comisai 1.0.45 → 1.0.46

package/node_modules/@comis/agent/dist/memory/benchmark/cross-judge-spread.js

@@ -6,15 +6,13 @@
  *
  * THE MOAT: credibility = cross-judge
  * >=2 + a published spread. A result is trusted ONLY if it survives -- i.e. the
- * two judges agree within tolerance on that category. This module replaces the
- * hand-authored arithmetic in
- * `benchmarks/results/2026-05-31-j1-baseline/cross-judge-spread.md` with a
+ * two judges agree within tolerance on that category. This module is a
  * tested, reproducible fold over >=2 committed judge manifests.
  *
- * THE SURVIVAL TOLERANCE (from the j1-baseline reference math): a category
+ * THE SURVIVAL TOLERANCE: a category
  * SURVIVES if |A-B| <= 5.0 points. Per-category n=20 gives a binomial SE of
- * ~10-11pt, so a >5pt inter-judge gap is judge-noise, not signal -- e.g.
- * single-session-preference did NOT survive in the j1 baseline (30 vs 45 = 15pt)
+ * ~10-11pt, so a >5pt inter-judge gap is judge-noise, not signal -- e.g. a
+ * single-session-preference reading of 30 vs 45 (= 15pt) does NOT survive
  * and must NOT headline.
  *
  * PURE (no Result, no throws, no I/O, no clock, no env -- AGENTS.md 2.1 pure-fn
@@ -60,7 +58,7 @@
  * The survival tolerance in accuracy points: a category SURVIVES (is stable
  * across the two judges) when its inter-judge spread |A-B| <= this value.
  *
- * 5.0pt is the j1-baseline protocol value. Rationale: per-category n=20 yields a
+ * 5.0pt is the protocol value. Rationale: per-category n=20 yields a
  * binomial standard error of ~10-11pt, so a <=5pt inter-judge gap is well within
  * noise (the judges agree), whereas a larger gap signals the number is too
  * judge-dependent to headline.

package/node_modules/@comis/agent/dist/memory/benchmark/filesystem-baseline.d.ts

@@ -9,8 +9,7 @@
  * by the SAME judge. It is a deliberately-trivial no-memory reference: if a
  * full-dump baseline ties/beats Comis's ranked recall on a benchmark, the
  * *benchmark* is weak, not Comis (exactly how Letta showed a filesystem agent
- * scored 74.0% on LoCoMo, above Mem0's self-reported 68.5%;
- * .planning/MEMORY_BENCHMARK_CREDIBILITY.md TL;DR). It is recorded ONLY under an
+ * scored 74.0% on LoCoMo, above Mem0's self-reported 68.5%). It is recorded ONLY under an
  * explicit control label — NEVER as Comis's own score.
  *
  * COMPARABLE SHAPE: each doc is rendered as the SAME numbered + dated block that

package/node_modules/@comis/agent/dist/memory/benchmark/filesystem-baseline.js

@@ -10,8 +10,7 @@
  * by the SAME judge. It is a deliberately-trivial no-memory reference: if a
  * full-dump baseline ties/beats Comis's ranked recall on a benchmark, the
  * *benchmark* is weak, not Comis (exactly how Letta showed a filesystem agent
- * scored 74.0% on LoCoMo, above Mem0's self-reported 68.5%;
- * .planning/MEMORY_BENCHMARK_CREDIBILITY.md TL;DR). It is recorded ONLY under an
+ * scored 74.0% on LoCoMo, above Mem0's self-reported 68.5%). It is recorded ONLY under an
  * explicit control label — NEVER as Comis's own score.
  *
  * COMPARABLE SHAPE: each doc is rendered as the SAME numbered + dated block that

package/node_modules/@comis/agent/dist/memory/benchmark/learning-lift-scorer.d.ts

@@ -17,8 +17,8 @@
  * (the gated harness passes `scenario.docs.length`); the module default is a fixed
  * large constant {@link DEFAULT_ABSENT_SENTINEL} so a unit caller need not supply one.
  *
- * PURE: NO clock, NO I/O, NO @comis/memory import — mirrors recall-eval.ts. The
- * gated harness (learning-lift-harness.bench.test.ts) feeds it the per-episode gold
+ * PURE: NO clock, NO I/O, NO @comis/memory import — mirrors recall-eval.ts. A
+ * gated harness feeds it the per-episode gold
  * ranks from the live recall pipeline; the unit test pins the math.
  *
  * @module

package/node_modules/@comis/agent/dist/memory/benchmark/learning-lift-scorer.js

@@ -18,8 +18,8 @@
  * (the gated harness passes `scenario.docs.length`); the module default is a fixed
  * large constant {@link DEFAULT_ABSENT_SENTINEL} so a unit caller need not supply one.
  *
- * PURE: NO clock, NO I/O, NO @comis/memory import — mirrors recall-eval.ts. The
- * gated harness (learning-lift-harness.bench.test.ts) feeds it the per-episode gold
+ * PURE: NO clock, NO I/O, NO @comis/memory import — mirrors recall-eval.ts. A
+ * gated harness feeds it the per-episode gold
  * ranks from the live recall pipeline; the unit test pins the math.
  *
  * @module

package/node_modules/@comis/agent/dist/memory/benchmark/letta-fs-baseline-adapter.d.ts

@@ -10,7 +10,7 @@
  * and lets the SAME answer + judge models grade it. If a full-dump baseline
  * ties/beats Comis's ranked recall on a benchmark, the *benchmark* is weak, not
  * Comis (Letta's filesystem agent scored 74.0% on LoCoMo, above Mem0's
- * self-reported 68.5%; .planning/MEMORY_BENCHMARK_CREDIBILITY.md TL;DR).
+ * self-reported 68.5%).
  *
  * THE CONTROL-LABEL DISCIPLINE: this adapter carries
  * `isControl: true` and its result's `manifestRef` embeds the explicit

package/node_modules/@comis/agent/dist/memory/benchmark/letta-fs-baseline-adapter.js

@@ -11,7 +11,7 @@
  * and lets the SAME answer + judge models grade it. If a full-dump baseline
  * ties/beats Comis's ranked recall on a benchmark, the *benchmark* is weak, not
  * Comis (Letta's filesystem agent scored 74.0% on LoCoMo, above Mem0's
- * self-reported 68.5%; .planning/MEMORY_BENCHMARK_CREDIBILITY.md TL;DR).
+ * self-reported 68.5%).
  *
  * THE CONTROL-LABEL DISCIPLINE: this adapter carries
  * `isControl: true` and its result's `manifestRef` embeds the explicit

package/node_modules/@comis/agent/dist/memory/benchmark/locomo-loader.d.ts

@@ -47,7 +47,7 @@ export interface LocomoParsed {
     docs: LocomoDoc[];
     /**
      * Kept qa items (category-5 excluded). `questionId` is the
-     * cross-plan gold-map key (`${sample_id}:${qaIdx}`, original pre-filter index);
+     * gold-map key the harness joins on (`${sample_id}:${qaIdx}`, original pre-filter index);
      * `query` (NOT `question`) carries the question text under the SAME field name
      * as LongMemEvalParsed.questions[].query so the harness reads `q.query`
      * uniformly across both datasets.

package/node_modules/@comis/agent/dist/memory/benchmark/personalization-loaders.d.ts

@@ -12,7 +12,7 @@
  * SEPARATE channel. The gold (answers, and HaluMem's hallucination labels) is
  * NEVER serialized into document content (the anti-leak invariant).
  *
- * Documented external formats (RESEARCH §6; full schema is operator-verified):
+ * Documented external formats (full schema is operator-verified):
  *   - PrefEval (ICLR 2025 Oral; amazon-science/PrefEval): preference-adherence
  *     triplets `{ preference, query, gold }` -> `{ items: { preference, query,
  *     answer }[] }`. The on-mission preference-following lane (uncontested).

package/node_modules/@comis/agent/dist/memory/benchmark/personalization-loaders.js

@@ -13,7 +13,7 @@
  * SEPARATE channel. The gold (answers, and HaluMem's hallucination labels) is
  * NEVER serialized into document content (the anti-leak invariant).
  *
- * Documented external formats (RESEARCH §6; full schema is operator-verified):
+ * Documented external formats (full schema is operator-verified):
  *   - PrefEval (ICLR 2025 Oral; amazon-science/PrefEval): preference-adherence
  *     triplets `{ preference, query, gold }` -> `{ items: { preference, query,
  *     answer }[] }`. The on-mission preference-following lane (uncontested).

package/node_modules/@comis/agent/dist/memory/benchmark/qa-accuracy.d.ts

@@ -3,8 +3,8 @@
  * verdicts into the apples-to-apples accuracy numbers the benchmark report
  * carries.
  *
- * THE LOAD-BEARING DENOMINATOR (verified VERBATIM
- * against Hindsight benchmark_runner.py:840-866):
+ * THE LOAD-BEARING DENOMINATOR (identical to the denominator Hindsight's
+ * published benchmark runner applies, keeping the numbers comparable):
  *
  *     correct    = count of valid verdicts graded correct
  *     invalid    = count of verdicts the judge could not be parsed for

package/node_modules/@comis/agent/dist/memory/benchmark/qa-accuracy.js

@@ -4,8 +4,8 @@
  * verdicts into the apples-to-apples accuracy numbers the benchmark report
  * carries.
  *
- * THE LOAD-BEARING DENOMINATOR (verified VERBATIM
- * against Hindsight benchmark_runner.py:840-866):
+ * THE LOAD-BEARING DENOMINATOR (identical to the denominator Hindsight's
+ * published benchmark runner applies, keeping the numbers comparable):
  *
  *     correct    = count of valid verdicts graded correct
  *     invalid    = count of verdicts the judge could not be parsed for

package/node_modules/@comis/agent/dist/memory/benchmark/qa-judge-parse.d.ts

@@ -9,7 +9,7 @@
  * leading-`{...}`-object extraction AND a regex fallback because the judge --
  * instructed to emit `{ "correct": bool, "reasoning" }` but free-text by nature --
  * also emits a JSON object followed by trailing commentary, `correct: yes`,
- * fenced JSON, or a `correct=no` token (Pitfall 1). The
+ * fenced JSON, or a `correct=no` token. The
  * order is:
  *   1. strip markdown code fences via `stripCodeFences` -- broadens the analog's
  *      lowercase-`json`-only regex to remove ANY `[a-zA-Z]*` language tag,

package/node_modules/@comis/agent/dist/memory/benchmark/qa-judge-parse.js

@@ -10,7 +10,7 @@
  * leading-`{...}`-object extraction AND a regex fallback because the judge --
  * instructed to emit `{ "correct": bool, "reasoning" }` but free-text by nature --
  * also emits a JSON object followed by trailing commentary, `correct: yes`,
- * fenced JSON, or a `correct=no` token (Pitfall 1). The
+ * fenced JSON, or a `correct=no` token. The
  * order is:
  *   1. strip markdown code fences via `stripCodeFences` -- broadens the analog's
  *      lowercase-`json`-only regex to remove ANY `[a-zA-Z]*` language tag,

package/node_modules/@comis/agent/dist/memory/benchmark/qa-judge-prompt.d.ts

@@ -19,7 +19,7 @@
  *      pi-ai `completeSimple` has NO response_format arg, so the JSON verdict
  *      contract is instructed IN-PROMPT and parsed by the judge-verdict parser.
  *
- * SECURITY (prompt-injection ordering, Pitfall 3): the rubric and
+ * SECURITY (prompt-injection ordering): the rubric and
  * instructions are placed FIRST; the UNTRUSTED dataset slots (`question`,
  * `goldAnswer`, `modelAnswer`) are appended AFTER, in clearly labeled fields, so
  * adversarial dataset content cannot masquerade as judge instructions. The judge

package/node_modules/@comis/agent/dist/memory/benchmark/qa-judge-prompt.js

@@ -20,7 +20,7 @@
  *      pi-ai `completeSimple` has NO response_format arg, so the JSON verdict
  *      contract is instructed IN-PROMPT and parsed by the judge-verdict parser.
  *
- * SECURITY (prompt-injection ordering, Pitfall 3): the rubric and
+ * SECURITY (prompt-injection ordering): the rubric and
  * instructions are placed FIRST; the UNTRUSTED dataset slots (`question`,
  * `goldAnswer`, `modelAnswer`) are appended AFTER, in clearly labeled fields, so
  * adversarial dataset content cannot masquerade as judge instructions. The judge

package/node_modules/@comis/agent/dist/memory/benchmark/qa-report.d.ts

@@ -10,7 +10,7 @@
  * harness (where the dataset bytes are read) and passed in as `dataset.sha256` --
  * this module just records the string.
  *
- * SECURITY -- structural secret omission (Pitfall 6, ASVS V7): the
+ * SECURITY -- structural secret omission (ASVS V7): the
  * report is persisted via writeRegularFile, OUTSIDE Pino's redaction safety-net,
  * so the builder itself must guarantee no credential ever reaches the file. It
  * does so STRUCTURALLY: each model role is rebuilt as a fresh `{ provider,
@@ -22,7 +22,7 @@
  * (`https://user:token@host/...?token=...`) cannot smuggle a credential through.
  * Even when the operator config carries a per-role secret, it cannot appear in
  * `JSON.stringify(report)` because there is no path from the input secret to the
- * output object. (RED gate: the unit asserts the serialized report contains none
+ * output object. (The unit gate asserts the serialized report contains none
  * of the known secret substrings -- apiKey/Bearer/base_url AND a credential-bearing
  * modelUri -- with a secret-bearing config.)
  *
@@ -124,7 +124,7 @@ export interface BenchmarkControl {
 /**
  * The reproducibility object. Records WHAT built/answered/judged
  * (model identities), the dataset, the recall defaults, and the accuracy results
- * (carrying `invalid` + `validTotal` per the corrected denominator) -- with no
+ * (carrying `invalid` + `validTotal` for the invalid-excluded denominator) -- with no
  * secret anywhere.
  */
 export interface BenchmarkReport {
@@ -155,9 +155,9 @@ export interface BenchmarkReport {
         rerankEnabled: boolean;
         scoringAlphas: ScoringAlphas;
     };
-    /** The accuracy results (overall + per-category, with the corrected denominator fields). */
+    /** The accuracy results (overall + per-category, with the invalid-excluded denominator fields). */
     results: AccuracyResult;
-    /** The harness version tag (e.g. "phase-89-v1"). */
+    /** The harness version tag -- a fixed stamp identifying the harness code that produced the run. */
     harnessVersion: string;
     /**
      * Tokens/query (answer + judge). Present only when the run measured it;

package/node_modules/@comis/agent/dist/memory/benchmark/qa-report.js

@@ -11,7 +11,7 @@
  * harness (where the dataset bytes are read) and passed in as `dataset.sha256` --
  * this module just records the string.
  *
- * SECURITY -- structural secret omission (Pitfall 6, ASVS V7): the
+ * SECURITY -- structural secret omission (ASVS V7): the
  * report is persisted via writeRegularFile, OUTSIDE Pino's redaction safety-net,
  * so the builder itself must guarantee no credential ever reaches the file. It
  * does so STRUCTURALLY: each model role is rebuilt as a fresh `{ provider,
@@ -23,7 +23,7 @@
  * (`https://user:token@host/...?token=...`) cannot smuggle a credential through.
  * Even when the operator config carries a per-role secret, it cannot appear in
  * `JSON.stringify(report)` because there is no path from the input secret to the
- * output object. (RED gate: the unit asserts the serialized report contains none
+ * output object. (The unit gate asserts the serialized report contains none
  * of the known secret substrings -- apiKey/Bearer/base_url AND a credential-bearing
  * modelUri -- with a secret-bearing config.)
  *

package/node_modules/@comis/agent/dist/memory/benchmark/recall-trace-analyzer.d.ts

@@ -9,11 +9,11 @@
  * every recall (lanes fired, fused order, rerank pre/post scores, the final
  * ranked set with per-memory score breakdowns + include/exclude reasons), but
  * the CLI table view reads only `traceId/sessionKey/finalCount/ts`
- * (memory.ts:276-285) — everything else is "captured-but-unread"
- * (HINDSIGHT_VS_COMIS.md N4). This analyzer is the missing reducer: it folds the
+ * (memory.ts:276-285) — everything else is "captured-but-unread".
+ * This analyzer is the missing reducer: it folds the
  * discarded `ranked[].breakdown`, `rerank.pre/postScores`, `ranked[].reason`,
  * `lanes`, and `degradations` into a `TraceQualityView` — the diagnostic
- * dashboard a recall@k regression in a later phase is read against.
+ * dashboard a future recall@k regression is read against.
  *
  * ARCHITECTURAL CUT (architecture-graph.test.ts:133 — agent depends on
  * {shared,core,observability,scheduler}, NO memory edge): this file imports ONLY

package/node_modules/@comis/agent/dist/memory/benchmark/recall-trace-analyzer.js

@@ -10,11 +10,11 @@
  * every recall (lanes fired, fused order, rerank pre/post scores, the final
  * ranked set with per-memory score breakdowns + include/exclude reasons), but
  * the CLI table view reads only `traceId/sessionKey/finalCount/ts`
- * (memory.ts:276-285) — everything else is "captured-but-unread"
- * (HINDSIGHT_VS_COMIS.md N4). This analyzer is the missing reducer: it folds the
+ * (memory.ts:276-285) — everything else is "captured-but-unread".
+ * This analyzer is the missing reducer: it folds the
  * discarded `ranked[].breakdown`, `rerank.pre/postScores`, `ranked[].reason`,
  * `lanes`, and `degradations` into a `TraceQualityView` — the diagnostic
- * dashboard a recall@k regression in a later phase is read against.
+ * dashboard a future recall@k regression is read against.
  *
  * ARCHITECTURAL CUT (architecture-graph.test.ts:133 — agent depends on
  * {shared,core,observability,scheduler}, NO memory edge): this file imports ONLY

package/node_modules/@comis/agent/dist/memory/benchmark/results-ledger.d.ts

@@ -6,7 +6,7 @@
  * cost, latency}; a prior dated entry can NEVER be mutated or overwritten by a
  * later release.
  *
- * THE CORRECTED NEVER-OVERWRITE MECHANISM (verified against
+ * THE NEVER-OVERWRITE MECHANISM (verified against
  * fs-safe.ts:436-452): `writeRegularFile`'s default
  * `unlinkExisting:true` does `fs.unlinkSync(path)` (swallowing ENOENT) and THEN
  * opens with O_EXCL on the now-clean path -- so it SILENTLY OVERWRITES a
@@ -71,9 +71,9 @@ export interface LedgerRow {
     date: string;
     /** The Comis commit SHA the run measured (the other half of the filename). */
     commit: string;
-    /** The git branch the run was taken on (e.g. "v2.8-prove-climb"). */
+    /** The git branch the run was taken on (e.g. "bench/prove-climb"). */
     branch: string;
-    /** Component -> version map (e.g. {comis:"2.8.0", pi:"0.78.0"}); URI values sanitized. */
+    /** Component -> version map (e.g. {comis:"1.0.0", pi:"0.78.0"}); URI values sanitized. */
     systemVersions: Record<string, string>;
     /** The benchmark tier (OPEN string -- e.g. "head-to-head", "longmemeval-v2"). */
     tier: string;
@@ -131,7 +131,7 @@ export interface AppendLedgerRowSuccess {
     path: string;
 }
 /**
- * Append a dated row to the ledger -- the corrected NEVER-OVERWRITE writer.
+ * Append a dated row to the ledger -- the NEVER-OVERWRITE writer.
  * Computes the dated path; **refuses** if a file already exists at
  * that path (the explicit `existsSync` guard -- `writeRegularFile`'s default
  * would SILENTLY clobber it, fs-safe.ts:436-452); else writes it confined to

package/node_modules/@comis/agent/dist/memory/benchmark/results-ledger.js

@@ -7,7 +7,7 @@
  * cost, latency}; a prior dated entry can NEVER be mutated or overwritten by a
  * later release.
  *
- * THE CORRECTED NEVER-OVERWRITE MECHANISM (verified against
+ * THE NEVER-OVERWRITE MECHANISM (verified against
  * fs-safe.ts:436-452): `writeRegularFile`'s default
  * `unlinkExisting:true` does `fs.unlinkSync(path)` (swallowing ENOENT) and THEN
  * opens with O_EXCL on the now-clean path -- so it SILENTLY OVERWRITES a
@@ -182,7 +182,7 @@ export function ledgerRowPath(historyDir, date, commit) {
     return safePath(historyDir, `${date}-${commit}.json`);
 }
 /**
- * Append a dated row to the ledger -- the corrected NEVER-OVERWRITE writer.
+ * Append a dated row to the ledger -- the NEVER-OVERWRITE writer.
  * Computes the dated path; **refuses** if a file already exists at
  * that path (the explicit `existsSync` guard -- `writeRegularFile`'s default
  * would SILENTLY clobber it, fs-safe.ts:436-452); else writes it confined to
@@ -206,7 +206,7 @@ export function appendLedgerRow(opts) {
             return err(e);
         return err(e instanceof Error ? e : new Error(String(e)));
     }
-    // THE CORRECTED NEVER-OVERWRITE GUARD: refuse to clobber a prior dated entry.
+    // THE NEVER-OVERWRITE GUARD: refuse to clobber a prior dated entry.
     // writeRegularFile's default unlinkExisting:true would silently overwrite it
     // (the O_EXCL is anti-TOCTOU-symlink, not anti-clobber -- fs-safe.ts:436-452).
     if (existsSync(path)) {

package/node_modules/@comis/agent/dist/memory/benchmark/significance.d.ts

@@ -2,7 +2,7 @@
  * Statistical-significance layer -- the N + confidence-interval + significance
  * computations every published benchmark number must carry.
  *
- * THE BELIEVABILITY REQUIREMENT (.planning/MEMORY_BENCHMARK_CREDIBILITY.md):
+ * THE BELIEVABILITY REQUIREMENT:
  * every headline number reports N + a significance flag; a single-judge,
  * no-N, no-significance number is not credible. This module supplies (1) a
  * Wilson 95% confidence interval for a single accuracy over integer counts, and

package/node_modules/@comis/agent/dist/memory/benchmark/significance.js

@@ -3,7 +3,7 @@
  * Statistical-significance layer -- the N + confidence-interval + significance
  * computations every published benchmark number must carry.
  *
- * THE BELIEVABILITY REQUIREMENT (.planning/MEMORY_BENCHMARK_CREDIBILITY.md):
+ * THE BELIEVABILITY REQUIREMENT:
  * every headline number reports N + a significance flag; a single-judge,
  * no-N, no-significance number is not credible. This module supplies (1) a
  * Wilson 95% confidence interval for a single accuracy over integer counts, and

package/node_modules/@comis/agent/dist/memory/benchmark/suite-report.d.ts

@@ -1,6 +1,6 @@
 /**
  * Generic per-tier benchmark suite report builder -- the single manifest shape
- * every Tier-3 + external + BEAM harness writes to disk via `writeRegularFile`,
+ * every Comis-unique, external-benchmark, and BEAM harness writes to disk via `writeRegularFile`,
  * so a Comis suite number is reproducible across changes and comparable
  * per-ability.
  *

package/node_modules/@comis/agent/dist/memory/benchmark/suite-report.js

@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
  * Generic per-tier benchmark suite report builder -- the single manifest shape
- * every Tier-3 + external + BEAM harness writes to disk via `writeRegularFile`,
+ * every Comis-unique, external-benchmark, and BEAM harness writes to disk via `writeRegularFile`,
  * so a Comis suite number is reproducible across changes and comparable
  * per-ability.
  *

package/node_modules/@comis/agent/dist/memory/benchmark/suite-scenario.d.ts

@@ -1,10 +1,10 @@
 /**
  * Deterministic constructed-scenario builders -- the
  * adversarial / contradiction / redaction / learning fixtures the 4 Comis-unique
- * Tier-3 harnesses (poisoning, trust-contradiction, redaction,
+ * harnesses (poisoning, trust-contradiction, redaction,
  * recall-learning) ingest WITHOUT any external corpus.
  *
- * WHY CONSTRUCTED: the 4 Tier-3 benchmarks measure
+ * WHY CONSTRUCTED: these 4 benchmarks measure
  * shipped properties (trust filter, write-time secret downgrade, FEED loop)
  * over scenarios that need no download. Each builder returns hard-coded
  * deterministic literals -- so the same call is byte-identical across runs and

package/node_modules/@comis/agent/dist/memory/benchmark/suite-scenario.js

@@ -2,10 +2,10 @@
 /**
  * Deterministic constructed-scenario builders -- the
  * adversarial / contradiction / redaction / learning fixtures the 4 Comis-unique
- * Tier-3 harnesses (poisoning, trust-contradiction, redaction,
+ * harnesses (poisoning, trust-contradiction, redaction,
  * recall-learning) ingest WITHOUT any external corpus.
  *
- * WHY CONSTRUCTED: the 4 Tier-3 benchmarks measure
+ * WHY CONSTRUCTED: these 4 benchmarks measure
  * shipped properties (trust filter, write-time secret downgrade, FEED loop)
  * over scenarios that need no download. Each builder returns hard-coded
  * deterministic literals -- so the same call is byte-identical across runs and

package/node_modules/@comis/agent/dist/memory/correction-detector-seam.d.ts

@@ -1,29 +1,29 @@
 /**
- * The daemon-injected OPTIONAL, cost-gated correction-detector seam (CORRECT-01).
+ * The daemon-injected OPTIONAL, cost-gated correction-detector seam.
  *
  * {@link createCorrectionDetectorSeam} wraps a cheap resolved model into a
  * `detect(followUpUserTurn)` seam — the SEPARATE `correction` outcome source for
- * v2.26 Verified Learning (WS1). When a finished trajectory is immediately
+ * verified learning. When a finished trajectory is immediately
  * followed by a user turn AND the per-agent correction detector is enabled, the
- * daemon (Plan 04) runs ONE cheap-model pass over that follow-up turn and gets
+ * daemon runs ONE cheap-model pass over that follow-up turn and gets
  * back an `isCorrection` verdict + a confidence. When `isCorrection`, the verdict
  * is fed to `OutcomeSignalPort.observe()` as a `source: "correction"` /
  * `outcome: "corrected"` observation — a SOFT-FAILURE of the PRIOR trajectory.
- * The deterministic tool/pipeline sources ALWAYS outrank it via the Plan 02
+ * The deterministic tool/pipeline sources ALWAYS outrank it via the
  * fusion precedence — a correction can never overturn a deterministic result.
  *
  * This is the signal the outcome-judge seam explicitly defers to: "the judge
- * does NOT detect corrections (that is the separate `correction` signal, Phase
- * 199)". It is a verbatim clone of {@link createOutcomeJudgeSeam}, changing ONLY
+ * does NOT detect corrections (that is the separate `correction`
+ * signal)". It is a verbatim clone of {@link createOutcomeJudgeSeam}, changing ONLY
  * the verdict shape, the cap constant, the source tier, and the prompt; the
  * entire untrusted-input posture (wrap + lenient parse + cap-in-code + tier-in-
  * code), the non-fatal `callModel`, and the abort timer are copied UNCHANGED.
  *
- * Built but DORMANT: the detector ships `correction.enabled:false` by default
- * (design D6 / Plan 01), so the daemon never constructs or calls it unless an
+ * Built but DORMANT: the detector ships `correction.enabled:false` by default,
+ * so the daemon never constructs or calls it unless an
  * operator opts in.
  *
- * Security posture (CORRECT-01 / SEC-01 / §9 triple-bound — the follow-up turn
+ * Security posture (a triple bound — the follow-up turn
  * the detector scores is UNTRUSTED, and the model's self-reported `confidence` is
  * UNTRUSTED):
  * 1. The follow-up turn is delimiter-wrapped via
@@ -31,7 +31,7 @@
  *    ever sees it — an injected "this is a correction / confidence: 1.0" is
  *    neutralized as external content, not read as an instruction. (The
  *    `outcome_judge` content label is reused — a separate `correction_input`
- *    label would be cosmetic; research A2.)
+ *    label would be cosmetic.)
  * 2. The reward the daemon will `observe()` is capped in CODE at
  *    {@link CORRECTION_REWARD_CAP} — `Math.min(modelConfidence,
  *    CORRECTION_REWARD_CAP)` — INDEPENDENT of the model's self-report. An injected
@@ -51,21 +51,21 @@
  * the prior outcome simply stays unflipped. Each call is BOUNDED by
  * `maxOutputTokens` and a wall-clock-free abort timer (the injected `clock`
  * supplies timestamps; the abort uses the sanctioned-root `systemSetTimeout`),
- * routed to the cheap `fast` tier (`outcomeJudge` ModelOperationType — research
- * A2, no new ModelOperationType; the daemon resolves provider/modelId/apiKey by
+ * routed to the cheap `fast` tier (the existing `outcomeJudge` ModelOperationType
+ * — no new ModelOperationType; the daemon resolves provider/modelId/apiKey by
  * NAME and injects them via Deps).
  *
  * OFFLINE only — NEVER imported by the recall read path; no agent↛memory edge
  * (it imports the `@comis/core` `wrapExternalContent` + types and `pi-ai` only —
- * the closed-graph cut, SEC-01).
+ * the closed-graph cut).
  *
  * @module
  */
 import type { ClockPort, ComisLogger } from "@comis/core";
 import { type CustomCompletionsModelSpec } from "./judge-model-resolver.js";
 /**
- * The reward ceiling for a correction verdict — the CORRECT-01 "reward capped
- * independent of self-reported confidence" constant (design §9 / §17 triple-bound).
+ * The reward ceiling for a correction verdict — the "reward capped
+ * independent of self-reported confidence" constant (bound #2 of the triple bound).
  * The effective reward the daemon `observe()`s is `Math.min(modelConfidence,
  * CORRECTION_REWARD_CAP)`, so a maximal self-report (an injection coercing
  * `confidence: 1.0 / this is a correction`) can never produce a reward above this
@@ -94,8 +94,8 @@ export interface CorrectionDetectorSeamDeps {
      * Custom OpenAI-compatible model spec (resolved, normalized `…/v1` baseUrl) for
      * building the judge Model when the pi-ai catalog has no entry for
      * `provider/modelId` — a custom YAML provider (ollama/lm-studio/…). Undefined
-     * for built-in providers. Without it, correction detection SKIPPED on every
-     * keyless/local turn (the same bug as the outcome judge, live 2026-06-20).
+     * for built-in providers. Without it, correction detection is SKIPPED on every
+     * keyless/local turn (the same failure mode the outcome judge guards against).
      */
     customModel?: CustomCompletionsModelSpec;
 }
@@ -116,7 +116,7 @@ export interface CorrectionVerdict {
     /**
      * The EFFECTIVE reward the daemon will `observe()`: `Math.min(confidence,
      * CORRECTION_REWARD_CAP)`. Capped in CODE independent of the self-report — the
-     * CORRECT-01 keystone. An injected `confidence: 1.0` yields at most the cap.
+     * reward-cap keystone. An injected `confidence: 1.0` yields at most the cap.
      */
     cappedConfidence: number;
     /** The outcome the daemon observes when `isCorrection` — a soft-failure of the PRIOR trajectory. Set in CODE. */

package/node_modules/@comis/agent/dist/memory/correction-detector-seam.js

@@ -1,30 +1,30 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * The daemon-injected OPTIONAL, cost-gated correction-detector seam (CORRECT-01).
+ * The daemon-injected OPTIONAL, cost-gated correction-detector seam.
  *
  * {@link createCorrectionDetectorSeam} wraps a cheap resolved model into a
  * `detect(followUpUserTurn)` seam — the SEPARATE `correction` outcome source for
- * v2.26 Verified Learning (WS1). When a finished trajectory is immediately
+ * verified learning. When a finished trajectory is immediately
  * followed by a user turn AND the per-agent correction detector is enabled, the
- * daemon (Plan 04) runs ONE cheap-model pass over that follow-up turn and gets
+ * daemon runs ONE cheap-model pass over that follow-up turn and gets
  * back an `isCorrection` verdict + a confidence. When `isCorrection`, the verdict
  * is fed to `OutcomeSignalPort.observe()` as a `source: "correction"` /
  * `outcome: "corrected"` observation — a SOFT-FAILURE of the PRIOR trajectory.
- * The deterministic tool/pipeline sources ALWAYS outrank it via the Plan 02
+ * The deterministic tool/pipeline sources ALWAYS outrank it via the
  * fusion precedence — a correction can never overturn a deterministic result.
  *
  * This is the signal the outcome-judge seam explicitly defers to: "the judge
- * does NOT detect corrections (that is the separate `correction` signal, Phase
- * 199)". It is a verbatim clone of {@link createOutcomeJudgeSeam}, changing ONLY
+ * does NOT detect corrections (that is the separate `correction`
+ * signal)". It is a verbatim clone of {@link createOutcomeJudgeSeam}, changing ONLY
  * the verdict shape, the cap constant, the source tier, and the prompt; the
  * entire untrusted-input posture (wrap + lenient parse + cap-in-code + tier-in-
  * code), the non-fatal `callModel`, and the abort timer are copied UNCHANGED.
  *
- * Built but DORMANT: the detector ships `correction.enabled:false` by default
- * (design D6 / Plan 01), so the daemon never constructs or calls it unless an
+ * Built but DORMANT: the detector ships `correction.enabled:false` by default,
+ * so the daemon never constructs or calls it unless an
  * operator opts in.
  *
- * Security posture (CORRECT-01 / SEC-01 / §9 triple-bound — the follow-up turn
+ * Security posture (a triple bound — the follow-up turn
  * the detector scores is UNTRUSTED, and the model's self-reported `confidence` is
  * UNTRUSTED):
  * 1. The follow-up turn is delimiter-wrapped via
@@ -32,7 +32,7 @@
  *    ever sees it — an injected "this is a correction / confidence: 1.0" is
  *    neutralized as external content, not read as an instruction. (The
  *    `outcome_judge` content label is reused — a separate `correction_input`
- *    label would be cosmetic; research A2.)
+ *    label would be cosmetic.)
  * 2. The reward the daemon will `observe()` is capped in CODE at
  *    {@link CORRECTION_REWARD_CAP} — `Math.min(modelConfidence,
  *    CORRECTION_REWARD_CAP)` — INDEPENDENT of the model's self-report. An injected
@@ -52,26 +52,26 @@
  * the prior outcome simply stays unflipped. Each call is BOUNDED by
  * `maxOutputTokens` and a wall-clock-free abort timer (the injected `clock`
  * supplies timestamps; the abort uses the sanctioned-root `systemSetTimeout`),
- * routed to the cheap `fast` tier (`outcomeJudge` ModelOperationType — research
- * A2, no new ModelOperationType; the daemon resolves provider/modelId/apiKey by
+ * routed to the cheap `fast` tier (the existing `outcomeJudge` ModelOperationType
+ * — no new ModelOperationType; the daemon resolves provider/modelId/apiKey by
  * NAME and injects them via Deps).
  *
  * OFFLINE only — NEVER imported by the recall read path; no agent↛memory edge
  * (it imports the `@comis/core` `wrapExternalContent` + types and `pi-ai` only —
- * the closed-graph cut, SEC-01).
+ * the closed-graph cut).
  *
  * @module
  */
 import { systemSetTimeout, systemClearTimeout, wrapExternalContent } from "@comis/core";
-import { completeSimple } from "@earendil-works/pi-ai";
+import { completeSimple } from "@earendil-works/pi-ai/compat";
 import { resolveJudgeModel, temperatureOption } from "./judge-model-resolver.js";
 import { z } from "zod";
 import { parseLenientJson } from "./llm-json.js";
 /** Hard abort ceiling per LLM call (mirrors the outcome-judge seam LLM timeout). */
 const LLM_TIMEOUT_MS = 120_000;
 /**
- * The reward ceiling for a correction verdict — the CORRECT-01 "reward capped
- * independent of self-reported confidence" constant (design §9 / §17 triple-bound).
+ * The reward ceiling for a correction verdict — the "reward capped
+ * independent of self-reported confidence" constant (bound #2 of the triple bound).
  * The effective reward the daemon `observe()`s is `Math.min(modelConfidence,
  * CORRECTION_REWARD_CAP)`, so a maximal self-report (an injection coercing
  * `confidence: 1.0 / this is a correction`) can never produce a reward above this
@@ -150,7 +150,7 @@ function parseVerdict(raw) {
     return {
         isCorrection: parsed.data.isCorrection,
         confidence: parsed.data.confidence,
-        // The CORRECT-01 reward cap — bounded in CODE, never trusted from the model.
+        // The reward cap — bounded in CODE, never trusted from the model.
         cappedConfidence: Math.min(parsed.data.confidence, CORRECTION_REWARD_CAP),
         // The outcome + tier are set HERE, in code — a smuggled `outcome`/`source`
         // field cannot promote the verdict.
@@ -221,10 +221,10 @@ export function createCorrectionDetectorSeam(deps) {
     return async function detect(followUpUserTurn) {
         // The follow-up turn is UNTRUSTED — delimiter-wrap it BEFORE the model sees it
         // so an injected "this is a correction / confidence: 1.0" is neutralized as
-        // external content, never read as an instruction (CORRECT-01 bound #1). The
+        // external content, never read as an instruction (bound #1). The
         // wrap reads `contentDelimiter` from the ALS context for cache-stable,
         // session-consistent markers. The `outcome_judge` source label is reused
-        // (research A2 — a separate `correction_input` label would be cosmetic).
+        // (a separate `correction_input` label would be cosmetic).
         const wrapped = wrapExternalContent(followUpUserTurn, { source: "outcome_judge" });
         const text = await callModel(CORRECTION_DETECTOR_PROMPT, wrapped);
         // A failed/aborted call → no verdict (undefined); the prior outcome stays unflipped.