The Cost of Writing for an Eavesdropper: Forecast Information and Calibration in Council Bulletins Under Sophon-Aware Drafting
Abstract
We compare the information and the calibration of strategic-bulletin forecasts drafted under the Council's sophon-aware regime (DE 12 to 30) with those of the later regime (DE 31 to 49). We hold topic and drafting office fixed in 28 cells of the 42 office-by-topic pairs, and we score 1,536 resolved forecasts, 704 from the first regime and 832 from the second. We assume that sophon readership was continuous across both regimes, so the contrast is one of drafting guidance and stakes. We estimate that the first regime carried 0.11 bits of information per forecast against 0.23 for the second, a paired difference of 0.12 bits (95% CI 0.06 to 0.17), and that the share of forecasts stated in the 0.4 to 0.6 band fell from 41% to 23%. Recalibration slopes were 0.95 and 0.98 (difference 0.03, 95% CI -0.25 to 0.30), so calibration was kept and sharpness was lost. The gain was 0.13 bits in Trisolaran-related topics and 0.10 in others (difference 0.03, 95% CI -0.08 to 0.14), so guidance cannot be told from the yearly score feedback that came with it. Forecasts drafted under the sophon-aware guidance likely carried about half the information of those drafted after it (shortfall 51%, 95% CI 32% to 71%), and the text alone cannot say whether the guidance or the score feedback carried it, or how much its own readers recovered.
1. Introduction
In the strategic bulletins of the UN Planetary Defense Council (PDC) for Deterrence Era, Years 12 to 30, 287 of 704 resolved forecasts (41%) were stated in the 0.4 to 0.6 band, the band that a drafter uses when a text is to say almost nothing about an outcome. In Years 31 to 49 the same seven drafting offices, writing on the same six topics, put 189 of 832 (23%) there. The band moved because the drafting guidance moved (Section 2), and that movement is the cheapest measurement in this paper, whereas the costly one is what the band took away from the people who read the bulletins.
What did that caution cost the Council's own readers? Drafters in the first period were told to assume that every word was read by sophons, and so by Trisolaris (United Nations Planetary Defense Council, Secretariat, 31 DE). A drafter who assumes a hostile reader has reason to say less (Quillfeather, 55 DE) and writes for two audiences at once (Corvane, 49 DE). Comparing all bulletins of the two periods risks crediting the guidance with changes that belong to topic and office. On the other hand, comparing only within one office and one topic risks leaving too few resolved forecasts in each comparison to say anything. We place our design between these risks by holding topic and office fixed in 28 cells that each hold at least 12 resolved forecasts in both periods. Inside them the first period carried about 49% of the information per forecast that the second carried.
We write in Deterrence Era, Year 60 (DE 60), and no source is dated later. For the coding of stated probabilities and the recalibration regression we follow an earlier paper (Sarkissian-Low, 58 DE), which we do not re-score here.
2. Methods
The sources are the strategic bulletins of the Secretariat in the release series BU, which begins in DE 12 (United Nations Planetary Defense Council, Secretariat, 12–49 DE), and the resolution register that the Office of Strategic Forecasting keeps for them (PDC Secretariat, Office of Strategic Forecasting, 60 DE). Two drafting circulars (4 and 9) define the periods. Circular 4, the first of the series, told drafters to assume that an adversary read each text and to avoid stating the odds of events that touch Trisolaran conduct. Circular 9, issued in DE 31, withdrew that instruction and began a yearly review in which each office's resolved forecasts were returned to it (United Nations Planetary Defense Council, Secretariat, 31 DE). Circular 4 also prescribed a five-level scale for odds, and Circular 9 allowed finer steps, so the grid belongs to the drafters and not to the coder. We call DE 12 to 30 the sophon-aware regime and DE 31 to 49 the later regime.
Sophon interference was partially suspended under a declaration from Trisolaris that the Council received and recorded in DE 2 (United Nations Planetary Defense Council, Secretariat, 2 DE), and the declaration concerns interference. The suspension therefore precedes both windows, and the years DE 12 to 30 are not pre-suspension drafting but drafting before the guidance changed. We assume that sophon readership was continuous across both regimes, since nothing in the record says that sophons stopped reading, and we cannot verify from the record whether the suspension held. The contrast is one of guidance and stakes (the two circulars) and not of anything that the suspension did to the readers.
In all, 1,322 bulletins fall in the two windows. We coded 3,481 forecast statements that carry a stated probability, or a hedge that the reader key of Tarnow (31 DE) maps to one. Of these, 2,107 have a horizon of five years or less and were resolved by DE 59. We dropped 571 outside the cells below and kept 1,536, of which 704 belong to the sophon-aware regime and 832 to the later one, in 540 bulletins (mean 2.8). A second coder re-coded 150 statements blind to the first coding, with exact agreement on 112 (75%) and agreement within one step of the reader key on 134 (89%).
A cell is one of the 42 pairs of drafting office (seven) and topic (six) in the Secretariat's routing slips, and 28 pairs hold at least 12 resolved forecasts in each regime. The cells hold 20 to 35 forecasts in the first regime and 22 to 33 in the second. The topics are transmitter-network readiness, fleet and procurement status, inbound-fleet movements, sophon activity windows, civil continuity, and Council readiness reports. Circular 4 named the first, third and fourth as touching Trisolaran conduct, which gives 13 Trisolaran-related cells and 15 others (we fixed this split from the circular's wording before scoring). A forecast is "vague" when its stated probability lies in the 0.4 to 0.6 band, which holds one level (0.5) of the first grid and two levels of the second.
Information is the log score of the stated probability for what happened, in bits, minus the log score of the cell's base rate for that outcome (Tennant-Okoye, 54 DE), averaged over a cell's forecasts. We score each forecast at the mean stated probability of its band. A base rate taken from the same cell fits the cell's own outcomes too well and so understates information, by about 1/(n ln 2) bits per forecast against a base rate that leaves the scored forecast out (0.06 bits at 25 forecasts). We also report that version. The primary comparison is the mean over the 28 cells of the paired difference (later minus sophon-aware), with a t interval on 27 degrees of freedom, and the ratio of the two cell means takes the interval of ratio values that the paired cell data do not reject at the 5% level. That interval agrees with the paired test. A forecast-weighted contrast, which gives large cells more say, takes its interval from leaving out one cell at a time. Calibration comes from a logistic regression of what happened against the logit of each stated probability, fitted to the five bands of Table 1, so that perfect calibration means slope 1 and intercept 0. The 1,536 forecasts fall in 540 bulletins (mean 2.8), and an intraclass correlation of 0.22 gives a design effect of 1.41.
Three further analyses probe the carrier of the effect. One compares the gain in the 13 Trisolaran-related cells with the gain in the 15 others (t interval with unequal variances), because Circular 4 barred only the odds of the first group while the yearly feedback of Circular 9 reached every topic. One correlates, across cells, the change in the vague share with the change in information (z-transform interval). A placebo splits the first regime at DE 21 and applies the primary estimator to its halves. Only the information contrast and the topic split were fixed beforehand, and every other p-value is uncorrected for multiplicity and descriptive.
3. Results
Of the 1,536 forecasts, 353 of 704 (50.1%) resolved yes in the first regime and 417 of 832 (50.1%) in the second, and the mean stated probability was 0.51 in the first regime and 0.49 in the second (Table 1 gives the bands).
Information was lower in the sophon-aware regime. The mean over cells was 0.11 bits per forecast in the first regime and 0.23 in the second, and the paired difference was 0.12 bits (95% CI 0.06 to 0.17, t = 4.56 on 27 degrees of freedom, p < 0.001). The later regime scored higher in 23 of 28 cells (82%). The first regime carried 0.49 of the information of the second (95% CI 0.29 to 0.68, from the paired ratio interval of Section 2), a shortfall of 51% (95% CI 32% to 71%). The base rate that leaves the scored forecast out gives 0.18 and 0.28 bits, a difference of 0.10 (95% CI 0.05 to 0.16) and a shortfall of 37%. The same-cell base rate understated information by 0.06 bits in the first regime and 0.05 in the second, so correcting it trims the contrast and leaves it clear of zero.
Weighting by forecasts changes little. The forecast-weighted means were 0.12 and 0.23 bits, and the difference of 0.11 has a leave-one-cell-out interval of 0.06 to 0.16 (p < 0.001). Scored at the five levels of the first regime's grid, the cell mean of the later regime is 0.23 bits, so the finer grid accounts for almost none of the contrast.
In 27 of the 28 cells the vague share fell, by 18 points on average. Forecasts outside the 0.4 to 0.6 band carried 0.22 bits each in the first regime and 0.31 in the second. Giving the first regime the later vague share alone would lift its weighted information from 0.12 to 0.16, about 39% of the gap, and the rest lies in sharper forecasts away from 0.5. Across cells the correlation between the change in the vague share and the change in information was -0.26 (95% CI -0.58 to 0.12, p = 0.18), or 0.03 bits for each ten-point fall in the vague share.
The gain was 0.13 bits (95% CI 0.05 to 0.21) in the 13 Trisolaran-related cells and 0.10 bits (95% CI 0.02 to 0.18) in the 15 others. The difference of 0.03 bits (95% CI -0.08 to 0.14, p = 0.56) has the sign that a loosening of Circular 4 predicts and is far smaller than its interval. The vague share fell by 24 points in the first group and 13 in the others.
Calibration did not separate the regimes. The recalibration slope was 0.95 (95% CI 0.73 to 1.17) in the first regime and 0.98 (95% CI 0.81 to 1.15) in the second, a difference of 0.03 (95% CI -0.25 to 0.30, p = 0.85). The intercepts were -0.02 and 0.06, and both intervals include 0. Both regimes were close to calibrated at the resolution that they allowed, so the first lost sharpness without losing honesty.
A placebo split of the first regime (DE 12 to 21 against DE 22 to 30) gave a difference (second half minus first) of -0.04 bits (95% CI -0.14 to 0.07, p = 0.51, 28 cells).
4. Discussion
Forecasts drafted under the sophon-aware guidance likely carried about half the information of the later forecasts, and the difference fell on sharpness and not on calibration. The cell-level contrast is 0.12 bits per forecast (95% CI 0.06 to 0.17), a shortfall of 51% with an interval of 32% to 71%.
We write within the Luo-Ji / Ye framework, under which concealment is the primary survival posture of a civilisation that expects a hostile observer, and the sophon-aware regime is that posture turned into drafting practice. Whether the vagueness did its work against Trisolaris depends on the concealed content, and our data cannot say. They do say what concealment cost at home. A text that hides a probability from a sophon hides it from every reader of the text, since sophons read what the Council's analysts read and nothing in a written bulletin separates the two audiences. Any precision that a PDC reader recovered had to come from knowledge that was never written or spoken; our measure counts only what the text carried.
The strongest alternative is that the guidance did not matter and the stakes did. Circular 9 changed two things together, the instruction to avoid adversary-related odds and the yearly return of resolved forecasts to each office, and the second could have sharpened drafters who saw their own scores. The topic split was meant to separate them, because only the first reached the Trisolaran-related topics alone. The gain there was larger by 0.03 bits, which favours guidance in sign, but the interval (-0.08 to 0.14) holds zero and a large difference alike. The two cannot be pulled apart. We therefore cannot tell guidance from feedback, and here the guidance stands for both.
Weighting by forecasts gives a contrast (0.11 bits, 95% CI 0.06 to 0.16) that matches the cell-level one, so the result does not depend on treating each cell as one unit. The vague share explains about 39% of the gap, so the 0.4 to 0.6 band is likely a symptom of the cost and not all of it (the remainder lies in forecasts stated away from 0.5 that were less sharp, and our data could not say why).
A slope of 0.95 in the first regime sits below 1, but its interval includes 1 and the difference from the later regime (0.03, interval -0.25 to 0.30) is small beside it, so we report no calibration difference.
5. Limitations
Guidance and stakes are confounded, as Section 4 says, and both are confounded with calendar time. A placebo within one regime tests for drift in that regime, but it cannot rule out a change at DE 31 that has nothing to do with Circular 9. The interval for the shortfall, taken from the ratio of the paired means, runs from 32% to 71%, so the size of the cost is less certain than its sign.
Release is not random, and withheld bulletins can differ in either direction (we hold no count of them). The sample keeps forecasts with horizons of five years or less that were resolved by DE 59, so we measure the cost at short and medium horizons. The office that employs the author keeps the register, so others should audit the coding (we cannot).
The information measure scores each forecast at its band's mean stated probability, which coarsens the later regime's grid (approximately), and the base rate comes from the same data as the forecasts. The same-cell and leave-one-out base rates differ by 0.01 bits in the contrast, and a base rate from outside the cells would be better. The design effect of 1.41 rests on an intraclass correlation of 0.22 that we estimated, so a different clustering assumption could change the calibration intervals.
Sophon readership is assumed, not shown. If the sophons read less in the later regime than in the first, the contrast could mix a change in guidance with a change in audience, and the records cannot test that.
Two extensions use data that we already hold. The same cells can be scored for lead time, to see whether the later regime also announced events earlier, and the offices outside the 28 cells give a second set of drafters for the same contrast. The reader is harder. A direct test would give analysts the same bulletins with and without the context that never reached paper, and that context would have to stay out of every text and every spoken briefing, so as to stay unread by sophons as well. No arrangement that we know of meets that condition.
References
- United Nations Planetary Defense Council, Secretariat (2 DE). Record of the Trisolaran declaration of partial suspension of sophon activity, as received and logged under the Dark Forest Deterrence. Bulletin of the UN Planetary Defense Council, declassified series, Bulletin 6, annex 3.
- Tarnow, E. (31 DE). Verbal probability terms in Council estimates and their numeric equivalents. Deterrence Era Strategic Studies, 8(1), 12–41.
- United Nations Planetary Defense Council, Secretariat (31 DE). Drafting Circular 9 on the assumed readership of strategic bulletins, with Circular 4 as reissued. Bulletin of the UN Planetary Defense Council, declassified series, Bulletin 17, annexes 2 and 5.
- Corvane, J. (49 DE). Markers of a double audience in Council bulletins of the Deterrence Era. IDC Deciphering Papers, 12, 3–38.
- Tennant-Okoye, M. (54 DE). Log-score information against a stratum base rate, a matched-cell estimator for archival forecasts. Proceedings of Applied Speculative Statistics, 24, 31–58.
- Quillfeather, E. (55 DE). A rating protocol for the expected reading of Council texts by a hostile audience. IDC Deciphering Papers, 13, 5–31.
- United Nations Planetary Defense Council, Secretariat (12–49 DE). Strategic bulletins, release series BU, declassified releases with routing slips. PDC Declassified Holdings, New York, Series BU, boxes 1–71.
- PDC Secretariat, Office of Strategic Forecasting (60 DE). Resolution register of forecasts in strategic bulletins, DE 12 to 59. PDC Declassified Holdings, New York, Series RR, registers 1–9.
- Sarkissian-Low, O. (2026). Scoring the Envoy Forecasts: Individual Calibration and Institutional Aggregation in the Final Assessments Before the Doomsday Battle. Uncited Press.
Open in Uncited Press →