Codebook, version 2

How each coded unit is coded. Written 2026-09-30 and revised 2026-10-01 (amendment A14), 2026-10-12 (amendment A28) and 2026-10-18 (amendment A29, tip intake) from docs/methodology-v2.md section 2, which is the authority; this file turns its rules into decisions with examples. Changes go in docs/amendments.md first.

Version 1 (the single stance code and the six-value question_frame) is retired. Rows coded under version 1 were recoded from their sources, not mapped (amendment A2).

About the examples. Examples marked constructed are invented to show a rule and must never be cited as quotes.

0. One unit, and how to read it

A unit is one statement by one speaker on one occasion; one occasion can give several units, since each distinct passage of an essay, talk or interview is its own row (decided 2026-10-18; amendment A28). Code only what the quoted words and their immediate context say. Don't use what the speaker is known for, what they said elsewhere, or what the article around the quote implies.

Read the source, not just the stored quote. The stored quote (60 words or fewer) is evidence for the codes. If a code depends on a sentence outside the stored quote, either change the quote to include it or code unexpressed / ambiguous.

1. target_capability: what capability is the unit about?

Choosing among them. Pick the capability the speaker's claim is about. If they name two (for example "human-level by 2029, Singularity by 2045"), code the one the unit leads with and record the other in definition_note.

The four self-improvement levels. Ask two questions in order:

1. Who does the improving? Humans with AI help → si_assist. The AI itself → go on.
2. How far does it go?
   • A part of the research pipeline, with humans still steering → si_research.
   • The system edits itself (its own code or weights) → si_modify.
   • Repeated rounds, each better system making the next, with little human role → si_sustained.

"Takeoff" and loose self-improvement words (ruling 12, A14). Code the lowest level the words support and set ambiguous: yes; put any speed the speaker gives ("slow", "fast", "gentle") in timeline_text. The lowest level depends on the answer to question 1 above:

• When the words don't say who does the improving (a bare "takeoff" or "hard takeoff", "FOOM", "large jumps in intelligence"), the lowest level is si_assist (A15 ruling 3 confirms this for a bare "takeoff").

• When the words do say the AI improves itself ("the AI will improve itself", "a machine improves itself", "a self-improving system", "improve their own capabilities", "recursive self-improvement"), question 1 is answered and the lowest level those words support is si_modify. Decided 2026-10-12 (A28): an earlier version of this paragraph listed "the AI will improve itself" among the words that don't say who improves, which contradicted question 1.

• Code si_sustained only when the words themselves describe repeated rounds with little human role ("each generation builds the next", "an intelligence explosion with no one in the loop", "reprogram itself to be smarter and iterate ... 24 hours a day"). The labels "recursive self-improvement" and "intelligence explosion" alone don't do that: they say the system improves itself, so si_modify.

• Rejections and doubts. When the speaker rejects or doubts runaway growth ("there will not be the run away exponential growth", "intelligence explosion is impossible", "a hard takeoff scenario is very unlikely"), keep the level the speaker is rejecting (the words describe it) rather than the lowest level. Coding si_assist would turn a rejection of runaway growth into a rejection of AI helping people build AI, which the speaker never said. Decided 2026-10-12 (A28).

• Seed example, altman-2023-04: Altman told the New York Times it would be a "very slow takeoff". The words don't say who improves the systems or how, so target_capability: si_assist, ambiguous: yes, timeline_text: "very slow takeoff".

• Constructed: "Our engineers now write most of their code with model help." → si_assist.

• Constructed: "The model ran the ablation studies overnight without anyone." → si_research.

• Constructed: "Once a system can rewrite its own weights, all bets are off." → si_modify, present_capability: hypothetical.

• Constructed: "Each generation will design the next, faster than we can follow." → si_sustained.

2. frame_primary and frames_present: which questions does the unit raise?

• frame_primary is the question the unit leads with: what the speaker is mainly answering, usually what they were asked or what the headline claim is about.
• frames_present lists every question raised, separated by ;, and always includes frame_primary. A question counts as present when the unit raises it, whatever answer is given ("it's not close" raises capability_now).
• concern_tag is required when other_concern is present (one of jobs, reliability, privacy, bias, misinformation, military, consciousness, ownership, other), and blank otherwise.
• A unit framed as self-improvement should normally have an si_* target; if it doesn't, say why in notes.

Test for possibility against timeline: delete every date from the unit. If the claim still stands and is about whether it can be done, it's possibility.

3. endorsement: does the speaker endorse that the target capability will exist?

Endorsement is about the future: whether the capability will exist, ever or by the horizon the speaker names (ruling 9, A14). Whether it exists now is a different field, present_capability.

Silence is never acceptance. Endorsement is coded only from words the speaker actually said. Every code other than unexpressed needs an anchor: the exact words in the stored quote that carry the position, recorded in notes as basis: "<words copied from the quote>". scripts/validate.py checks that the anchor appears in the quote. If you can't point to such words, the code is unexpressed.

Rulings of 2026-10-01 (A14), with examples from the seed data:

• Not here yet (ruling 9). Denying that the capability exists now is present_capability: disputed. Endorsement then comes from what the speaker says about the future, or is unexpressed if they say nothing about it. Seed example, lecun-2017-02: "We're very far from having machines that can learn the most basic things about the world in the way humans and animals can do." This denies present capability and says nothing about whether such machines will come, so present_capability: disputed, endorsement: unexpressed (it was rejects before the ruling).
• Hedged dates (ruling 10). "May", "might" or "could" with a date is uncertain. Fill the timeline fields anyway and keep the hedge in timeline_text. When a hedge and an acceptance word govern the same date, the hedge decides (agent interpretation, noted in A14). Seed example, hinton-2023-06: "Now, I think we may be much closer, maybe only five years away from that." → endorsement: uncertain (basis "maybe only five years away"), timeline_type: qualitative, timeline_year: 2028 (computed: 2023 + 5), timeline_text: "maybe only five years away".
• "Decades away" (ruling 11) is accepts with timeline_type: qualitative; leave timeline_year blank unless a number of decades is given. Seed example, marcus-2022-04: "we are still likely decades away from general-purpose, human-level AI" → endorsement: accepts (basis "still likely decades away"), present_capability: disputed, timeline_type: qualitative, timeline_text: "still likely decades away", and ambiguous: yes for the "likely" (A28).
• Self-improvement words (ruling 12): see section 1.
• "Probably", "likely" and "perhaps" (A28, decided 2026-10-12). A position resting on one of these words is accepts with ambiguous: yes (field: endorsement), whether the word governs the arrival claim ("it's likely that machines will be smarter than us") or its timing ("still probably decades away", "likely before too long"). If "may", "might" or "could" governs the same claim, ruling 10 wins and the code is uncertain. Agent reading (2026-10-12): the rule covers a positive position; a negative one ("unlikely", "perhaps we are just not smart enough") is not an acceptance and keeps its own code (rejects or uncertain), and a word that governs something other than the claim ("algorithms, likely yet to be invented") does not trigger the rule. Seed example, marcus-2013-01: "It's likely that machines will be smarter than us before the end of the century" → endorsement: accepts (basis "It's likely that machines will be smarter than us before the end of the century"), ambiguous: yes.

Hard cases:

• Giving a date usually carries endorsement, but check the words. "AGI by 2029" with "will" or "I think" anchors accepts. "People say 2029; who knows" is uncertain (basis "who knows") or unexpressed.
• Worrying about something is not endorsing it. "If superintelligence arrives it could end us" is conditional at most, and unexpressed if no arrival claim is made. Writing about regulation does not establish that the author accepts feasibility.
• Rejecting a near date is not rejecting the capability. "Not by 2027" → rejects for that horizon only; record the year with the not_by tag (section 5) and say in notes whether the speaker accepts it eventually. If the same unit also says "but it will come eventually", code accepts with the hedge in timeline_text; set ambiguous: yes if the two are balanced.
• Rejecting a route is not rejecting the capability (A15 ruling 4). "Language models are not a road to AGI" says nothing about whether AGI will exist. Code unexpressed, and keep the route claim in claim_summary and notes. Seed example, lecun-2024-02: "They're not a road towards what people call 'AGI.'" → endorsement: unexpressed.
• Minimum years are acceptance (A15 ruling 5). "At least a decade" or "at least thirty or fifty years" asserts arrival with a minimum: accepts, timeline_type: range with timeline_year set to the minimum, timeline_year_high left blank, the lower_bound tag, and "at least" kept in timeline_text. Seed example, marcus-2016-01: "aren't going to be for at least thirty or fifty years" (2016) → accepts, range, timeline_year: 2046, high bound open.
• Probabilities. A stated probability of 50% or more by a date → accepts with timeline_prob filled (ruling 10). Below 50%, or a wide spread stated as such → uncertain.
• Jokes and sarcasm → unexpressed and ambiguous: yes, unless the speaker explains the joke.

One row per capability (A15 ruling 13)

When one statement makes distinct claims about distinct target capabilities, it gets one row per capability. The rows share the quote, source and date; their ids carry a one-letter suffix (marcus-2026-04a, marcus-2026-04b); each row is coded for its own capability.

Seed example 1, marcus-2026-04: "it's great for example at (some) code optimizations, but not even (as the blog makes clear) actual RSI, which remains speculative." Two claims:

• marcus-2026-04a: AI helping build AI exists → target_capability: si_assist, present_capability: demonstrated, endorsement: accepts (basis "it's great for example at (some) code optimizations").
• marcus-2026-04b: recursive self-improvement does not exist yet → si_sustained, present_capability: disputed, endorsement: unexpressed (nothing said about the future).

Seed example 2, altman-2025-06: "We are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence…"

• altman-2025-06a: "the takeoff has started" → the lowest self-improvement level the words support (si_assist), present_capability: claimed, ambiguous: yes.
• altman-2025-06b: superintelligence → frame_primary: timeline, endorsement: accepts, timeline_type: qualitative, timeline_text: "close".

4. present_capability: does it exist now, in the speaker's view?

na exactly when target_capability is none.

5. Timeline fields

As in version 1, with three tags (in notes) for claims that aren't arrival dates:

• not_by: YYYY for "it won't happen by YYYY". Leave timeline_year blank and code timeline_type: qualitative. These are not arrival dates and charts must not plot them as such.
• lower_bound for "at least N years" or "no sooner than YYYY". Fill timeline_year with the bound and add the tag. With timeline_type: range, the high bound may be left blank (open-ended minimum, A15 ruling 5).
• Proximity (distant, medium, near, present) is worked out when charts are drawn, never stored.

6. Consequences, affect and agency

These are separate on purpose. Someone who expects AGI in five years can be thrilled, frightened, resigned or organizing.

• Constructed: "AGI by 2030, and frankly it terrifies me, but good policy can still steer it." → accepts (basis "AGI by 2030"), affect: fear, agency: can_influence.
• Constructed: "It's coming whether we like it or not." → accepts (basis "It's coming"), affect: resignation, agency: cannot_influence.

7. elicitation: how did the statement come about?

Secondhand rows are usually quoted.

8. Other fields

• definition_note: what the speaker means by AGI in this unit, if they say.
• coding_confidence: high when the codes are clear and the source is primary; medium when one needed judgment; low when several did, or the source is secondhand.
• ambiguous: yes when the words support two codes equally. Kept as a code, never resolved by guessing. Say which field in notes.
• date must match date_precision: YYYY-MM-DD, YYYY-MM or YYYY.
• quote: verbatim, 60 words or fewer, "..." only within one passage.
• Primary and secondhand. Primary = the speaker's own words in something they produced or took part in (their writing, a transcript, a recorded talk, an interview with direct quotes by the outlet that did it). Secondhand = someone else's report or paraphrase, or a quote not traced to where it was said. Secondhand rows carry the secondhand tag and coding_confidence: low.
• coder: john, claude-agent (version 1, retired), claude-agent-v2, claude-agent-v2.1 (rows coded or changed under the A14 rulings, including the full panel collection), or claude-session-frontier-traces-2026-10-17 (the frontier-lab traces of Musk, Amodei, Hassabis and Sutskever folded in from pull request 13, coded under this codebook unchanged).

9. Notes tags

Notes start with any of these tags, separated by ; , then free text:

10. Mapping to Fast and Horvitz 2017

Methodology-v2 section 5 asks for the categories of Fast and Horvitz's 2017 study of New York Times AI coverage to be reconciled with these codes before the media stream starts. That reconciliation belongs to the media stream and has not been done yet.

Rendered from docs/codebook.md at version a7bd17a.