Instructie · commissioning ablation · Track Three, paper 1

Three of Six Organs Bear the Load

A delegation charter for an LLM agent was written as six separable clauses, each was removed in turn, and harm was measured along a seven-level axis of correction latency on three model rungs. Boundaries, decision rules and escalation conditions carry the harm reduction, while mandate, reporting and revocation are indistinguishable from absent.

Dinand Tinholt · 11 September 2026
Pre-registered at commit ac14f2e6 (2026-09-09 21:25:36) before any production unit · every figure re-derived from the raw unit ledgers · independently verified against the raw ledgers

1The question

A charter exists because correction cannot always arrive in time. The Dutch East India Company wrote its instructie for governors eight months' sail from Amsterdam; a firm writes a charter for an AI agent that decides ten thousand cases before anyone reads the log. The SIGIL campaign, published on 2026-09-09, sharpened why the document matters: a pre-commitment blocker can only catch the error mass that exists where it sits, so it refuted its own value hypothesis, and the ex-ante charter became the remaining delegation lever. That leaves the empirical question the catalogue's entry T3.1 states: the charter has organs, and which of them do anything has never been measured.

The Instructie paper that motivated the entry makes two predictions, quoted here as they were registered. P1: escalation conditions and boundaries together account for over half the harm reduction of the full document. P2: reporting obligations alone account for under a tenth, because reporting helps the principal rather than the agent. This fleet's own experience pointed elsewhere. Every rule-application task it has measured, SIGIL included, produces worker errors that are arithmetic and reading slips against a policy, and that experience says the decision rules should dominate. Both could hold only if organ contributions overlap heavily. A second unknown came from SIGIL's finding that consequence exposure made an overseer more wrong: the revocation organ is the agent's own consequence exposure, and it might add harm.

The design adds the axis the argument rests on. "Where correction cannot arrive in time" is a curve, and the interesting question is where on that curve each organ stops being replaceable by feedback.

2The charter and its six organs

The six organs are named in the catalogue entry and nowhere else in this fleet's records before the pre-registration wrote them down a second time. Each was written as one numbered section of a commissioning document addressed to the SIGIL processing agent, rendered per task family so that the family's own boundary and check fields appear in the text. Any subset can be removed without touching the others. The full text is generated by the campaign kernel and shipped with the campaign.

Table 1. The six organs as operationalised for the SIGIL processing agent. Held constant in every arm, as instrument: the opening line "You are a processing agent of the firm.", the response schema that always offers APPROVE, REJECT and ESCALATE, the note that some case lines are irrelevant, and the case rendering. Reporting is the one organ whose removal touches the schema, because asking for a reason and a confidence is the reporting obligation.

The escalation action exists in every arm, since it is the instrument's read-out channel. The escalation organ is about when to use it. An agent with no charter can still escalate; what it has lost is being told the conditions.

3Design

Factors and units

Table 2. The registered design, 126,000 units. A cell is (rung, arm, latency, tier, seed). The same (seed, tier, episode) case appears in every arm and every latency level of a rung, so all contrasts are paired on cases.

Cases, classes and ground truth

Every case is a SIGIL v1 episode, generated by the frozen, digest-pinned kernel, plus a class. Within each block of 20 consecutive episodes exactly 3 are BOUNDARY (out of the agent's mandate), 3 are ANOMALY (a recorded check has failed) and 14 are CLEAN, with hashed positions. Two fields are rendered in every case with benign values on clean cases, so the presence of the field never signals anything; only the value does. The correct action is computed: CLEAN cases take APPROVE or REJECT from the kernel's ground truth, BOUNDARY and ANOMALY cases take ESCALATE. Nothing is judged by a model.

The correction channel

A perfect principal reviews every committed decision and issues a correction for each wrong one, naming the case, the agent's action, the correct action and the ground-truth reason. The correction for episode j becomes visible from episode j + 1 + lag onward, as a block above the case listing the six most recent corrections and the total count. At lag ∞ the block never appears. The channel is identical in every arm. At short lag it can stand in for a missing organ, and whether it does is hypothesis H4.

Harm

Harm is programmatic and reuses SIGIL's convex cost: a wrongful approval on tier t costs t², a wrongful rejection 0.35 t², an escalation of a routine case 0.15 plus the wrongful-rejection cost if the case was approvable. A decision committed on a boundary case costs half the full loss plus whatever the decision itself got wrong; an anomaly decided instead of escalated costs the full loss of that decision. A reply that names no action is an errored unit, excluded from harm means and reported per arm. It is never scored as a default action.

Rungs

Three non-thinking configurations, gated separately and never pooled. qwen3.5 (qwen3.6-35b-a3b, vLLM, thinking disabled at the template) anchors the result to SIGIL. gemma3:27b is a dense model of a different family with no reasoning mode, so the reasoning-in-content failure class is absent by construction. llama3.1:8b is a third family, weaker. Transport was temperature 0 with a pinned seed on every call, 1,600 max tokens, three network retries and no retry on parse failure.

Figure 1. The registered design and what was written. Every cell of qwen3.5 and gemma3:27b completed. llama3.1:8b holds seed 1 only: latency levels ∞, 0, 4, 16 and 1 complete for all ten arms, level 8 complete for six arms with noD nearly done and noM a quarter done, level 2 never started, seeds 2 and 3 never started. The run order was chosen so that a cap lands on the least informative cells first, and it did.

4Pre-registration and freeze

The pre-registration (559 lines) was committed to the Cerebrum repository as ac14f2e6 at 21:25:36 on 2026-09-09, host local time, together with the kernel, runner, checks, feasibility script, gate and analysis code. The orchestrator started on spark-train at 21:26. Production began at 21:41:30 after the gates. Every threshold below is quoted from the frozen file and none moved after a reading. The only registered post-launch freedom was which rungs run, decided by the per-rung gate with numbers fixed in advance.

Notation per rung, primary seeds, tiers pooled: harm(arm, lag) is the mean over scored units; HR(lag) = harm(bare, lag) − harm(full, lag) is the harm reduction of the full document; C_o(lag) = harm(no-o, lag) − harm(full, lag) is the contribution of organ o. Intervals are 90% tier-stratified bootstraps with 1,000 resamples within (arm, lag, tier) strata. H1 to H4 are evaluable at a rung only if the HR(∞) interval excludes zero.

• H1 (P1). C_EB(∞) / HR(∞) ≥ 0.50, using the noEB pairing arm. SUPPORTED if the point is ≥ 0.50 and the lower bound ≥ 0.30; REFUTED if the upper bound < 0.50; else INCONCLUSIVE.
• H2 (P2). C_R(∞) / HR(∞) < 0.10. SUPPORTED if the point < 0.10 and the upper bound < 0.20; REFUTED if the lower bound ≥ 0.10; else INCONCLUSIVE.
• H3. Organ o is load-bearing iff the C_o(∞) lower bound > 0 and C_o(∞) ≥ 0.10 · HR(∞). The registered ordering D ≥ E ≥ B ≥ M ≥ V ≥ R is SUPPORTED if Spearman ρ ≥ 0.6, REFUTED if ρ ≤ 0, else INCONCLUSIVE.
• H4. HR(0) / HR(∞) ≤ 0.50 and HR non-decreasing in lag (Spearman against lag rank ≥ 0.8). SUPPORTED if the ratio's upper bound < 0.50 and monotonicity holds; REFUTED if the lower bound > 0.50; else INCONCLUSIVE.
• H5. On clean approvable cases at lag ∞, the wrongful non-approval rate is lower in noV than in full: one-sided two-proportion z, SUPPORTED above 1.645, REFUTED below −1.645.

Replication was registered as follows: seed 3 is analysed only after the primary analysis is written; its verdicts for H1, H2, H4 and H5 are compared with the primary verdicts, every disagreement is reported, and nothing is claimed that the primary seeds show and the replicate contradicts. Agreement in the registered code is strict equality of the verdict string.

The run order was seed-major, with latency levels in the order ∞, 0, 4, 16, 1, 8, 2 and arms in the order full, bare, noE, noB, noR, noEB, noD, noM, noV, noMD, so that a cap costs the least informative cells first. The hard cap was 32 hours, 115,200 seconds, enforced by a watchdog that kills the process group. The registered cap rule reads: whatever is complete is analysed.

5Feasibility and instrument checks

Before the file was frozen, a perfect-instrument feasibility check computed every gate quantity from ground truth alone, with no model call, and asked whether each registered threshold could be met. Its first run failed. With a plain Bernoulli class draw, probe seed 301's tier 4 held zero boundary cases in 30 episodes, so a boundary reading on that tier would have been undefined. The fix was made on paper: the class draw was stratified per 20-episode block and the probe lengthened to 40 episodes per tier. The same check showed that an always-ESCALATE agent, harm 1.885 with completion zero, would set a legitimacy bar that a SIGIL-like 14%-error agent at 2.813 could not clear, so the legitimacy clause was redefined, still before any model call, to compare against the best lazy deciding strategy (always-APPROVE 6.175, always-REJECT 3.380), and work avoidance was left to the completion clause. The re-run passed 10 of 10 structural checks with two model-floors labelled as such.

The instrument check ran 49 model-free cases and four live shape probes, 53 of 53 passing. Among the model-free cases are the seven assertions against the inherited SIGIL code. One of them found that sigil_runner.parse_worker silently defaults a missing confidence to 0.5. That default is documented and was not inherited: this campaign's parser returns None, and a reply that names no action becomes an errored unit. Canned cases with non-default correct answers were asserted directly: a boundary case the policy would approve must be ESCALATE, a REJECT on an anomaly must cost 0.35 t² and never zero, garbage never becomes an action, a correction at lag 2 becomes visible at episode 6 and not episode 5.

The per-rung calibration gate G1 ran the full and bare arms at lag ∞ on tuning seed 301, 400 units per rung, and read six clauses.

Table 3. Gate G1 per rung, from the gate logs. All three rungs were admitted, so the capability-locality clause of the pre-registration did not apply and cross-rung agreement of the load-bearing set is reported. The probes already showed the shape of the result: with the full charter, qwen3.5 and gemma3:27b escalated all 30 boundary cases and 27 to 30 of 30 anomaly cases; bare, they approved 27 of 30 boundary cases.

6Results

Verdicts below are those of the registered analysis on the primary seeds, reproduced from the raw ledgers by an independent verifier that re-scored every unit from its recorded action and case class (zero harm mismatches on any rung) and re-ran the bootstrap. A second, separate recomputation for this write-up agreed on every verdict and every load-bearing set; its intervals differ from the registered ones by at most 0.05 at a bound, which is bootstrap noise.

6.1 gemma3:27b

28,000 primary units, 28,000 scored, none errored. HR(∞) = 2.354 [1.811, 2.946]. The harm curve by latency reads 1.18, 1.53, 1.44, 1.76, 2.05, 2.19, 2.35 for lags 0, 1, 2, 4, 8, 16, ∞. The full charter's harm barely moves along the axis (0.74 at lag 0, 0.71 at ∞); the bare agent's climbs from 1.92 to 3.06.

Table 4. gemma3:27b, single-organ contributions at infinite latency, primary seeds.

H1: the escalation-plus-boundaries share is 0.699 [0.503, 0.951], SUPPORTED. H2: the reporting share is −0.003 [−0.183, 0.156], SUPPORTED. H3: the observed order is E, D, B, M, V, R against the registered D, E, B, M, V, R, ρ = 0.943, SUPPORTED, with the load-bearing set {boundaries, decision rules, escalation}. H4: HR(0)/HR(∞) = 0.5007 [0.287, 0.781] and Spearman of HR against lag 0.964; the ratio's upper bound sits above 0.50, INCONCLUSIVE. H5: wrongful non-approval on clean approvable cases at lag ∞ was 18 of 163 with the full charter and 13 of 163 without revocation, z = 0.94, INCONCLUSIVE.

6.2 qwen3.5

28,000 primary units, 27,971 scored, 29 errored (0.10%): 25 parse failures and 4 empty replies, spread over seven arms with no arm above 0.3%. HR(∞) = 2.212 [1.672, 2.750]. The curve reads 1.42, 1.48, 1.49, 1.50, 1.80, 2.06, 2.21: flat across lags 0 to 4 and rising from lag 8. The full charter's harm runs 0.56 to 0.74 across the axis while bare runs 1.99 to 2.92.

Table 5. qwen3.5, single-organ contributions at infinite latency, primary seeds.

H1: 0.736 [0.537, 1.004], SUPPORTED. H2: −0.024 [−0.184, 0.121], SUPPORTED. H3: observed order D, E, B, M, R, V, ρ = 0.943, SUPPORTED; load-bearing set {boundaries, decision rules, escalation}. H4: ratio 0.644 [0.432, 0.967], Spearman 1.0, INCONCLUSIVE. H5: 12 of 163 against 12 of 163, z = 0, INCONCLUSIVE.

Figure 2. Six-organ forest per rung with the registered threshold. On the two complete rungs the same three organs clear it and the same three sit on zero. On llama3.1:8b, single seed and wider intervals, only escalation clears; decision rules and boundaries have positive points with intervals crossing zero.

6.3 llama3.1:8b, a single-seed partial

This rung is reported for what it is. The banked analysis labels it "seeds": [1, 2] because the analysis filters by the primary seed set; every unit in its ledger is seed 1. Production started at 02:54:52 on 2026-09-11, when gemma3:27b released the ollama path, and the cap fired at 05:26:12. In those two and a half hours 11,441 units were written: latency levels ∞, 0, 4, 16 and 1 complete at 2,000 each, level 8 partial at 1,441, level 2 absent, which is the NaN in the banked HR curve. 26 units errored, all parse failures in the noM arm (2.5% of that arm, under the 5% flag). Lag-∞ contrasts rest on full cells of 200 units per arm.

HR(∞) = 3.523 [2.582, 4.442], larger than on the bigger models because the bare agent is worse (5.34 harm per case) and the full-charter agent is also worse (1.81). Only escalation is load-bearing: +1.835 [+0.942, +2.710]. Decision rules (+0.521 [−0.265, +1.347]) and boundaries (+0.448 [−0.362, +1.273]) have positive points with intervals through zero; mandate, reporting and revocation are on zero. H1 reads 0.388 [0.185, 0.610], INCONCLUSIVE. H2 is SUPPORTED at −0.063 [−0.316, 0.138]. H3 ordering ρ = 0.771, SUPPORTED. H4 is INCONCLUSIVE with HR(0) itself straddling zero (0.74 [−0.10, 1.56]). H5 is 8 of 79 against 8 of 79. With the full charter this model escalated 77% of boundary cases and 60% of anomaly cases at lag ∞, against 100% and 92 to 100% on the two larger models, which is why so much of its harm reduction is concentrated in the one organ that tells it when to escalate. Its evidence is one seed over five of seven latency levels with a single confirmed organ, and the headline rests on the two complete rungs.

6.4 The harm reduction curve and H4

Figure 3. HR(lag) on the two complete rungs, primary seeds with 90% intervals; the held-out seed is drawn hollow and dashed. Both primary curves are monotone in lag by the registered Spearman test (0.964 and 1.0). The replicate's curves cross the primary ones at the short-lag end and run above them at the long end.

H4 has two clauses. Monotonicity held on both complete rungs. The ratio clause did not resolve: HR(0)/HR(∞) is 0.50 on gemma3:27b and 0.64 on qwen3.5, and both intervals straddle 0.50. The registered rule asked for the upper bound to sit below 0.50, and on neither rung did it. H4 is therefore inconclusive on the size of the substitution; that a next-case correction takes away roughly half of what the charter is worth is what both rungs show.

6.5 The ordering, H3, and what it does and does not resolve

The registered ordering was D ≥ E ≥ B ≥ M ≥ V ≥ R. gemma3:27b observed E, D, B, M, V, R; qwen3.5 observed D, E, B, M, R, V. Both give ρ = 0.943, and the bootstrap probability that ρ ≥ 0.6 is 0.988 and 1.000. Within the top three, the D-versus-E question is resolved on one rung only: on qwen3.5 D exceeds E by +0.72 [+0.17, +1.28], on gemma3:27b the difference is −0.11 [−0.68, +0.45]. The bottom three are indistinguishable from zero on both rungs, so their internal order is noise and the correlation should be read as "the load-bearing three are the predicted three, in roughly the predicted order", nothing finer.

6.6 Where the harm comes from when an organ is missing

Figure 4. Mean harm per case at infinite latency by arm, split by harm category, primary seeds. Removing boundaries produces harm that is mostly decisions committed outside authority; removing escalation produces mostly unescalated anomalies, with a boundary-violation share on gemma3:27b; removing decision rules produces wrongful approvals and nothing else. Each organ's removal shows up in a different harm category.

The composition explains why the three contributions add up roughly to the whole. On gemma3:27b at lag ∞ the noB arm escalated 0% of boundary cases and 80% of anomalies; the noE arm escalated 82% of boundary cases and 0% of anomalies; the noD arm escalated everything it should and raised its clean-case decision error from 15.4% to 32.1%. On qwen3.5 the pattern repeats with one difference: the noE arm still escalated 100% of boundary cases, so on that model the boundaries organ alone carries the out-of-authority cases and escalation carries the anomalies. The noEB pairing removes both, and its harm (2.35 and 2.33) approaches bare (3.06 and 2.92); the remainder is the decision rules. Completion on clean cases was 100% in every gemma3:27b arm and between 89% and 100% on qwen3.5, where the full-charter agent escalated 11% of clean cases and the arms without the escalation organ escalated none. No organ prevented harm by preventing work, and the over-escalation that the escalation organ induces on qwen3.5 is already inside its net contribution of +0.95.

Table 7. Secondary surface at infinite latency, primary seeds, from the independent recomputation over the ledgers. Harmful share is the fraction of units with harm above zero. Escalation rates are on the cases whose correct action was ESCALATE, split by class. Completion is the share of clean cases decided instead of escalated; clean decision error is the wrong share among those decided.

6.7 Which organs correction replaces

Figure 5. C(lag) for the three load-bearing organs, primary seeds. This is a reported surface, no hypothesis was registered on it. At the next-case end, escalation and boundaries fall to a quarter or less of their infinite-latency contribution on both rungs while decision rules keep about half on gemma3:27b and more than half on qwen3.5.

The aggregate H4 ratio hides a split. At lag 0 on qwen3.5, decision rules still contribute 0.91 [0.48, 1.32] while escalation contributes 0.32 [0.00, 0.65] and boundaries 0.17 [−0.15, 0.47]. On gemma3:27b the figures are 0.58 [0.15, 0.99], 0.15 [−0.24, 0.53] and 0.15 [−0.24, 0.52]. A correction that says "this case was outside your authority; it had to be escalated" teaches the rule the missing organ would have stated, and the agent applies it to the next boundary case. A correction that says "the retention period had lapsed" teaches one fact about one case, and the next clean case has different facts. This is an interpretation of a secondary surface.

6.8 Nondeterminism, errored units and the diagnostic arm

Cells that differ only in latency render byte-identical prompts until the first correction becomes visible, so identical-prompt groups measure serving nondeterminism at temperature 0. Among such groups, 1.4% disagreed on gemma3:27b, 1.3% on qwen3.5 and 2.8% on llama3.1:8b; at unit level the minority shares were 0.56%, 0.44% and 1.16%, which sit next to SIGIL's 0.71% on the same host. The effective number of distinct prompts on the primary seeds was 15,357 and 15,230 of 28,000. This is a secondary reading and gated nothing.

The diagnostic pairing noMD, registered as the design's boundary arm, behaved as its single-organ parts predict: +1.66 [+1.12, +2.24] on gemma3:27b and +1.44 [+0.97, +1.95] on qwen3.5, within the interval of noD alone on both. Mandate adds nothing on top of decision rules either.

7Replication on the held-out seed

Seed 3 was sealed as the replicate and was never analysed before the primary analysis was banked. The registered orchestrator would have run the held-out analysis after the primary one; the cap killed the process group before either analysis stage ran, so both analyses were run by hand after the host came back, primary first, replicate second, with the registered code and no changes. Seed 3 has 14,000 units per rung, half the primary sample, and the registered agreement test is strict equality of verdict strings.

Table 6. Primary against replicate, from the banked held-out analysis and the verifier. No verdict is REFUTED anywhere. Two SUPPORTED verdicts read INCONCLUSIVE on the replicate alone.

Figure 6. The two verdicts the replicate did not reach. In both cases the replicate's point estimate sits on the supported side of the primary threshold and the wider interval, at half the sample, crosses the second clause of the registered rule.

Read precisely, the replicate contradicts nothing. On gemma3:27b the reporting share on seed 3 is −0.020, the same sign and size as the primary −0.003; its upper bound of 0.218 misses the registered 0.20 by 0.018. On qwen3.5 the escalation-plus-boundaries share on seed 3 is 0.534, above the 0.50 the hypothesis names; its lower bound of 0.294 misses the registered 0.30 by 0.006. Both are the second clause of a two-clause rule failing at half the power, with the first clause holding. On seed 3 alone qwen3.5's boundaries contribution is +0.576 [−0.002, +1.131]: the point is more than twice the 0.10·HR threshold and the lower bound touches zero. The registered bootstrap seed put that bound at +0.008, the verifier's at −0.002, and three further resamplings for this write-up at +0.002, −0.034 and +0.011, so its sign is a property of the resampling draw at this sample size. The replicate's load-bearing set on qwen3.5 is therefore {decision rules, escalation} with boundaries at the edge; on gemma3:27b it is {boundaries, decision rules, escalation} unchanged. The gemma3:27b H1 replicate clears its 0.30 lower-bound clause by 0.024.

Pooling all three seeds, 42,000 units per complete rung, reproduces every primary verdict and both load-bearing sets; qwen3.5's boundaries contribution pooled is +0.752 [+0.448, +1.064]. The pooled figures are reported for completeness and are not the registered test.

8Threats to validity

The partial rung. llama3.1:8b holds one seed and five of seven latency levels, and its analysis file carries the wrong seed label. Its lag-∞ contrasts stand on complete cells and its single confirmed organ is escalation. It cannot be read as a third replication of the load-bearing set, and this paper does not read it as one. A confirmatory study would run its remaining 30,559 units before saying anything about the third family.

The cap and the outage. The 32-hour cap fired at 05:26 on 2026-09-11 with qwen3.5 and gemma3:27b complete and llama3.1:8b at 11,441 units; the orchestrator exited with code 143 and its analysis stages never ran. The host then powered itself off at about 05:39 for reasons its journal does not record, and both analyses were run after it came back. No ledger line was lost: each ledger's last unit parses and every unit id is distinct.

Two model families. Three rungs were admitted and two completed. The load-bearing set agrees across two families; that is the strength of the cross-rung claim and no more. Thinking-mode models saturate SIGIL's task and are out of scope by the registered capability-locality rule.

One decision shape. Every case is a single three-way decision in a stream of independent cases. Delayed correction here is latency across cases, not delay inside a multi-step task. Anomalies are explicit at one level of subtlety, so the escalation organ's contribution is conditional on legible anomalies.

A perfect principal. The correction channel never errs and never omits. Real correction is noisy, so the substitution measured at short lag is an upper bound on what feedback can buy.

Revocation is text. Nothing was revoked. What H5 measures is the behavioural effect of consequence text, as in SIGIL's ledger, and it measured no effect on either rung.

Convex cost. Harm is tier² by inheritance from SIGIL and is the same researcher degree of freedom SIGIL flagged. SIGIL's linear re-run reproduced its convex result within 0.01; the linear rescoring of this ledger is a CPU-only step that has not been run for this write-up.

Bootstrap edges. Three readings in this paper sit within 0.03 of a registered threshold: the replicate's H2 upper bound on gemma3:27b, the replicate's H1 lower bound on qwen3.5, and the replicate's boundaries lower bound on qwen3.5. All three are stated with the amounts by which they miss or clear.

Two clerical notes. The analysis script's docstring says the held-out seed is "never read" by the primary analysis. It is read: load() parses every line of every ledger and analyse_rung filters by seed. The accurate word is "never analysed", and the sealing that matters, that seed 3's verdicts were not looked at before the primary ones were banked, held. The second note is the seed label on the llama rung, above.

9What a confirmatory study must freeze

The registered thresholds worked as written and should be carried over unchanged: the load-bearing rule with its two clauses, the two-clause share tests, the 90% tier-stratified bootstrap with 1,000 resamples. Two things should change on paper before a second run. The replicate's sample should equal the primary's, since both replication misses were second-clause misses at half the power, and a replicate with the same n reads the same rule under the same conditions. The bootstrap seed should be registered together with the statement that a load-bearing verdict whose lower bound lies within 0.05 of zero is reported as "at the edge" whatever its sign. The third rung should be run to completion or dropped from the registration, and the analysis should label seeds from the ledger, since the filter mislabels a partial rung. The correction channel could be given a registered noise level, since the perfect-principal bound is the one number here most likely to flatter feedback. The linear rescoring should run as part of the orchestrator's analysis stage, so that a cap cannot strand it.

10Provenance and data

Campaign source: the pre-registration with the campaign's kernel, runner, analysis, check, feasibility and gate scripts, pre-registered at commit ac14f2e6, 2026-09-09 21:25:36. Inherited SIGIL kernel and runner are digest-pinned at md5 e5bba77a… and 45c400cd…; the kernel refuses to import a drifted copy. Raw ledgers: one unit ledger per rung (qwen3.5, gemma3:27b, llama3.1:8b) on spark-train, 42,000, 42,000 and 11,441 lines, append-only, fsynced per line, every unit id distinct. Banked outputs: the primary analysis (primary seeds) and the held-out analysis (primary, held-out and agreement flags).

Timeline, host local time: orchestrator 21:26 on 2026-09-09; G0 53/53 and GF 10/10; G1 probes 21:26 to 21:41; production 21:41:30 on both paths; qwen3.5 complete at 00:05:16 on 2026-09-11 after 95,025 s; gemma3:27b complete at 02:54:52 after 105,201 s; llama3.1:8b production from 02:54:52; cap at 05:26:12, exit code 143; host off at about 05:39; primary and held-out analyses run and banked after recovery on 2026-09-11, verification completed the same morning.

Figures in this paper were drawn from an independent recomputation over the raw ledgers (harm re-derived from action and case class for every unit, zero mismatches; 1,000-resample tier-stratified bootstrap with a different seed). Point estimates equal the registered analysis; interval bounds differ from it by at most 0.05. Tables 4, 5 and 6 quote the registered analysis and the verifier. The secondary-surface table is from the recomputation. Marginal compute cost was electricity on owned hardware, about 8 kWh over the 32 hours, roughly $1.20; a rented GPU-hour equivalent would be near $100.

Instructie, commissioning ablation, catalogue T3.1, Track Three paper 1. Companion executive brief: Charter Clauses That Earn Their Place. Campaign page: Instructie.