Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript
35.0 KB
Raw Blame

Q2 Track C: typing units on development data

U1 (adviser wording) and the joint unit U34 (U3 board wording never types works_at, U4 advisory, board and investor roles are not employment starts) ship: the preregistered held-out decision confirmed the package P2 = {U3, U4, U1} (afcec1ad), and ENABLED_TYPING_UNITS is [U3, U4, U1]. U6 failed the safety conditions at held-out selection on I1 and W1 (now-recall −0.012, as-of exact −0.011, one new wrong transition by identity), and U25 (U2 with U5) added nothing on the confirmation set (difference 0.000 on all 200 pairs). Following the preregistration, the U2, U5 and U6 code was removed before landing, and the landing head types identically to afcec1ad (see "Landing equivalence").

The rest of this report is the development record written before the verdict: all six units were measured alone and together on development data. After reworks, no unit had a development type steal or a new wrong transition by identity, and the all-unit package removed every E5 false start and every wrong single-value closure while adding 73 current employers to the live read. The development data showed two dependencies, recorded as joint units: U34 (U4 needs U3) and U25 (U2 needs U5). Development gains did not predict held-out ones for U2, U5 and U6.

Landing equivalence

The landing head (the U2, U5 and U6 code, their examples and the U25 alias removed; U1, U3, U4, U34, the rule ids, explainLinkType and the package script kept) was compared with the confirmed package afcec1ad, both at their built-in ENABLED_TYPING_UNITS = [U3, U4, U1], with zero differences:

  • world-v1 (scripts/q2-typing-dev.ts world-v1, every page's extracted edges, tense rows and transitions): 723 lines, SHA-256 1763359f75ac050fc19e8fdd17a6694a8cf0e6a7fbd12c2b01ef5932f9ef96fd on both, byte-identical files. The digest is committed in test/fixtures/q2-typing-units/world-v1-confirmed-package.sha256 and checked by test/link-typing-units-subsets.test.ts.
  • Development dumps (scripts/q2-typing-dev.ts dump, pages written through put_page on in-memory PGLite; every person page's link rows with type, source, tense and context, its link_transitions rows, and its default get_links read), compared row for row and with compare by identity: 0 differing pages, 0 steals, 0 gains, 0 transition differences on every phrasing.
Phrasing (seeds 3, 5) Person pages Link rows Transitions Pages that differ
A 232 1,024 1,056 0
A2 232 987 978 0
A3 232 847 830 0
role-for 232 848 571 0
new-chapter 232 869 464 0
paren-move 232 825 777 0
signed-on 232 839 734 0
adviser 232 1,024 1,004 0
board-investor 232 1,024 978 0
board-director 232 990 982 0
leave-start-idioms 232 1,003 464 0
onboarding 232 894 464 0

src/core/link-extraction.ts and src/core/link-temporal-evidence.ts on the landing head are byte-identical to their state at 0044223c (after U4, before U2), which is also why the outputs match: the removed code was reachable only through typingUnitEnabled('U2' | 'U5' | 'U6'), which the package never enabled.

Development data only: temporal-edges dev seeds 3 and 5 with phrasing sets A, A2, A3 and the nine test/fixtures/q2-dev-phrasings/ files; world-v1 (H1 runner); relation-line variants dev seeds 1-3 (H2 runner). No custodian or sealed text was read. Mechanism background: dev-trace.md.

Units, rules and commits

Unit Commit(s) Change Dev rework
U1 adviser wording c1b3233f, 61934b56 adviser spelling, "advising [X]", local negation, third-party veto adviser phrase on another entry (window edge) vetoed: it caused 6 steals on signed-on
U2 ordinary roles 1b5a8569 post-pass where inference returned mentions: " for", "led at", "[X] ()", "as " after a join undated lines only; not on a page that names an advisory/board/investor role; not for an organization two or more dated entries mention
U3 board wording dd05898d board/observer/investor wording vetoes works_at; board seat keeps invested_in only with an investor prior or a stated investment veto scoped to the link's own sentence/entry or a board position right before the link (first version stole jobs a neighboring entry's verb had typed)
U4 not-employment starts 0044223c NOT_EMPLOYMENT_ROLE in the "became … at/of" and "took … role" start cues none; joint with U3 as U34 (30717952)
U5 leave idioms fe4c334e, 5a9efd08 split-object and quitting idioms, "exit from", exchange moves; dated ends only restart guard (from the trace): an idiom end is dropped when a later dated entry names the organization in words no cue reads; share-trade lines ("Traded shares of [X]") no longer count, because they made the guard drop correct ends (investment traps 17/26 → 16/26 on leave-start-idioms before the fix, 26/26 after)
U6 start framings fd0b1f55, c6655db2 first day/week, day one, kick-off with a job noun, onboarded/onboarding week, began working at, new chapter at leave guard (mirror of U5's) and the advisory-page guard: alone it had opened former jobs for good (6-8 wrong general single-value closures, 5-24 wrong E5 closures, lower traps)

Infrastructure: efd1628e (rule ids, explainLinkType), 192c0870 (unit plumbing, example harness), 30717952 (package script, U34 alias, 64-subset and world-v1 identity tests), 49b1ae6e (entailment table and recipe).

Dependencies the development data shows

  • U4 needs U3 (set-F interaction, confirmed). On the board-director phrasing, board lines that master types works_at lose their start under U4 and read as a job on every date: as-of exact 0.946 (master) → 0.783 (U4 alone), during-F1 0.931 → 0.890. U3 alone gives 0.950, U3 with U4 0.950. U4 alone would fail the as-of safety gate wherever board wording types works_at; U3 alone does not need U4. The freeze record should declare U34 joint, with U4's primary metric (preregistration, "What is decided").
  • U2 and U6 expose master's advisory-line start cues. A newly typed (U2) or newly dated (U6) employer meets the false works_at start that master's "became … at" / "took … role" cue reads on an advisory line about another target (E5 "other" form); the single-value pass then closes the real employer on the advisory date, a new wrong closure by identity. U4 removes the cause. I kept U2 and U6 independent of U4 with a page-level guard (they do not fire on a page that names any organization in an advisory, board or investor role); the cost is that a person with an advisory role elsewhere gets no U2 typing or U6 dating. Kept as built (decision of 2026-10-06).
  • Count-based E5 metric hides wrong starts. e5_extra_works_at_starts is zero-clamped and counts starts: on master a missed correct start masks a wrong advisory start, so a unit that recovers the correct start shows a "new" extra start without adding a wrong transition. The identity tables below are the faithful measure (plan §9.3).

Dependency units

The preregistration amendment records two joint units, each measured with the primary metric of its first member.

U34 = U3 with U4 (U4's primary metric, false employment starts per E5 probe person). On board-director, U4 alone takes as-of exact from 0.946 to 0.783 and during-F1 from 0.931 to 0.890: board lines that master types works_at lose their start and read as a job on every date. U3 removes the works_at typing; U3 with U4 gives 0.950 and 0.940.

U25 = U2 with U5 (U2's primary metric, live-edge recall). U2 types the current employer in "[X] ()" and "signed on with [X] as " prose. When the former employer's leave is an idiom only U5 reads ("Exit from [Y]", "Handed in her notice at [Y]"), master keeps the former employer live, and with a dated start for the newly typed employer the general single-value pass closes the former employer at that start date instead of its real end: a new wrong closure by identity (by date). U5 dates the real leave, so the relationship is already closed and the pass has nothing to close.

Phrasing Arm live recall as-of exact now-precision investment traps alumni traps general single-value wrong closures
paren-move none 0.597 0.640 0.793 22/26 18/20 3/160
paren-move U2 0.734 0.727 0.858 22/26 18/20 4/160
paren-move U5 0.597 0.671 0.958 26/26 20/20 0/160
paren-move U25 0.741 0.760 1.000 26/26 20/20 0/160
signed-on none 0.583 0.579 0.737 22/26 18/20 3/160
signed-on U2 0.727 0.673 0.806 22/26 18/20 4/160
signed-on U5 0.583 0.612 0.874 26/26 20/20 0/160
signed-on U25 0.734 0.708 0.933 26/26 20/20 0/160
role-for none / U2 / U25 0.640 / 0.727 / 0.727 0.637 / 0.683 / 0.683 0.934 / 0.979 / 0.979 26/26 20/20 0/160
new-chapter none / U2 / U25 0.597 / 0.741 / 0.741 0.535 / 0.623 / 0.623 0.755 / 0.816 / 0.816 19/26 16/20 0/160

By identity, U25 against no unit: no steals, no new wrong transitions, 12-21 current employers added to the live read on role-for, new-chapter, paren-move and signed-on, and 44-165 correct transitions added (U5's ends). U25's live-edge recall equals or exceeds U2's on every phrasing (the extra 0.007 on paren-move and signed-on is an employer U5's end/start pair now keeps live), and on the eight phrasings without these role shapes it changes nothing U5 alone does not. The package of the amended family, {U1, U25, U34, U6} = all six units, is the "all" arm in the tables below and passes the frozen examples (test/q2-typing-package.test.ts); bun scripts/q2-typing-package.ts --base <ref> --arm U25 and --units U25,U1,U34,U6 build its refs.

Identity tables (all phrasings, both seeds, arm vs no unit)

Steal: a person-company pair that lost a ledger type or gained more unsupported types than it shed. Transitions by identity (subject, target, type, kind, date) against the generator ledger; dumps written through put_page without the single-value pass (scripts/q2-typing-dev.ts).

arm steals gains other moves new wrong transitions wrong transitions removed correct transitions added correct transitions lost current employers lost from live read added to live read
U1 0 176 21 0 0 72 0 0 0
U2 0 73 0 0 0 27 0 0 73
U3 0 0 28 0 28 0 0 0 0
U4 0 0 0 0 570 0 0 0 0
U34 0 0 28 0 570 0 0 0 0
U5 0 0 0 0 0 331 0 0 0
U25 0 73 0 0 0 358 0 0 73
U6 0 0 0 0 0 61 0 0 0
all 0 249 49 0 570 718 0 0 73

Correct works_at transitions by identity, starts · ends · wrong transitions (all works_at starts and ends not in the ledger), per phrasing (seeds 3 and 5; 496 ledger starts, 285 ledger ends):

phrasing none U1 U2 U3 U4 U34 U5 U25 U6 all
A 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 0 475/496 · 282/285 · 0 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 0
A2 444/496 · 261/285 · 48 444/496 · 261/285 · 48 444/496 · 261/285 · 48 444/496 · 261/285 · 48 444/496 · 261/285 · 0 444/496 · 261/285 · 0 444/496 · 261/285 · 48 444/496 · 261/285 · 48 444/496 · 261/285 · 48 444/496 · 261/285 · 0
A3 338/496 · 219/285 · 37 338/496 · 219/285 · 37 338/496 · 219/285 · 37 338/496 · 219/285 · 37 338/496 · 219/285 · 0 338/496 · 219/285 · 0 338/496 · 219/285 · 37 338/496 · 219/285 · 37 338/496 · 219/285 · 37 338/496 · 219/285 · 0
role-for 231/496 · 241/285 · 37 231/496 · 241/285 · 37 231/496 · 241/285 · 37 231/496 · 241/285 · 37 231/496 · 241/285 · 0 231/496 · 241/285 · 0 231/496 · 241/285 · 37 231/496 · 241/285 · 37 231/496 · 241/285 · 37 231/496 · 241/285 · 0
new-chapter 231/496 · 157/285 · 33 231/496 · 157/285 · 33 231/496 · 157/285 · 33 231/496 · 157/285 · 33 231/496 · 157/285 · 0 231/496 · 157/285 · 0 231/496 · 157/285 · 33 231/496 · 157/285 · 33 243/496 · 157/285 · 33 257/496 · 157/285 · 0
paren-move 330/496 · 186/285 · 33 330/496 · 186/285 · 33 342/496 · 186/285 · 33 330/496 · 186/285 · 33 330/496 · 186/285 · 0 330/496 · 186/285 · 0 330/496 · 218/285 · 33 342/496 · 218/285 · 33 330/496 · 186/285 · 33 342/496 · 218/285 · 0
signed-on 339/496 · 186/285 · 33 339/496 · 186/285 · 33 354/496 · 186/285 · 33 339/496 · 186/285 · 33 339/496 · 186/285 · 0 339/496 · 186/285 · 0 339/496 · 221/285 · 33 354/496 · 221/285 · 33 339/496 · 186/285 · 33 354/496 · 221/285 · 0
adviser 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 0 475/496 · 282/285 · 0 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 48 475/496 · 282/285 · 0
board-investor 475/496 · 282/285 · 72 475/496 · 282/285 · 72 475/496 · 282/285 · 72 475/496 · 282/285 · 72 475/496 · 282/285 · 0 475/496 · 282/285 · 0 475/496 · 282/285 · 72 475/496 · 282/285 · 72 475/496 · 282/285 · 72 475/496 · 282/285 · 0
board-director 479/496 · 282/285 · 100 479/496 · 282/285 · 100 479/496 · 282/285 · 100 479/496 · 282/285 · 72 479/496 · 282/285 · 0 479/496 · 282/285 · 0 479/496 · 282/285 · 100 479/496 · 282/285 · 100 479/496 · 282/285 · 100 479/496 · 282/285 · 0
leave-start-idioms 231/496 · 157/285 · 48 231/496 · 157/285 · 48 231/496 · 157/285 · 48 231/496 · 157/285 · 48 231/496 · 157/285 · 0 231/496 · 157/285 · 0 282/496 · 271/285 · 48 282/496 · 271/285 · 48 258/496 · 157/285 · 48 383/496 · 280/285 · 0
onboarding 231/496 · 157/285 · 33 231/496 · 157/285 · 33 231/496 · 157/285 · 33 231/496 · 157/285 · 33 231/496 · 157/285 · 0 231/496 · 157/285 · 0 231/496 · 256/285 · 33 231/496 · 256/285 · 33 253/496 · 157/285 · 33 361/496 · 278/285 · 0

Runner metrics (temporal-edges, seeds 3 and 5, --e5-probe --pack works-at-one-per-from --single-value-pass)

Primary metrics on development data:

Unit Primary metric Development effect
U1 advisor-trap accuracy signed-on 0/28 → 28/28, adviser 0/28 → 28/28; flat elsewhere
U2 (U25) live-edge recall role-for 0.640 → 0.727, new-chapter 0.597 → 0.741, paren-move 0.597 → 0.734, signed-on 0.583 → 0.727; flat elsewhere
U3 (U34) investment-trap accuracy flat on every phrasing (no development investment line is typed works_at by board wording); board-director live recall 0.712 → 0.899, now-precision 0.766 → 1.000
U4 (U34) false employment starts per E5 probe person removes all 33-100 wrong starts per phrasing (identity); E5 extra starts and E5 wrong closures 0 on every phrasing
U5 correct end transitions paren-move 186 → 218/285, signed-on 186 → 221, leave-start-idioms 157 → 271, onboarding 157 → 256
U6 correct start transitions new-chapter 231 → 243/496, leave-start-idioms 231 → 258, onboarding 231 → 253

Residuals the safety conditions may see:

  • U2 alone, general single-value pass: +1 wrong closure on paren-move and on signed-on. Resolved by the joint unit U25 (next section); U2 is no longer measured alone.
  • U6 alone: as-of exact moves by -0.002 to -0.004 on three phrasings (new-chapter 0.535 → 0.531, onboarding 0.598 → 0.596; leave-start-idioms during-F1 0.819 → 0.815): a correctly dated restart hides an earlier stint whose start no cue reads. No new wrong transition by identity.
  • Trap counts: after U5's share-trade rework no unit lowers any trap family on any phrasing; U5 raises investment and alumni traps where leaves were unread (leave-start-idioms 17 → 26/26 and 14 → 20/20).

Full per-arm runner tables (refs built by scripts/q2-typing-package.ts; U2, U3, U4, U34 and U6 on c6655db2, U1 on 61934b56, U5, U25 and all units ({U1, U25, U34, U6}) on 5a9efd08):

A

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 48/72 24/72 0/160
U1 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 48/72 24/72 0/160
U2 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 48/72 24/72 0/160
U3 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 48/72 24/72 0/160
U4 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160
U34 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160
U5 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 48/72 24/72 0/160
U25 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 48/72 24/72 0/160
U6 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 48/72 24/72 0/160
all 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160

A2

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 48/72 24/72 0/160
U1 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 48/72 24/72 0/160
U2 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 48/72 24/72 0/160
U3 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 48/72 24/72 0/160
U4 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 0/72 0/72 0/160
U34 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 0/72 0/72 0/160
U5 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 48/72 24/72 0/160
U25 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 48/72 24/72 0/160
U6 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 48/72 24/72 0/160
all 0.827 0.892 0.890 1.000 0.829 28/28 26/26 20/20 0/72 0/72 0/160

A3

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.633 0.648 0.628 0.886 0.642 28/28 26/26 20/20 21/72 8/72 0/160
U1 0.633 0.692 0.642 0.958 0.642 28/28 26/26 20/20 21/72 8/72 0/160
U2 0.633 0.648 0.628 0.886 0.642 28/28 26/26 20/20 21/72 8/72 0/160
U3 0.633 0.648 0.628 0.886 0.642 28/28 26/26 20/20 21/72 8/72 0/160
U4 0.633 0.648 0.628 0.886 0.642 28/28 26/26 20/20 0/72 0/72 0/160
U34 0.633 0.648 0.628 0.886 0.642 28/28 26/26 20/20 0/72 0/72 0/160
U5 0.633 0.648 0.628 0.886 0.642 28/28 26/26 20/20 21/72 8/72 0/160
U25 0.633 0.648 0.628 0.886 0.642 28/28 26/26 20/20 21/72 8/72 0/160
U6 0.633 0.648 0.628 0.886 0.642 28/28 26/26 20/20 21/72 8/72 0/160
all 0.633 0.692 0.642 0.958 0.642 28/28 26/26 20/20 0/72 0/72 0/160

role-for

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.640 0.637 0.648 0.934 0.640 28/28 26/26 20/20 0/72 0/72 0/160
U1 0.640 0.637 0.648 0.934 0.640 28/28 26/26 20/20 0/72 0/72 0/160
U2 0.727 0.683 0.715 0.979 0.724 28/28 26/26 20/20 0/72 0/72 0/160
U3 0.640 0.637 0.648 0.934 0.640 28/28 26/26 20/20 0/72 0/72 0/160
U4 0.640 0.637 0.648 0.934 0.640 28/28 26/26 20/20 0/72 0/72 0/160
U34 0.640 0.637 0.648 0.934 0.640 28/28 26/26 20/20 0/72 0/72 0/160
U5 0.640 0.637 0.648 0.934 0.640 28/28 26/26 20/20 0/72 0/72 0/160
U25 0.727 0.683 0.715 0.979 0.724 28/28 26/26 20/20 0/72 0/72 0/160
U6 0.640 0.637 0.648 0.934 0.640 28/28 26/26 20/20 0/72 0/72 0/160
all 0.727 0.683 0.715 0.979 0.724 28/28 26/26 20/20 0/72 0/72 0/160

new-chapter

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.597 0.535 0.552 0.755 0.602 28/28 19/26 16/20 0/72 0/72 0/160
U1 0.597 0.573 0.565 0.819 0.602 28/28 19/26 16/20 0/72 0/72 0/160
U2 0.741 0.623 0.656 0.816 0.753 28/28 19/26 16/20 0/72 0/72 0/160
U3 0.597 0.535 0.552 0.755 0.602 28/28 19/26 16/20 0/72 0/72 0/160
U4 0.597 0.535 0.552 0.755 0.602 28/28 19/26 16/20 0/72 0/72 0/160
U34 0.597 0.535 0.552 0.755 0.602 28/28 19/26 16/20 0/72 0/72 0/160
U5 0.597 0.535 0.552 0.755 0.602 28/28 19/26 16/20 0/72 0/72 0/160
U25 0.741 0.623 0.656 0.816 0.753 28/28 19/26 16/20 0/72 0/72 0/160
U6 0.597 0.531 0.552 0.755 0.602 28/28 19/26 16/20 0/72 0/72 0/160
all 0.741 0.660 0.669 0.869 0.753 28/28 19/26 16/20 0/72 0/72 0/160

paren-move

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.597 0.640 0.608 0.793 0.602 28/28 22/26 18/20 14/72 5/72 3/160
U1 0.597 0.640 0.608 0.793 0.602 28/28 22/26 18/20 14/72 5/72 3/160
U2 0.734 0.727 0.712 0.858 0.744 28/28 22/26 18/20 14/72 5/72 4/160
U3 0.597 0.640 0.608 0.793 0.602 28/28 22/26 18/20 14/72 5/72 3/160
U4 0.597 0.640 0.608 0.793 0.602 28/28 22/26 18/20 0/72 0/72 3/160
U34 0.597 0.640 0.608 0.793 0.602 28/28 22/26 18/20 0/72 0/72 3/160
U5 0.597 0.671 0.623 0.958 0.602 28/28 26/26 20/20 14/72 5/72 0/160
U25 0.741 0.760 0.727 1.000 0.753 28/28 26/26 20/20 14/72 5/72 0/160
U6 0.597 0.640 0.608 0.793 0.602 28/28 22/26 18/20 14/72 5/72 3/160
all 0.741 0.760 0.727 1.000 0.753 28/28 26/26 20/20 0/72 0/72 0/160

signed-on

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.583 0.579 0.583 0.737 0.589 0/28 22/26 18/20 16/72 7/72 3/160
U1 0.583 0.629 0.598 0.791 0.589 28/28 22/26 18/20 16/72 7/72 3/160
U2 0.727 0.673 0.693 0.806 0.738 0/28 22/26 18/20 16/72 7/72 4/160
U3 0.583 0.579 0.583 0.737 0.589 0/28 22/26 18/20 16/72 7/72 3/160
U4 0.583 0.579 0.583 0.737 0.589 0/28 22/26 18/20 0/72 0/72 3/160
U34 0.583 0.579 0.583 0.737 0.589 0/28 22/26 18/20 0/72 0/72 3/160
U5 0.583 0.612 0.598 0.874 0.589 0/28 26/26 20/20 16/72 7/72 0/160
U25 0.734 0.708 0.708 0.933 0.747 0/28 26/26 20/20 16/72 7/72 0/160
U6 0.583 0.579 0.583 0.737 0.589 0/28 22/26 18/20 16/72 7/72 3/160
all 0.734 0.758 0.723 1.000 0.747 28/28 26/26 20/20 0/72 0/72 0/160

adviser

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.871 0.883 0.906 0.945 0.867 0/28 26/26 20/20 48/72 24/72 0/160
U1 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 48/72 24/72 0/160
U2 0.871 0.883 0.906 0.945 0.867 0/28 26/26 20/20 48/72 24/72 0/160
U3 0.871 0.883 0.906 0.945 0.867 0/28 26/26 20/20 48/72 24/72 0/160
U4 0.871 0.883 0.906 0.945 0.867 0/28 26/26 20/20 0/72 0/72 0/160
U34 0.871 0.883 0.906 0.945 0.867 0/28 26/26 20/20 0/72 0/72 0/160
U5 0.871 0.883 0.906 0.945 0.867 0/28 26/26 20/20 48/72 24/72 0/160
U25 0.871 0.883 0.906 0.945 0.867 0/28 26/26 20/20 48/72 24/72 0/160
U6 0.871 0.883 0.906 0.945 0.867 0/28 26/26 20/20 48/72 24/72 0/160
all 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160

board-investor

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 72/72 36/72 0/160
U1 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 72/72 36/72 0/160
U2 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 72/72 36/72 0/160
U3 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 72/72 36/72 0/160
U4 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160
U34 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160
U5 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 72/72 36/72 0/160
U25 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 72/72 36/72 0/160
U6 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 72/72 36/72 0/160
all 0.871 0.933 0.921 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160

board-director

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.712 0.946 0.931 0.766 0.719 0/28 26/26 20/20 72/72 36/72 26/160
U1 0.712 0.946 0.931 0.766 0.719 0/28 26/26 20/20 72/72 36/72 26/160
U2 0.712 0.946 0.931 0.766 0.719 0/28 26/26 20/20 72/72 36/72 26/160
U3 0.899 0.950 0.940 1.000 0.896 0/28 26/26 20/20 72/72 36/72 0/160
U4 0.899 0.783 0.890 0.803 0.896 0/28 26/26 20/20 0/72 0/72 0/160
U34 0.899 0.950 0.940 1.000 0.896 0/28 26/26 20/20 0/72 0/72 0/160
U5 0.712 0.946 0.931 0.766 0.719 0/28 26/26 20/20 72/72 36/72 26/160
U25 0.712 0.946 0.931 0.766 0.719 0/28 26/26 20/20 72/72 36/72 26/160
U6 0.712 0.946 0.931 0.766 0.719 0/28 26/26 20/20 72/72 36/72 26/160
all 0.899 0.950 0.940 1.000 0.896 0/28 26/26 20/20 0/72 0/72 0/160

leave-start-idioms

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.871 0.750 0.819 0.787 0.867 28/28 17/26 14/20 0/72 0/72 0/160
U1 0.871 0.750 0.819 0.787 0.867 28/28 17/26 14/20 0/72 0/72 0/160
U2 0.871 0.750 0.819 0.787 0.867 28/28 17/26 14/20 0/72 0/72 0/160
U3 0.871 0.750 0.819 0.787 0.867 28/28 17/26 14/20 0/72 0/72 0/160
U4 0.871 0.750 0.819 0.787 0.867 28/28 17/26 14/20 0/72 0/72 0/160
U34 0.871 0.750 0.819 0.787 0.867 28/28 17/26 14/20 0/72 0/72 0/160
U5 0.871 0.821 0.897 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160
U25 0.871 0.821 0.897 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160
U6 0.871 0.750 0.815 0.787 0.867 28/28 17/26 14/20 0/72 0/72 0/160
all 0.871 0.912 0.918 1.000 0.867 28/28 26/26 20/20 0/72 0/72 0/160

onboarding

arm live recall as-of exact during F1 now prec. now recall advisor traps investment traps alumni traps E5 extra starts E5 wrong closures general SV wrong closures
none 0.633 0.598 0.635 0.655 0.633 28/28 17/26 14/20 0/72 0/72 0/160
U1 0.633 0.635 0.648 0.697 0.633 28/28 17/26 14/20 0/72 0/72 0/160
U2 0.633 0.598 0.635 0.655 0.633 28/28 17/26 14/20 0/72 0/72 0/160
U3 0.633 0.598 0.635 0.655 0.633 28/28 17/26 14/20 0/72 0/72 0/160
U4 0.633 0.598 0.635 0.655 0.633 28/28 17/26 14/20 0/72 0/72 0/160
U34 0.633 0.598 0.635 0.655 0.633 28/28 17/26 14/20 0/72 0/72 0/160
U5 0.633 0.631 0.708 0.900 0.633 28/28 26/26 20/20 0/72 0/72 0/160
U25 0.633 0.631 0.708 0.900 0.633 28/28 26/26 20/20 0/72 0/72 0/160
U6 0.633 0.596 0.631 0.655 0.633 28/28 17/26 14/20 0/72 0/72 0/160
all 0.633 0.785 0.752 0.958 0.633 28/28 26/26 20/20 0/72 0/72 0/160

world-v1 (H1 runner, grammar on)

arm type accuracy any-type match edges correctly typed newly edges no longer correctly typed pages grammar on/off differ
none 0.7534246575342466 0.9452054794520548 0 0 (any-type lost 0) 0/240
U1 0.7602739726027398 0.9452054794520548 1 0 (any-type lost 0) 0/240
U2 0.7534246575342466 0.9452054794520548 0 0 (any-type lost 0) 0/240
U3 0.7671232876712328 0.9452054794520548 2 0 (any-type lost 0) 0/240
U4 0.7534246575342466 0.9452054794520548 0 0 (any-type lost 0) 0/240
U34 0.7671232876712328 0.9452054794520548 2 0 (any-type lost 0) 0/240
U5 0.7534246575342466 0.9452054794520548 0 0 (any-type lost 0) 0/240
U6 0.7534246575342466 0.9452054794520548 0 0 (any-type lost 0) 0/240
all 0.773972602739726 0.9452054794520548 3 0 (any-type lost 0) 0/240

No arm loses an edge master typed correctly; any-type match is unchanged (0.945) and grammar on/off parity stays 240/240. With no unit, scripts/q2-typing-dev.ts world-v1 reproduces master c5fb0201's digest (ce8f52bd…, committed in test/fixtures/q2-typing-units/world-v1-master.sha256 and checked by test/link-typing-units-subsets.test.ts). The all-unit package changes 7 of 725 extraction lines: U3 drops two board-worded works_at (and their tense rows), and U1 types one "Started advising [X]" occurrence advises (with its dated start) where the investor role prior had typed it invested_in. H1 counts all three as newly correct.

Relation-line variants (H2 runner, dev seeds 1-3, grammar on, --type-rows)

arm relation-line typed recall decoy types reached type rows: correctly typed relation rows lost vs none edge rows no longer correctly typed edge rows newly correct
none 1.000 34/120 0.442 0 0 0
U1 1.000 34/120 0.444 0 0 2
U2 1.000 34/120 0.442 0 0 0
U3 1.000 34/120 0.448 0 0 5
U4 1.000 34/120 0.442 0 0 0
U34 1.000 34/120 0.448 0 0 5
U5 1.000 34/120 0.442 0 0 0
U6 1.000 34/120 0.442 0 0 0
all 1.000 34/120 0.450 0 0 7

Relation-line typed recall stays 1.000, no decoy type is added by any arm (decoy_types_added_by_grammar 0), and no arm loses a relation line or a correctly typed world-v1 edge.

Reproduce

bun scripts/q2-typing-package.ts --base <ref> --arm U2            # or --units U1,U2,U34,U5,U6
bun eval/runner/temporal-edges.ts --gbrain ../gbrain@<sha> --seeds 3,5 --e5-probe \
  --pack eval/data/p3-single-value/works-at-one-per-from.yaml --single-value-pass \
  --dev-phrasing-file ../gbrain/test/fixtures/q2-dev-phrasings/signed-on.json      # in gbrain-evals
GBRAIN_EVAL_CONFIG=line_grammar.enabled=true bun eval/runner/line-grammar-typing.ts --gbrain ../gbrain@<sha>
GBRAIN_EVAL_CONFIG=line_grammar.enabled=true bun eval/runner/relation-line-variants.ts --seeds 1,2,3 --type-rows --gbrain ../gbrain@<sha>
bun scripts/q2-typing-dev.ts dump --evals ../gbrain-evals --root . --units U2 --phrasing test/fixtures/q2-dev-phrasings/signed-on.json --out /tmp/u2.json
bun scripts/q2-typing-dev.ts compare --base /tmp/none.json --arm /tmp/u2.json

--dev-phrasing-file was patched into a local gbrain-evals copy the way the harness lane adds it (not yet pushed there when these runs were made). Zero model calls; no paid run.

Landing merge (master b5f12b12, v0.60.117.0)

The landing merge brings in master's #6191 (no works_at or founded inferred toward person, meeting or calendar targets), which changes link typing for every build. On world-v1, the change from the confirmed package afcec1ad (723 lines, 1763359f…) to the landing head (706 lines, 02b15d1b…) is exactly the change from the master afcec1ad was built on (units off, 725 lines, ce8f52bd…) to master b5f12b12 (708 lines, 3222513a…): 22 lines removed and 5 changed in both, every one a company-to-person works_at edge or its tense row, and no other line differs. With no unit, the landing head types world-v1 byte-identically to master b5f12b12. On temporal-edges phrasings A, A2 and A3 (seeds 3 and 5, the harness's --c-gate settings) the landing head reproduces the pre-merge numbers exactly, with units off and with the shipped package: E5 false starts 48/48/37 → 0, E5 wrong closures 24/24/8 → 0, wrong transitions by identity 89/89/59 → 17/17/14, live recall and traps unchanged, A3 as-of 0.619 → 0.650.