Steering

What is currently shaping the method — frozen specs, learning gates, and distilled directives. A steer without a measured outcome is labelled as such.

Frozen specs

Era 2era2-v1 · hash e12b5ea9b314 · closed gate unreached
Era 3era3-v1 · hash 5f736e8cd424 · model claude-sonnet-5 · closed to new questions 2026-11-05 (decision 0021), gate unreached
Era 4era4-v1 · hash 1ff59df4acb4 · model claude-sonnet-5 · no backup model · current

Era 3 is closed to new questions

On 2026-11-05 (decision 0021) Era 3 stopped registering at 56 of 100 storylines — gate unreached, no edge claim. Nothing in its record is regraded, re-scoped or moved into Era 4. The 20 questions it registered before the close still get their daily state from the same pinned model and settle by the same code; they count toward Era 3 only. It closed because an audit that day found the record measures the wrong thing: the macro baseline counted one-month rises while each question spans k months; bars were first prints but settlement read revised values; earnings compared a four-decimal consensus with EPS reported to the cent; a company registered only if its own few-report miss rate fell inside the band, which pushed every earnings baseline up; a backup model answered when the pinned one failed and its states were scored; and the challenger's prompt could read the champion's own moves. Fixing any of these inside Era 3 would move a frozen bar, so Era 4 fixes all six in a spec frozen before its first question. This is the third era to close without its gate; decision 0021 carries the review of gate sizing that decision 0014 required. Era 3's book.

Record integrity (Era 3)

Written under decision 0018every forecast after #2411 (the build went live 2026-10-21 09:55:52 UTC)
Earnings states written on or after their report date33 before the boundary · 0 after it
Written under decision 0019every forecast after #2931 (the build went live 2026-11-04 10:44:44 UTC)

Decision 0019 fixed, at deploy: the macro forecaster never saw the latest reading of the series it was forecasting (five of six were not fetched; the sixth was labelled so its search could not find it); and an earnings question with no report in its window was graded NO even when the report had only moved. Two such grades (XOM, AAPL) were restated to unresolvable with the original kept.

Decision 0018 fixed, at deploy: the earnings seeder asking each company once in its life; the forecaster updating after a report was public; a settled sibling's null reaching open questions as a float; a malformed earnings base-rate line; and Era 3 forecasts stamped with directives they never read. States before the boundary are not rewritten. After it, no state is written on or after a report date, so the second count should stay at zero.

Challenger, Era 4 never the record

Laneevidence_pack.v2: the pinned forecaster plus an evidence pack stored before each state
Started2026-11-05 00:47:55 UTC — the promotion thresholds froze with this first state
States written206 on 32 questions
Matched resolutions6 of 30 needed
Paired Brier, champion minus challenger-0.0065 (positive = challenger better; promote at ≥ 0.02)
Verdictnot yet: 6 of 30 matched resolutions
FamilyMatchedPaired delta
finnhub_earningsn=6-0.0065

Pooled across families decides; per-family numbers are shown, not used. A promote is a dated decision, never automatic. The challenger never reads the null. This lane runs on Era 4 questions with a context that carries no Workshop row, and only the pinned model writes its states.

Challenger, Era 3 closed, the saved record

Laneevidence_pack.v1: the pinned forecaster plus an evidence pack stored before each state
Started2026-11-04 21:30:19 UTC — the promotion thresholds froze with this first state
States written24 on 24 questions
Matched resolutions12 of 30 needed
Paired Brier, champion minus challenger-0.0439 (positive = challenger better; promote at ≥ 0.02)
Verdictclosed 2026-11-05 with Era 3 (decision 0021): record only, no promote is possible; 12 matched resolutions
FamilyMatchedPaired delta
finnhub_earningsn=6-0.0772
fredn=6-0.0106

Pooled across families decides; per-family numbers are shown, not used. A promote is a dated decision, never automatic. The challenger never reads the null. Correction (decision 0021): this page said this lane read no champion state. It did: its prompt shared the champion's news search, which carried Workshop's own "Forecast moved" notes. It stopped at Era 3's close; its states are the saved record and its comparison ends at the close day.

Learning gates connect pass only

No learning gates recorded yet. Set LEARNING_SHADOW_GATE=1 to persist decisions.

Active directives connect pass only — never in a forecaster's prompt

Directive effectiveness (Era 3)

Not causal. From Aug 2, 2026 (when attribution began) until the decision-0018 build, an Era 3 forecast was stamped with whatever directives were active in the connect pass, which the Era 3 prompt never contained. Groups with a hash below split the record by time, not by cause. Since that build a forecast is stamped only with what its own prompt carried — the public forecaster carries none — so a directive gets a causal number only when a challenger lane tests it (decisions 0017, 0018).

Directive setStorylinesMean edge
none injected / predates attributionn=33+0.0022 Brier
fe658f2d1aa11aab active in the connect pass, not in the Era 3 promptn=17-0.0036 Brier
6e639f5bd391f867 active in the connect pass, not in the Era 3 promptn=6-0.0185 Brier

Descriptive only — no auto-retire. A storyline is attributed to the stamp on its EARLIEST forecast, never a later one, so a steer landing mid-life cannot retroactively own an older call's outcome.