judge of parts; the relevance judge; the scheme classifier; the linking step auto-connect; extraction's claim passes; the title writer; the reported-speech check; the outside reader of proposals if you find it in docs/research/outside-checks/README.md). Columns: what it decides or writes · checks by code · checks by a second, separate call · what is recorded about each call · what a person has read or labelled (how many) · measured figures · where it stands (proposing / writes marked / trusted) and by whose word · file:line for each fact.\n2. A list \"Controls that exist, by kind\", each with BUILT / MEASURED / RULED / PROPOSED and file:line: e.g. reasoning before verdict; a refusal asked first as a gate; asking a second way and counting who changes side; both orders; a separate checker not shown the draft; a check by code of names and numbers; write marked and reversible; hold what failed its check; undo; record of every call; a person reads before a rate is proven; the review page and its note box; sealed part of a labelled set; random draws in every sitting; second model family; admission of a model in four steps; depth capped by measured error; version of a pass by fingerprint of its instructions.\n3. For each control in 2: does it carry over to a task that writes a DEEPER step under a link (a step whose inputs are themselves machine-written sentences)? Answer yes / with a change / no, one sentence why. Name any control the deeper task needs that no existing task has.\n4. \"Contradictions and gaps\": places where two documents or a document and the code disagree about how a task is checked, with file:line on both sides.\n5. \"What the corpus graph already held\": for each note you opened, one line: what `around` told you that mattered, or \"nothing\".\nKeep it under about 2,200 words. Quote sparingly and exactly, with file:line. Mark anything you infer and did not read as \"inferred\".\n\n---\n\n## Depth, chains, error and volume\n\nYou are a read-only reader in the repository the repository (a research corpus plus a Python codebase for \"Deliberus\", a platform that maps arguments as a graph of claims and links). Thoroughness: very thorough. Do not edit or create any file, do not call any model, do not run anything that writes. You may run read-only commands (rg, sed, cat, and `uv run python scripts/corpus_graph.py around <note-name-or-ledger-row-id>`).\n\nBACKGROUND. Read this note in full first: docs/research/the-chain-inside-a-link.md. It designs how the machine writes out what lies under a link between two claims (the intermediate steps of reasoning), level by level. Today the founder ruled on it, verbatim: \"It may and should write deeper steps. But the microtasks in charge of doing so need to be quality-controlled, just like the other microtasks that are in play in Deliberus.\" (ledger row AG6 in docs/research/session26-open-rulings.md). An earlier ruling (ledger T1, 2026-09-29) says: a pass with a measured error rate keeps the same share of wrong content at any volume; only machine work built on machine work compounds, so the measurement caps how deep that building may go. An earlier ruling still (2026-08-25, docs/research/kartpaus-auto-enrichment-design.md) has the machine work out every claim's descent eagerly, recursively, fed by research, into a draft stage the default map does not show, stopped where nothing more could change what a reader sees.\n\nYOUR PART: depth, chains, error and volume. Find every passage in the corpus that bears on what happens when a machine writes a step, then a step under that step, and so on: how error grows with depth; how a long chain of steps that each look sound can hide an error; how strength passes or fades along a chain; when splitting something finer makes checking better and when worse; what stops a machine from writing forever; how much a deeper level costs in calls and in claims; what was measured about chains written in one go.\n\nRead these (say how far you got in each; read the cited sections in full):\n- docs/research/proving-a-pass-reliable.md (whole)\n- docs/research/major-transitions-heredity-and-evolvability.md (the sections on error threshold, and §§ 5, 5b, 7, 8)\n- docs/research/convergence-wager-red-team.md and docs/research/red-team-synthesis-2026-07.md (obfuscated arguments; stability when a decomposition is made again)\n- docs/research/weakest-link-arithmetic-and-the-merge-hunch.md (§§ 0, 1, 7, 9, 9d, 9e)\n- docs/research/the-squirrel-and-the-strength-calculus.md (whole) and docs/research/strength-calculus-literature/ (the files there; find what they say about chains and about impact fading)\n- docs/research/decomposition-axes.md (§ 2b and § 4) and docs/research/the-jhu-decomposition-line.md and docs/research/reading-the-text-itself.md\n- docs/research/scheme-bounded-decomposition-and-evidence-as-subgraph.md and docs/research/self-similar-decomposition-and-claim-ontology.md\n- docs/research/first-pipeline-run-analysis.md, docs/research/extraction-virtual-run.md, docs/research/extraction-improvement-research.md (bundled claims; paraphrase over several hops)\n- docs/research/kartpaus-auto-enrichment-design.md and docs/research/active-inference-context-acquisition-and-deliberus.md (the stopping rule) and deliberus/question_ordering.py and deliberus/hinge.py (docstrings)\n- docs/research/derivability-of-missing-considerations.md, docs/research/drowning-in-claims.md, docs/research/saturation-and-the-long-tail.md, docs/research/latent-scaffolding-and-the-type-token-split.md\n- docs/research/the-reuse-flywheel-prior.md, docs/research/graph-daemons-design-space.md, docs/research/the-analogy-daemon.md\n- docs/research/how-far-a-premise-is-from-its-claim.md, docs/research/what-a-link-rests-on.md, docs/research/writing-out-the-unwritten.md (§§ 6, 8, 9, 10), docs/research/philosophy-of-the-unwritten-step.md (§§ 4, 7, 10), and the two outside-reading files in docs/research/writing-out-the-unwritten/\n- the \"Standing Epistemic Threat Model\" section of CLAUDE.md\nAlso search the whole of docs/ with rg for: \"depth\", \"compound\", \"chain\", \"multi-hop\", \"obfuscat\", \"re-decompos\", \"attenuat\", \"fades\", \"second-order\", \"error rate\", \"per step\", and open any note not listed above that has a passage on the question.\n\nBefore opening each note, run `uv run python scripts/corpus_graph.py around <note-name-without-.md>` and note what it tells you.\n\nDELIVER, as your final message (it will be saved word for word, so make it self-contained, plain, no preamble):\n1. A table of bearings. Columns: corpus passage (file:line, with a short exact quote) · which sentence or section of the-chain-inside-a-link.md, or which ledger row (AG1 to AG6, T1, AD19, AD29), it bears on · does that note already relate it? (grep the note; yes / partly / no) · what it would change in a design where the machine writes deeper steps under quality control. Aim for the 15 to 25 bearings that matter most, the unrelated ones first.\n2. \"Every number about depth and chains that the corpus holds\", each with file:line, what was measured, on how many items, by whom (a person, a model, code), and what it does not show.\n3. \"What must be true for writing deeper to be safe\", as a short list of conditions that the passages you read support, each with its passage.\n4. \"Contradictions\": places where passages disagree with each other, or with the ruling of today, or with the note, file:line on both sides.\n5. \"What the corpus graph already held\": for each note you opened, one line: what `around` told you that mattered, or \"nothing\".\nKeep it under about 2,200 words. Quote exactly, with file:line. Mark anything you infer and did not read as \"inferred\".\n\n---\n\n## People, showing, and what checks cannot see\n\nYou are a read-only reader in the repository the repository (a research corpus plus a Python codebase for \"Deliberus\", a platform that maps arguments as a graph of claims and links). Thoroughness: very thorough. Do not edit or create any file, do not call any model, do not run anything that writes. You may run read-only commands (rg, sed, cat, and `uv run python scripts/corpus_graph.py around <note-name-or-ledger-row-id>`).\n\nBACKGROUND. Read this note in full first: docs/research/the-chain-inside-a-link.md. It designs how the machine writes out what lies under a link between two claims (the intermediate steps of reasoning), level by level, as questions with moves beside them, hidden by default and shown when somebody contests. Today the founder ruled on it, verbatim: \"It may and should write deeper steps. But the microtasks in charge of doing so need to be quality-controlled, just like the other microtasks that are in play in Deliberus.\" (ledger row AG6 in docs/research/session26-open-rulings.md).\n\nYOUR PART: people, showing, and what checks cannot see. Find every passage in the corpus that bears on (a) who checks machine-written steps and whether their checking can be trusted or decays; (b) what happens to a reader who is shown a machine's reasoning (anchoring, deference, over-trust, entrenchment, disengagement); (c) which kinds of error no check by a machine or by a sample can see (what the corpus omits; the frame it was built inside; one model family all the way down); (d) how a machine judge can be moved by wording or pressure, and what that means for a machine checking a machine's deeper steps; (e) when writing the unsaid is harmful or unwanted (things left unsaid on purpose; sacred matters; a person's own steps); (f) how machine work is shown on the map and what the founder has ruled about showing it.\n\nRead these (say how far you got in each; read the cited sections in full):\n- docs/research/what-human-judgment-is-for.md\n- docs/research/the-ratification-record-and-the-triage-feed.md\n- docs/research/cognitive-commons-and-deliberus.md\n- docs/research/fragile-checkers-and-the-verification-bottleneck.md\n- docs/research/the-scrutiny-gap.md\n- docs/research/steadying-an-outside-judge.md\n- docs/research/the-residual-error-taxonomy.md\n- docs/research/the-harness-conviction.md\n- docs/research/does-descending-change-the-descender.md\n- docs/research/the-load-bearing-unsaid.md\n- docs/analysis-and-attunement.md and docs/research/parity-and-the-zone-interior.md (the sacred ruling)\n- docs/research/engagement-gradient-priors.md, docs/research/curiosity-as-growth-fuel.md, docs/research/incentives-analysis.md\n- docs/research/the-tidying-line.md (§ 4 and what follows) and docs/research/earning-the-right-to-write.md\n- docs/research/understanding-after-the-proof-machines.md, docs/research/game-feel-and-the-first-screen.md, specs/map-view/design.md, docs/ux-principles.md\n- docs/research/store-the-uses-read-the-meaning.md, docs/research/what-counts-as-evidence.md (the parts on independence, shared origin and on a located encounter with the world)\n- docs/research/the-motte-and-the-periphery.md, docs/research/grounded-critics-and-quarantined-imagination.md, docs/research/bengio-safety-from-honesty-and-deliberus.md\n- docs/research/what-a-rule-rests-on.md § 13 and § 14, and the three reports in docs/research/what-a-rule-rests-on/older-corpus-bearings-*.md\n- the \"Standing Epistemic Threat Model\" section and the \"Principles in One Breath\" section of CLAUDE.md\nAlso search the whole of docs/ with rg for: \"anchor\", \"defer\", \"over-trust\", \"same family\", \"same-hand\", \"rubber\", \"ratification\", \"left unsaid\", \"uninvited\", \"show on contest\", \"wall of text\", and open any note not listed above that has a passage on the question.\n\nBefore opening each note, run `uv run python scripts/corpus_graph.py around <note-name-without-.md>` and note what it tells you.\n\nDELIVER, as your final message (it will be saved word for word, so make it self-contained, plain, no preamble):\n1. A table of bearings. Columns: corpus passage (file:line, with a short exact quote) · which sentence or section of the-chain-inside-a-link.md, or which ledger row (AG1 to AG6), it bears on · does that note already relate it? (grep the note; yes / partly / no) · what it would change in a design where the machine writes deeper steps under quality control. Aim for the 15 to 25 bearings that matter most, the unrelated ones first.\n2. \"What a person's check of a machine-written step can and cannot be relied on for\", each point with its passage and, where the corpus has one, its number.\n3. \"What no check sees\", as a short list, each with its passage, and for each: what the corpus says is the only remedy.\n4. \"The founder's own words that bear on this\", verbatim with file:line and date, most recent first (rulings and positions on machine scaffolding, on showing machine work, on who must speak first, on what must never be done uninvited).\n5. \"Contradictions\": places where passages disagree with each other, or with the ruling of today, or with the note, file:line on both sides.\n6. \"What the corpus graph already held\": for each note you opened, one line: what `around` told you that mattered, or \"nothing\".\nKeep it under about 2,200 words. Quote exactly, with file:line. Mark anything you infer and did not read as \"inferred\".\n"}