Garry's Opinionated OpenClaw/Hermes Agent Brain
Private forecast journal (wave 4b W7)
Status: all four slices in #6886. Plan: gbrain-evals docs/plans/2026-10-07-10x-memory-advantage/PLAN.md, wave 4, row W7.
W7 makes each prediction in a brain auditable for its owner: what was predicted, when, by when it had to come true, how it was resolved and on what evidence, and how many predictions are still open. It extends takes_scorecard and gbrain takes scorecard instead of adding a track_record op. Everything new is private, which means trusted local only. takes_scorecard keeps returning exactly what it returns today to remote callers. Nothing public about real people ships from this wave.
Slice 1 (this change): the horizon
The horizon of a take is the last day its claim can come true: 2024-12-31 for "by the end of 2024", 2024-09-30 for "by Q3 2024". Without it, the journal cannot tell a prediction that is overdue from one that is still open.
Where it lives. It goes in a table of its own, take_horizons, one row per take (migration v251 for now; the number is taken when the merge slot is called). It does not go in either of the two obvious places:
- The takes fence. Changing the fence format would make sync hold files. Garry's decision is an additive DB column, not a fence-format change.
- A column on
takes.takes_listandtakes_searchreturnt.*, so a new column would reach remote callers. That would break the rule that remote output stays the same, and it would leak a private field.
| Column | Type | Meaning |
|---|---|---|
take_id |
BIGINT PRIMARY KEY REFERENCES takes(id) ON DELETE CASCADE |
The take. The row is deleted when its take is. |
claim_sha256 |
TEXT NOT NULL |
The sha256 of the claim the horizon was set for. If the claim is later edited, the row stays as written but reads as not current, so a horizon never carries over to a different proposition. |
horizon_date |
DATE NOT NULL |
The last day the claim can come true. |
basis |
TEXT NOT NULL, one of stated, stated_relative, owner |
How the horizon was set (see below). |
set_at |
TIMESTAMPTZ NOT NULL DEFAULT now() |
When it was set. |
The three bases:
statedis an absolute deadline in the claim's own words: "by the end of 2024", "by the end of Q1 2025", "by Q3 2024", "by March 2025", "before 2025", "by 2026".stated_relativeis "within N months" or "within N years", counted fromsince_date. It is only as certain assince_date. On a source without recording, that date is the uncertified one, and the journal labels ituncertified(Garry, 2026-10-17).owneris set by the owner. A scan never overwrites it.
No page text is stored. The table holds dates, an enum, a hash and a timestamp. That keeps it outside the purge surface (text_free_ledger in the deletion inventory). Graduation copies it as user data. Postgres's RLS event trigger covers it.
Nothing is written on upgrade.
gbrain takes horizon scanlists each take's stated horizon. It writes them only with--apply.gbrain takes horizon set <slug> --row N --date YYYY-MM-DDandclearare the owner's own edits.- None of these has an MCP op. A thin client refuses them as host-bound.
- The parser returns nothing in three cases: when a claim states no deadline, when it states two different deadlines, or when it states a relative deadline without a usable
since_date. It does not guess.
What stays the same. takes_list and takes_scorecard return byte-identical output, locally and to remote callers, whether or not horizons exist. Tests on PGLite and Postgres check this.
Slice 2: prediction time and the journal
src/core/forecast-journal.ts reads every active take that is not a fact as a prediction.
Prediction time. The prediction time is when the belief ledger recorded the exact claim (claim, kind and holder) on its page:
- A take added after its page was created has a live
take.assertedevent. Its first one gives the time. - A take the page was created with has no event of its own, because a page creation is one
page.publishedsnapshot (BELIEF_LEDGER.md section 4). Such a take gets the snapshot's time, but only when the ledger has recorded the source without a stop or restore since then, with no drift mark on the page. Otherwise the take could have been added while the ledger was not recording, so it falls back. - A take whose first event is
take.transferredcame from another page. It gets no recorded time.
A recorded time is certified only where coverageAt certifies that instant, which means a managed source. Every other take shows its since_date, labelled uncertified. Tests on PGLite and Postgres cover four cases: a managed source (certified), an unmanaged one (recorded, uncertified), a take added later, and a coverage stop. A forced probe fails the stop case when the guard is removed.
The journal. gbrain takes journal (trusted local) shows each prediction with:
- its time and certification;
- its claim, holder and weight;
- its horizon and basis;
- its state:
resolved,overdue,openorundated; - its resolution: quality,
resolved_byandresolved_source.
It filters by --holder, --about <slug-prefix>, --state, --since, --until and --source-id, and --limit caps the list. The totals always count every match, so the unresolved denominator survives any filter on state or limit. A thin client refuses it as host-bound. There is no MCP op. Each entry also carries grade: the newest verdict grade_takes cached for the take (W6, the one gbrain takes grades lists), applied or not, with judge_measured saying whether the W6 eval measured its judge and prompt.
Slice 3: takes_scorecard {group_by}
With group_by set to holder or source, a trusted local caller gets the plain card unchanged plus two additions.
groups has one row per group, with:
n,correct,incorrect,partialandunresolvable;accuracy;accuracy_shrunk: a Beta prior of 10 pseudo-observations centred on the pooled accuracy;accuracy_wilson95;accuracy_recent: weighted by a 365-day half-life since resolution;brierandbrier_skill: skill is measured against the group's own base rate;citations: resolutions that name a source;open,overdue,undatedandunresolved;low_n: belowmin_n, which defaults to 10.
journal holds the journal totals.
A remote caller that passes group_by or min_n is refused with scorecard_group_by_local_only. Without them, its output is what it was. gbrain takes scorecard --group-by holder|source [--min-n N] is the CLI form, and thin clients refuse it.
Slice 4: the $0 gate
evals/forecast-journal/: the preregistration was frozen before the scored run. The gate uses 20 seeded cohorts of placeholder holders with known skill, 5,622 predictions in all. It passed:
- median Spearman ρ of 0.905;
- who-was-right set-F1 of 1.000;
- exact denominator,
low_nand labelling counts; - 8/8 remote probes.
The details are in that directory's README.
The resolution policy for auto-resolve is out of W7's scope. Garry's decision (2026-10-16) is recorded with the take-grading follow-ups in gbrain-evals: a take whose numeric target was missed is never auto-applied, neither as partial nor as incorrect. It is held for manual review until the real sample settles which.