@event4u/agent-config 14.11.0 → 14.13.0

package/src/scripts/_lib/config_cost.ts

@@ -0,0 +1,209 @@
+/**
+ * Delivered-cost primitives (`road-to-delivered-cost-truth` Phase 1).
+ *
+ * Three answers a consumer of this suite cannot currently get: what does my
+ * configuration cost per session, which assets produce that cost, and is any of
+ * it taking effect. This module owns the first; `_lib/asset_delivery_ledger.ts`
+ * owns the second.
+ *
+ * PAYLOAD AMPLIFICATION, NOT "NET NEGATIVE" — the naming is a council verdict,
+ * not a preference. AI council 2026-08-28 (2/2 convergent) resolved the
+ * `what-the-net-negative-threshold-is` blocker as option (b), a ratio against
+ * the active profile's declared payload, AND both seats refused the name the
+ * roadmap used: a ratio measures how much payload is delivered per unit the
+ * profile declared. It cannot establish that the configuration is net-negative,
+ * because it never observes what was RETURNED. A 10:1 amplification delivering
+ * critical capability is not net-negative; a 1:1 delivering nothing is. So the
+ * figure informs the question and is not allowed to answer it, and the field is
+ * named for what it measures.
+ *
+ * THE DENOMINATOR IS THE INTENDED PAYLOAD, NOT THE MEASURED ONE. A first cut
+ * declared the profile's payload as its own current measurement, which made the
+ * ratio 1.00 by construction — a number that cannot move is not a measurement.
+ * The declaration is what the profile INTENDS to deliver (its budget); the
+ * numerator is what the tree actually delivers; the ratio is the gap, which is
+ * the whole point.
+ *
+ * AND IT IS VERIFIED, NOT TRUSTED. Both council seats independently raised
+ * denominator gaming: a profile self-declaring a large payload makes any
+ * delivered cost look good. So a declaration above the measured tree's own total
+ * is refused — a profile cannot intend to deliver more than exists — and an
+ * absent, non-positive or over-ceiling declaration yields `unknown_profile`,
+ * never a ratio computed from a number nobody checked.
+ */
+
+export type AmplificationVerdict = 'measured' | 'unknown_profile';
+
+export interface ProfileDeclaration {
+    profile: string;
+    /**
+     * INTENDED standing payload per category — the profile's budget, not its
+     * current measurement. Per category so the numerator and denominator
+     * measure the same things; a single total would let a rules figure be
+     * compared against a rules-plus-skills delivery.
+     */
+    declared: { rules_tokens: number; skill_catalogue_tokens: number };
+    /**
+     * How far ABOVE the measured tree a declaration may sit before it is
+     * refused as gamed. A profile cannot intend to deliver more than exists.
+     */
+    sanity_headroom: number;
+}
+
+export interface AmplificationResult {
+    verdict: AmplificationVerdict;
+    /** `null` whenever `verdict` is not `measured`. Never a placeholder number. */
+    ratio: number | null;
+    delivered_tokens: number;
+    declared_tokens: number | null;
+    reason: string;
+}
+
+/** A declaration is usable when it is present, positive, and not above the tree. */
+export function amplification(
+    deliveredTokens: number,
+    measured: { rules_tokens: number; skill_catalogue_tokens: number },
+    declaration: ProfileDeclaration | null,
+): AmplificationResult {
+    if (declaration === null) {
+        return {
+            verdict: 'unknown_profile',
+            ratio: null,
+            delivered_tokens: deliveredTokens,
+            declared_tokens: null,
+            reason: 'the active profile declares no intended standing payload, so there is no denominator to compare against',
+        };
+    }
+    const declaredTotal = declaration.declared.rules_tokens + declaration.declared.skill_catalogue_tokens;
+    if (!Number.isFinite(declaredTotal) || declaredTotal <= 0) {
+        return {
+            verdict: 'unknown_profile',
+            ratio: null,
+            delivered_tokens: deliveredTokens,
+            declared_tokens: null,
+            reason: `profile '${declaration.profile}' declares a non-positive payload (${String(declaredTotal)}), which cannot be a denominator`,
+        };
+    }
+
+    // The gaming direction, and the only one that matters: a declaration ABOVE
+    // what the tree can deliver drives the ratio below 1 and makes any delivery
+    // look disciplined. A declaration BELOW the delivery is not gaming — it is
+    // the finding, and it is exactly what this ratio exists to surface.
+    const measuredTotal = measured.rules_tokens + measured.skill_catalogue_tokens;
+    const ceiling = measuredTotal * (1 + declaration.sanity_headroom);
+    if (declaredTotal > ceiling) {
+        return {
+            verdict: 'unknown_profile',
+            ratio: null,
+            delivered_tokens: deliveredTokens,
+            declared_tokens: declaredTotal,
+            reason:
+                `profile '${declaration.profile}' declares ${String(declaredTotal)} tok of intended payload against a ` +
+                `measured tree of ${String(measuredTotal)} tok — a profile cannot intend to deliver more than exists, ` +
+                'so the declaration is refused rather than used as a denominator',
+        };
+    }
+
+    return {
+        verdict: 'measured',
+        ratio: deliveredTokens / declaredTotal,
+        delivered_tokens: deliveredTokens,
+        declared_tokens: declaredTotal,
+        reason:
+            `${String(deliveredTokens)} tok delivered against ${String(declaredTotal)} tok intended by profile ` +
+            `'${declaration.profile}'`,
+    };
+}
+
+// ─────────────────────────────────────────────────── the billable-input split
+
+export interface BillableSplit {
+    fresh_input_tokens: number;
+    cache_read_tokens: number;
+    cache_creation_tokens: number;
+    billable_input_tokens: number;
+    /** cache_read / billable_input. `null` when nothing was billed. */
+    cache_read_share: number | null;
+}
+
+export function billableSplit(fresh: number, cacheRead: number, cacheCreation: number): BillableSplit {
+    const total = fresh + cacheRead + cacheCreation;
+    return {
+        fresh_input_tokens: fresh,
+        cache_read_tokens: cacheRead,
+        cache_creation_tokens: cacheCreation,
+        billable_input_tokens: total,
+        // A share over a zero denominator is not 0 % — it is absent, and
+        // rendering 0 % would read as "the cache never helped".
+        cache_read_share: total > 0 ? cacheRead / total : null,
+    };
+}
+
+// ──────────────────────────────────────────────────── re-read cost (step 1.3)
+
+export interface RereadLeg {
+    /** Repo-relative path that was re-read. */
+    file: string;
+    /** Total reads of this file in the leg. */
+    total_reads: number;
+    /** Reads beyond the first. */
+    duplicate_reads: number;
+    /** Measured size of the file in tokens. */
+    file_tokens: number;
+}
+
+export interface RereadCost {
+    legs: RereadLeg[];
+    /** Tokens paid for content already in context. */
+    wasted_tokens: number;
+    /** `null` when nothing was re-read — never a zero standing in for absence. */
+    worst_file: string | null;
+}
+
+/**
+ * Price the re-reads `hot_context_hook` already counts.
+ *
+ * The hook notices a repeated read and says so; nothing turned that into a
+ * number, so a re-read was visible and unpriced. The cost is
+ * `duplicate_reads x file_tokens` — a figure DERIVED from the file's measured
+ * size, never a constant, because a constant would rank a re-read of a one-line
+ * config beside a re-read of a 3,000-token rule.
+ */
+export function rereadCost(legs: readonly RereadLeg[]): RereadCost {
+    const wasted = legs.reduce((n, l) => n + l.duplicate_reads * l.file_tokens, 0);
+    let worst: RereadLeg | null = null;
+    for (const l of legs) {
+        const cost = l.duplicate_reads * l.file_tokens;
+        if (cost > 0 && (worst === null || cost > worst.duplicate_reads * worst.file_tokens)) worst = l;
+    }
+    return { legs: [...legs], wasted_tokens: wasted, worst_file: worst === null ? null : worst.file };
+}
+
+// ────────────────────────────────────────────── the summary line (step 2.3)
+
+/**
+ * The one-line end-of-task figure. TOKENS ONLY.
+ *
+ * No currency, and this is a hard property rather than a style choice: the suite
+ * does not know the consumer's contract — subscription, per-token, committed
+ * spend, a reseller rate — so any monetary figure it printed would be
+ * extrapolated from a rate it invented. A fabricated cost is worse than none,
+ * because it is actionable and wrong.
+ *
+ * {@link CURRENCY_MARKERS} is what the test asserts against, so the ban is
+ * checkable rather than merely intended.
+ */
+export const CURRENCY_MARKERS = ['$', '€', '£', '¥', 'USD', 'EUR', 'GBP', '/1k', 'per token', 'per-token'] as const;
+
+export function summaryLine(deliveredTokens: number, split: BillableSplit | null): string {
+    const share =
+        split === null || split.cache_read_share === null
+            ? 'cache-read share unavailable'
+            : `cache-read share ${(split.cache_read_share * 100).toFixed(1)}%`;
+    return `Delivered standing payload: ${String(deliveredTokens)} tok · ${share}`;
+}
+
+/** True when a rendered string carries no currency figure or per-token rate. */
+export function isCurrencyFree(text: string): boolean {
+    return !CURRENCY_MARKERS.some((m) => text.includes(m));
+}

package/src/scripts/_lib/council_settings_block.ts

@@ -94,6 +94,12 @@ export function synthesizeAiCouncilBlock(cfg: CouncilConfig): Dict {
             strong_threshold: cfg.consensus_scoring.strong_threshold,
             minority_threshold: cfg.consensus_scoring.minority_threshold,
             lenses: [...cfg.consensus_scoring.lenses],
+            // Phase 1B. Omitting it here is not a missing default — it is the key
+            // becoming UNREADABLE: `inlineFindingsActive` reads this projected
+            // dict, not the typed config, so an absent key resolves to `false` and
+            // the feature is silently off no matter what the YAML says. A live
+            // analysis run on 2026-08-30 is what found it, after the tests passed.
+            inline_findings: cfg.consensus_scoring.inline_findings,
         },
         cli_call_budget: {
             max_calls_per_day: _mapToObject(cfg.cli_call_budget.max_calls_per_day),

package/src/scripts/_lib/curator_ops.ts

@@ -0,0 +1,172 @@
+/**
+ * The curator's operation set — SEVEN ops per decision E6 — and the
+ * deterministic near-duplicate pre-stage that runs before any model judgment.
+ *
+ * `road-to-governed-harness-evolution` Phase 5, step 5.5.
+ *
+ * > *Curator operation set per E6. Candidates only, never promotions. Run
+ * > `src/scripts/_lib/shingle_similarity.ts` as a deterministic pre-stage
+ * > before any model judgment.*
+ * > verify: **a near-duplicate candidate is caught by the similarity stage with
+ * > zero model calls.**
+ *
+ * ## Seven, and why the middle answer was rejected
+ *
+ * E6 (decided 2026-08-30, AI council, anthropic + openai, 2/2) settled this on
+ * a contradiction rather than a preference: step 7.6 of the same roadmap
+ * already specifies the verdict set `KEEP / REVISE / MERGE / SPLIT / RETIRE`,
+ * so a four-op curator would emit verdicts it cannot execute. The six-op middle
+ * — deferring `SPLIT` — was rejected explicitly: it still contradicts 7.6, so
+ * it buys nothing the four-op answer does not also cost. The algebra argument
+ * is the one worth carrying: `MERGE` is n→1 and `SPLIT` is 1→n, and without
+ * `SPLIT` an overgrown rule becoming two must be written as
+ * `RETIRE + 2 x ADD`, which loses the semantic link and makes the intent
+ * unreadable in the audit log. {@link validateOp} encodes exactly that arity.
+ *
+ * ## Candidates only — carried by the type, not by a promise
+ *
+ * Every screened proposal carries `lifecycle: 'candidate'` as a LITERAL type,
+ * so there is no value a curator can construct that names any other lifecycle.
+ * Phase 7 is gated on the OPEN `merge-authority` blocker; a curator that could
+ * express a promotion would be one edit from performing one.
+ *
+ * ## Zero model calls, and how that is established rather than asserted
+ *
+ * {@link screenNearDuplicates} is synchronous, imports only
+ * `shingle_similarity`, and returns `model_calls: 0` as a literal-typed field.
+ * `tests/scripts/curator_ops.test.ts` proves the property two ways: it scans
+ * this module and its one dependency for network and model-client constructs in
+ * both polarities, and it runs the screen with `globalThis.fetch` replaced by a
+ * throwing stub, so a call would surface as a thrown error rather than as an
+ * unchecked assumption. An assertion by inspection would be worth nothing here
+ * — the whole point of a deterministic pre-stage is that it is cheaper than the
+ * judgment it precedes.
+ */
+import { shingleOverlap } from './shingle_similarity.js';
+
+/** E6, option B. The exact set — arity and membership are both pinned by the test. */
+export const CURATOR_OPS = [
+    'KEEP',
+    'ADD',
+    'MERGE',
+    'REPLACE',
+    'SPLIT',
+    'RETIRE',
+    'SKIP',
+] as const;
+
+export type CuratorOp = (typeof CURATOR_OPS)[number];
+
+export interface CuratorProposal {
+    id: string;
+    op: CuratorOp;
+    /** Existing artifact ids the op reads or consumes. */
+    targets: readonly string[];
+    /** Artifact ids the op would produce. Empty for read-only ops. */
+    produces: readonly string[];
+    /** The proposed text, which is what the similarity stage compares. */
+    text: string;
+}
+
+/** Arity per op. `null` means the op places no constraint on that side. */
+const OP_ARITY: Record<CuratorOp, { targets: [number, number | null]; produces: [number, number | null] }> = {
+    KEEP: { targets: [1, 1], produces: [0, 0] },
+    ADD: { targets: [0, 0], produces: [1, 1] },
+    MERGE: { targets: [2, null], produces: [1, 1] },
+    REPLACE: { targets: [1, 1], produces: [1, 1] },
+    SPLIT: { targets: [1, 1], produces: [2, null] },
+    RETIRE: { targets: [1, 1], produces: [0, 0] },
+    SKIP: { targets: [0, null], produces: [0, 0] },
+};
+
+function arityError(label: string, n: number, [lo, hi]: [number, number | null]): string | null {
+    if (n < lo) return `${label}: expected at least ${String(lo)}, got ${String(n)}`;
+    if (hi !== null && n > hi) return `${label}: expected at most ${String(hi)}, got ${String(n)}`;
+    return null;
+}
+
+/** `null` when the proposal's arity matches its op, otherwise the reason it does not. */
+export function validateOp(p: CuratorProposal): string | null {
+    const spec = OP_ARITY[p.op];
+    return (
+        arityError(`${p.op} targets`, p.targets.length, spec.targets) ??
+        arityError(`${p.op} produces`, p.produces.length, spec.produces)
+    );
+}
+
+/**
+ * Containment-overlap percent at which a proposal is a near-duplicate.
+ *
+ * A STATED default, not a measured optimum: `lint_originality`'s own gate uses
+ * shingle containment on the same primitive, and 70 % is the conservative end
+ * of the range a reskin scores there. `revisit-if` a screening run rejects a
+ * proposal a curator then has to re-add by hand, or admits one a human calls a
+ * duplicate.
+ */
+export const NEAR_DUPLICATE_THRESHOLD = 70;
+
+export interface CorpusEntry {
+    id: string;
+    text: string;
+}
+
+export interface NearDuplicateFinding {
+    proposal_id: string;
+    /** The corpus entry it duplicates. */
+    against: string;
+    overlap_percent: number;
+}
+
+export interface ScreenedProposal {
+    proposal: CuratorProposal;
+    /** Literal. There is no other value this field can hold. */
+    lifecycle: 'candidate';
+}
+
+export interface ScreenResult {
+    admitted: readonly ScreenedProposal[];
+    near_duplicates: readonly NearDuplicateFinding[];
+    malformed: readonly { proposal_id: string; reason: string }[];
+    /** Literal 0. The stage is deterministic; a model call here would be a defect. */
+    model_calls: 0;
+}
+
+/**
+ * The deterministic pre-stage: reject malformed ops and near-duplicates before
+ * anything expensive runs.
+ *
+ * Synchronous on purpose — a synchronous function cannot await a network call,
+ * so the zero-model-call property is a consequence of the signature rather than
+ * a claim the body has to keep.
+ */
+export function screenNearDuplicates(
+    proposals: readonly CuratorProposal[],
+    corpus: readonly CorpusEntry[],
+    threshold: number = NEAR_DUPLICATE_THRESHOLD,
+): ScreenResult {
+    const admitted: ScreenedProposal[] = [];
+    const nearDuplicates: NearDuplicateFinding[] = [];
+    const malformed: { proposal_id: string; reason: string }[] = [];
+
+    for (const p of proposals) {
+        const shapeError = validateOp(p);
+        if (shapeError !== null) {
+            malformed.push({ proposal_id: p.id, reason: shapeError });
+            continue;
+        }
+        let worst: NearDuplicateFinding | null = null;
+        for (const entry of corpus) {
+            const overlap = shingleOverlap(p.text, entry.text);
+            if (overlap >= threshold && (worst === null || overlap > worst.overlap_percent)) {
+                worst = { proposal_id: p.id, against: entry.id, overlap_percent: overlap };
+            }
+        }
+        if (worst !== null) {
+            nearDuplicates.push(worst);
+            continue;
+        }
+        admitted.push({ proposal: p, lifecycle: 'candidate' });
+    }
+
+    return { admitted, near_duplicates: nearDuplicates, malformed, model_calls: 0 };
+}

package/src/scripts/_lib/delivery_arm_experiment.ts

@@ -0,0 +1,266 @@
+/**
+ * The three-arm delivery experiment — `road-to-governed-harness-evolution` 6.1.
+ *
+ * WHAT 6.1 ASKS, AND WHY IT IS NOT THE PRICE GRID. Step 6.1 says: measure
+ * `eager-all` against `thin` against `delivery` — the three values
+ * `_lib/lean_projection_mode.ts:19` already declares — BEFORE any new retrieval
+ * component is written. `model_rule_injection.ts` already prices those three
+ * shapes, and a price grid is a cost model rather than a delivery measurement:
+ * it reports what each shape COSTS and never what each shape DELIVERS. The
+ * three arms differ on exactly one observable, and it is not cost:
+ *
+ *   for a given prompt, is the labelled rule's BODY in the model's context?
+ *
+ * That question is decidable from the tree, so this module answers it for all
+ * three arms over one corpus with one matcher, and reports the cost figures
+ * beside it so the arms are commensurable on one table.
+ *
+ * WHAT EACH ARM PUTS IN CONTEXT, read off the two consumers of the mode rather
+ * than assumed. `condense.ts:1124` calls `writesThinFiles(mode)`, so `thin` and
+ * `delivery` project pointer stubs where `eager-all` projects bodies;
+ * `hooks/rule_inject_hook.ts:230` calls `deliversBodies(mode)`, so only
+ * `delivery` binds the concern that injects matched bodies back at prompt time.
+ * Hence:
+ *
+ *   eager-all  every projected body is standing.
+ *   thin       kernel bodies + no-trigger residue bodies are standing; every
+ *              other body is a pointer and is NOT in context. Nothing is
+ *              injected, because the concern is not bound.
+ *   delivery   thin's standing set, PLUS the bodies `selectForInjection`
+ *              returns for this prompt under the concern's byte cap.
+ *
+ * The `thin` standing set mirrors `standingCorpora` in `model_rule_injection.ts`
+ * — kernel or triggerless keeps a body, anything else becomes a pointer — so the
+ * delivery half and the cost half classify each rule the same way. Two
+ * classifications would be the "measures nothing" defect `_lib/rule_injection.ts`
+ * names in its own header.
+ *
+ * NO SECOND MATCHER, WHICH IS 6.2'S CONSTRAINT AND THIS MODULE'S TOO. Every
+ * match here comes from `matchTierRules`, and every selection from
+ * `selectForInjection` — both in `_lib/rule_injection.ts`, which owns no matcher
+ * of its own and wraps `_lib/router_match.ts`. This module imports nothing else
+ * that could answer "which rules fire on this prompt?", and
+ * `tests/scripts/single_matcher_preserved.test.ts` fails if it ever does.
+ *
+ * ZERO MODEL CALLS, and the property follows from the imports rather than from
+ * good behaviour: the only import is `_lib/rule_injection.js`, which reads files.
+ * Token counting and the standing-corpus sizes are PARAMETERS, not imports, so
+ * this module carries no tokenizer and no network surface at all.
+ *
+ * WHAT IT DOES NOT MEASURE, stated because the gap is the interesting half.
+ * It does not measure whether a session that RECEIVES a body behaves like a
+ * session that HAD it standing. That is not an expense this run declined to pay
+ * — the instrument is closed by ADR-202 (`docs/decisions/ADR-202-anchor-scoring-as-thin-quality-instrument.md`),
+ * whose status reads "instrument not achievable with available evaluators" at an
+ * inter-evaluator kappa of 0.472 against a registered floor of 0.800, with no
+ * third attempt licensed. Nothing printed here is evidence about quality.
+ */
+import {
+    kernelIds,
+    loadRuleBody,
+    matchTierRules,
+    selectForInjection,
+    triggerlessRuleIds,
+    type Router,
+} from './rule_injection.js';
+
+/** The three arms, named exactly as `LeanProjectionMode` declares them. */
+export const ARMS = ['eager-all', 'thin', 'delivery'] as const;
+export type ArmName = (typeof ARMS)[number];
+
+/** The corpus fields this experiment reads. Structurally a `CorpusCase`. */
+export interface LabelledCase {
+    readonly rule: string;
+    readonly prompt: string;
+    readonly openFiles: string[] | null;
+    readonly label: 'positive' | 'near_miss';
+}
+
+export interface ArmRow {
+    readonly arm: ArmName;
+    /** Positives whose labelled rule body is in context under this arm. */
+    readonly delivered: number;
+    readonly positives: number;
+    /** Near-miss prompts whose labelled rule body is nevertheless in context. */
+    readonly falseContext: number;
+    readonly nearMisses: number;
+    /** Positives the matcher fired on but the byte cap then dropped. */
+    readonly capDropped: number;
+    /** Exact-BPE tokens standing in context before the first prompt. */
+    readonly standingTokens: number;
+    /** Mean exact-BPE tokens injected per positive prompt. */
+    readonly injectedMeanTokens: number;
+    /** Same, at the 90th percentile. */
+    readonly injectedP90Tokens: number;
+}
+
+export interface ArmExperimentInput {
+    readonly repoRoot: string;
+    readonly router: Router;
+    readonly cases: readonly LabelledCase[];
+    /** The concern's per-prompt byte cap. The sensitivity handle. */
+    readonly capBytes: number;
+    /** Standing exact-BPE tokens per arm, supplied by the caller. */
+    readonly standing: Readonly<Record<ArmName, number>>;
+    /** Exact-BPE tokenizer, injected so this module imports none. */
+    readonly tokensOf: (text: string) => number;
+}
+
+/**
+ * Ids whose FULL body is standing under `thin` and `delivery`.
+ *
+ * Kernel is standing by definition; a rule with no trigger cannot be delivered,
+ * so the projector keeps it eager under the thin shapes too.
+ */
+export function thinStandingIds(router: Router): Set<string> {
+    const ids = new Set<string>(kernelIds(router));
+    for (const id of triggerlessRuleIds(router)) ids.add(id);
+    return ids;
+}
+
+/** Nearest-rank quantile over a copy. Deterministic. */
+function quantile(values: readonly number[], q: number): number {
+    if (values.length === 0) return 0;
+    const s = [...values].sort((a, b) => a - b);
+    const idx = Math.min(s.length - 1, Math.max(0, Math.ceil(q * s.length) - 1));
+    return s[idx] as number;
+}
+
+function mean(values: readonly number[]): number {
+    if (values.length === 0) return 0;
+    return values.reduce((a, b) => a + b, 0) / values.length;
+}
+
+/** Ids projected with a real body file — the `eager-all` standing set. */
+function hasBody(repoRoot: string, id: string): boolean {
+    return loadRuleBody(repoRoot, id) !== null;
+}
+
+/**
+ * Run the experiment. Pure over its inputs; two runs are byte-identical.
+ *
+ * One walk over the corpus computes all three arms, so no arm can be scored on
+ * a different case set than another.
+ */
+export function runArmExperiment(input: ArmExperimentInput): ArmRow[] {
+    const { repoRoot, router, cases, capBytes, standing, tokensOf } = input;
+    const thinStanding = thinStandingIds(router);
+
+    const delivered: Record<ArmName, number> = { 'eager-all': 0, thin: 0, delivery: 0 };
+    const falseContext: Record<ArmName, number> = { 'eager-all': 0, thin: 0, delivery: 0 };
+    let positives = 0;
+    let nearMisses = 0;
+    let capDropped = 0;
+    const injectedTokens: number[] = [];
+
+    for (const c of cases) {
+        const inEager = hasBody(repoRoot, c.rule);
+        const inThin = thinStanding.has(c.rule) && inEager;
+
+        const matches = matchTierRules(router, c.prompt, c.openFiles);
+        const sel = selectForInjection(repoRoot, matches, capBytes);
+        const injectedIds = new Set(sel.selected.map((m) => m.id));
+        const inDelivery = inThin || injectedIds.has(c.rule);
+
+        if (c.label === 'positive') {
+            positives += 1;
+            if (inEager) delivered['eager-all'] += 1;
+            if (inThin) delivered.thin += 1;
+            if (inDelivery) delivered.delivery += 1;
+            if (sel.dropped.some((m) => m.id === c.rule)) capDropped += 1;
+            let tokens = 0;
+            for (const m of sel.selected) tokens += tokensOf(loadRuleBody(repoRoot, m.id) ?? '');
+            injectedTokens.push(tokens);
+        } else {
+            nearMisses += 1;
+            if (inEager) falseContext['eager-all'] += 1;
+            if (inThin) falseContext.thin += 1;
+            if (inDelivery) falseContext.delivery += 1;
+        }
+    }
+
+    const injMean = Math.round(mean(injectedTokens));
+    const injP90 = quantile(injectedTokens, 0.9);
+    return ARMS.map((arm) => ({
+        arm,
+        delivered: delivered[arm],
+        positives,
+        falseContext: falseContext[arm],
+        nearMisses,
+        capDropped: arm === 'delivery' ? capDropped : 0,
+        standingTokens: standing[arm],
+        injectedMeanTokens: arm === 'delivery' ? injMean : 0,
+        injectedP90Tokens: arm === 'delivery' ? injP90 : 0,
+    }));
+}
+
+function ratio(hits: number, total: number): string {
+    return total === 0 ? '—' : (hits / total).toFixed(3);
+}
+
+function signed(n: number): string {
+    return n > 0 ? `+${n}` : String(n);
+}
+
+/**
+ * Render the report. The pairwise deltas are printed, not left to the reader:
+ * "measured against one another" is a comparison, and a table of three
+ * independent rows is three measurements rather than one comparison.
+ */
+export function renderArmExperiment(rows: readonly ArmRow[], capBytes: number): string[] {
+    const by = new Map<ArmName, ArmRow>(rows.map((r) => [r.arm, r]));
+    const eager = by.get('eager-all') as ArmRow;
+    const thin = by.get('thin') as ArmRow;
+    const delivery = by.get('delivery') as ArmRow;
+
+    const out: string[] = [];
+    const push = (s: string): void => {
+        out.push(s);
+    };
+
+    push('── three-arm delivery experiment (road-to-governed-harness-evolution 6.1) ──');
+    push(
+        `arms = LeanProjectionMode (_lib/lean_projection_mode.ts:19) · ` +
+            `one corpus · one matcher (_lib/router_match.ts) · cap ${capBytes} B`,
+    );
+    push('');
+    push('  arm          delivery(pos)    false-context(near-miss)   standing tok   injected tok/prompt');
+    for (const r of rows) {
+        const inj =
+            r.arm === 'delivery'
+                ? `mean ${r.injectedMeanTokens} p90 ${r.injectedP90Tokens}`
+                : '0';
+        push(
+            `  ${r.arm.padEnd(12)} ${ratio(r.delivered, r.positives)} (${r.delivered}/${r.positives})`.padEnd(
+                33,
+            ) +
+                `${ratio(r.falseContext, r.nearMisses)} (${r.falseContext}/${r.nearMisses})`.padEnd(27) +
+                `${r.standingTokens}`.padEnd(15) +
+                inj,
+        );
+    }
+    push('');
+    push('── pairwise deltas ──');
+    push(
+        `  thin      vs eager-all: delivery ${signed(thin.delivered - eager.delivered)} positives · ` +
+            `standing ${signed(thin.standingTokens - eager.standingTokens)} tok`,
+    );
+    push(
+        `  delivery  vs eager-all: delivery ${signed(delivery.delivered - eager.delivered)} positives · ` +
+            `standing ${signed(delivery.standingTokens - eager.standingTokens)} tok`,
+    );
+    push(
+        `  delivery  vs thin:      delivery ${signed(delivery.delivered - thin.delivered)} positives · ` +
+            `standing ${signed(delivery.standingTokens - thin.standingTokens)} tok`,
+    );
+    push(`  cap-dropped under delivery: ${delivery.capDropped} of ${delivery.positives} positives`);
+    push('');
+    push(
+        'This measures WHICH BODIES REACH CONTEXT and what they cost. It does not',
+    );
+    push(
+        'measure whether a delivered body behaves like a standing one — that instrument',
+    );
+    push('is closed by ADR-202 and this run does not reopen it.');
+    return out;
+}