Skip to content

Garry's Opinionated OpenClaw/Hermes Agent Brain

TypeScript
9.2 KB
Raw Blame

Q2 Track C: development mechanism trace

Where did set H's current employers go? On fresh development phrasings, the build that set H measured (583a851ee) loses a current employer from the default get_links read in two ways, and neither is the one the published summary suggests. A current employer whose stint resumed after an idiomatic leave is closed by the new leave cue because the restart wording is not a start cue (5 of 7 lost edges; the edge stays typed works_at, so the live read shows only the mentions row). A new investor spelling on one dated timeline line attaches to the link on the next dated line and outranks works_at by verb precedence (2 of 7). Stated types, pack rules and role priors played no part.

Development data only: temporal-edges dev seeds 3 and 5, phrasing sets A, A2 and A3, and eight new phrasing files written for this trace (test/fixtures/q2-dev-phrasings/). No custodian or sealed text was read.

Setup

  • Builds: gbrain master c5fb0201 (the Track C base), 583a851ee (set H's build) and 27b1b6af (= 780a4fc5^, the last state before the amendment-3 removal). 583a851ee and 27b1b6af have identical src/core/link-extraction.ts and src/core/link-temporal-evidence.ts; master differs from both by exactly the git show 780a4fc5 diff in those files, so every difference below comes from the removed rules.
  • Metrics: gbrain-evals eval/runner/temporal-edges.ts on branch capy/q2-parser-gaps, --seeds 3,5 --e5-probe --pack eval/data/p3-single-value/works-at-one-per-from.yaml --single-value-pass, with --phrasing A|A2|A3 or a development phrasing file (--dev-phrasing-file, patched locally the way the harness lane adds it).
  • Edge identity: bun scripts/q2-typing-dev.ts dump writes the same worlds through put_page on each build and records every person page's link rows (type, source, tense, context), its link_transitions rows and its default get_links read; compare prints type steals and transitions by identity against the generator ledger.

New development phrasings (each a full PhrasingTemplates file):

File Shapes
role-for " for " (current, join, move)
new-chapter a join described as starting a new chapter
paren-move a move or join with the role in parentheses; "now at [X] ()"
signed-on "signed on with [X] as "; "serves as an adviser to"; split-object leave
adviser "is an adviser to"; "now advising [X]"
board-investor "took a board seat at [X] as an investor"; observer at a board meeting; advisory board
leave-start-idioms quitting idioms, an exchange move, "first day at", "began working at"; "traded shares of" and "first day of the [X] summit" look-alikes
onboarding onboarding starts, "stepped back from", "parted company with"; an onboarding call as a look-alike

Lost current employers, by mechanism

A lost edge is a current employer (ledger) that master's default get_links lists as works_at and 583a851ee's does not. 11 phrasing sets × 2 seeds, 2,552 person pages per build (232 per phrasing set, E5 probe people included).

Mechanism Lost Where Detail
Stated type or pack rule (link-extraction.ts ~708) 0 — no relation lines in prose renders; every base-pack rule is ner_only
Attachment (attachedVerb) + VERB_RULES precedence 2 A2 s5, board-investor s5 (same person) "…'s Series B as an angel investor" / "…as an investor" on the line before "Was hired by [Stark] as VP engineering": the new INVESTED_RE alternatives match, no link sits between the match and [Stark], so the match attaches, and invested_in outranks the works_at that master found after the link ("VP engineering")
Role prior 0 —
Temporal evidence 5 onboarding s3 ×2, s5 ×3 a rejoined current stint: "Parted company with [X]" or "Stepped back from [X]" (new end cues) dates the earlier exit, the later restart "Onboarding week at [X] as …" is not a start cue on any build, so the relationship stays closed. Master recognized neither, so the edge stayed undated and live. The end transitions are correct by identity; the missing restart is missing on both builds
Live-read visibility of an unchanged edge 0 —

Every "lost" edge shows up in the live read of 583a851ee as a mentions row only, which is how the published summary ("typed mentions where master typed works_at") reads when seen from get_links. On development text, 5 of 7 are a temporal closure, not a typing change.

Type steals and transitions by identity (583a851ee vs master)

A steal is a person→company pair where the build drops a type the ledger supports or adds one it does not.

Phrasing Steals Gains New wrong transitions Correct transitions fixed Missing correct Live lost / gained
A 0 0 0 0 0 0 / 0
A2 8 37 0 29 1 1 / 0
A3 0 85 0 81 0 0 / 69
role-for 0 0 0 0 0 0 / 0
new-chapter 0 24 0 24 0 0 / 0
paren-move 0 0 0 0 0 0 / 0
signed-on 0 0 0 35 0 0 / 0
adviser 0 52 0 0 0 0 / 0
board-investor 23 0 0 0 1 1 / 0
leave-start-idioms 0 0 0 300 0 0 / 0
onboarding 0 24 0 198 0 5 / 0

The 31 steals, by cause:

Cause Pairs Ledger relation of the pair
investor spelling ("as an angel investor", "as an investor") on another dated line attaches across the line break 16 advisee 10, current employer 6
observer or board wording ("…board meeting as an observer") types invested_in 15 former employer (an alumni-trap line) 15

The three published set-H shapes did not reproduce a loss: " for ", "()" after a move and most "new chapter" joins type mentions on both builds (live recall 0.640, 0.597 and 0.597 on master and 583a851ee alike), because the removed ROLE_AT_RE reads " at" and "as ", not "for" or parentheses. They are recall gaps, not regressions.

Runner metrics that moved (seeds 3 and 5; master → 583a851ee):

Phrasing live recall as-of exact advisor traps investment traps alumni traps E5 extra starts E5 wrong closures
A 0.871 → 0.871 0.933 → 0.933 28/28 26/26 20/20 48 → 0 24 → 0
A2 0.827 → 0.820 0.892 → 0.900 28/28 26/26 20/20 48 → 0 24 → 0
A3 0.633 → 0.827 0.648 → 0.790 28/28 26/26 20/20 21 → 0 8 → 0
signed-on 0.583 → 0.583 0.579 → 0.612 0/28 → 0/28 22 → 26/26 18 → 20/20 16 → 0 7 → 0
adviser 0.871 → 0.871 0.883 → 0.915 0/28 → 28/28 26/26 20/20 48 → 0 24 → 0
board-investor 0.871 → 0.863 0.933 → 0.929 28/28 26/26 20/20 72 → 0 36 → 0
leave-start-idioms 0.871 → 0.871 0.750 → 0.863 28/28 17 → 20/26 14 → 19/20 0 → 0 0 → 0
onboarding 0.633 → 0.597 0.598 → 0.698 28/28 17 → 26/26 14 → 20/20 0 → 0 0 → 0

583a851ee also applied 5 single-value closures on leave-start-idioms (sv_wrong_closures 0 → 5/160).

What the units change because of this

  • U3 adds no investor or observer spelling to INVESTED_RE. Its only typing changes are suppressions (board, observer and investor wording never types works_at; "board seat at" keeps invested_in only with an investor prior), so it cannot produce the cross-line invested_in steal. Observer wording stays mentions.
  • U5 (leave idioms and exchange moves) gets a restart guard: an end it adds is dropped when a later dated line refers to the same organization in plain words that no cue reads (a possible unrecognized rejoin), outside event and qualified lines. Without it, U5 reproduces the onboarding losses.
  • U6 covers the restart framings the trace found ("onboarding week at", "onboarded at", "first day at", "began working at", "kicked off a new role at", "a new chapter at").
  • U2 targets the shapes that were recall gaps on every build (" for", "()" after the link, "as " after a join, move, sign-on or new chapter), on the link's own line only, so a role on another timeline line never attaches.
  • U1 carries the adviser spelling and "advising [X]"; the signed-on and adviser sets show master at 0/28 advisor traps for those spellings.
  • U3/U4 dependency: see dev-units.md (board-director development phrasing). The trace above has no board line that master types works_at, so it cannot show the set-F interaction by itself.

Reproduce (from a gbrain checkout with gbrain-evals beside it):

bun eval/runner/temporal-edges.ts --gbrain ../gbrain@583a851ee --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/onboarding.json --output /tmp/te   # in gbrain-evals
bun scripts/q2-typing-dev.ts dump --evals ../gbrain-evals --root ../gbrain-evals/.gbrain-overlays/583a851ee \
  --phrasing test/fixtures/q2-dev-phrasings/onboarding.json --out /tmp/583.json                      # in gbrain
bun scripts/q2-typing-dev.ts compare --base /tmp/master.json --arm /tmp/583.json