{
  "consolidations": [
    {
      "student": "s-mesh-1",
      "commission": "Where, exactly, in GLideN64's depth/clipping path can mesh improvement be inserted so that intersecting geometry stops cutting through itself — and what is the measured step-1 plan to do it? (Founder verdicts 2026-11-07 06:07/06:13; faces are the first proving ground.)",
      "method": "Insertion-point mapping by pinned-source recount: draft the map from the research line's citations, then re-read EVERY citation at its file/line in the pinned source tree before filing; a citation counts as evidence only after the recount (two of the research line's citations failed the first check and were corrected by it). The verified map (gSPVertex / gSPProcessVertex shadow-stash / gSPClipVertex / gSPTriangle polygon-clip seam / drawTriangles flat subdivision, all in the pinned GLideN64 tree) and the faces-first step-1 plan live in the student's filed workspace, STEP1-PLAN.md.",
      "evidence": {
        "commission": "/home/hatch/workspace/flowchi-school/work-agent/students/s-mesh-1/commission.json",
        "ledger": "/home/hatch/workspace/flowchi-school/work-agent/students/s-mesh-1/ledger.json",
        "evaluation": {
          "madeBetter": true,
          "learnedBetter": true,
          "measured": 5,
          "baseline": 0,
          "triesSequence": [
            2,
            1
          ],
          "graduate": true
        }
      },
      "consolidatedAt": "2026-11-07T12:34:02.397Z"
    }
  ],
  "entries": [
    {
      "method": "Insertion-point mapping by pinned-source recount: draft the map from the research line's citations, then re-read EVERY citation at its file/line in the pinned source tree before filing; a citation counts as evidence only after the recount (two of the research line's citations failed the first check and were corrected by it). The verified map (gSPVertex / gSPProcessVertex shadow-stash / gSPClipVertex / gSPTriangle polygon-clip seam / drawTriangles flat subdivision, all in the pinned GLideN64 tree) and the faces-first step-1 plan live in the student's filed workspace, STEP1-PLAN.md.",
      "provenance": "student s-mesh-1",
      "question": "Where, exactly, in GLideN64's depth/clipping path can mesh improvement be inserted so that intersecting geometry stops cutting through itself — and what is the measured step-1 plan to do it? (Founder verdicts 2026-11-07 06:07/06:13; faces are the first proving ground.)"
    },
    {
      "method": "Digesting an external architecture articulation in three passes. (a) Convergence pass: name what restates what is already live and refuse to count it as improvement (in Matrix 2.0: the versioned method registry is method memory; LEARN and EVOLVE are the record phase and the student/proposal loops; the lineage layer and evidence-gated promotion already stand). (b) Delta pass: extract only the portable remainder and scope each delta to the smallest form that survives the constitution - a routing policy inside DIRECT, never around the trunk; evidence expiration as assumption-naming plus a named invalidation list; self-benchmarking on the founder's own metrics, the supervision ratio included. (c) Refusal pass: strike what contradicts ratified law, with the law cited, and file the surviving deltas as ONE Gate proposal instead of enacting them. Recorded from task 0005-skill-tree-digest; machinery unchanged; proposal matrix-2-deltas pending at the Gate.",
      "provenance": "task 0005-skill-tree-digest - SI core deliberation over OMNI SKILL MATRIX 2.0 under the parent verdict of 2026-11-07 08:01 EST, accepted by the founder 08:03",
      "question": "Which parts of the external OMNI SKILL MATRIX 2.0 articulation would make Omni faster, sharper, and more accurate - and which parts must be refused to keep the constitution intact?",
      "kind": "design-knowledge",
      "digested": {
        "source": "/home/hatch/workspace/user/files/OMNI_SKILL_MATRIX_2_0_Architecture_48_l33j.md",
        "verdictSource": "/home/hatch/workspace/your_files/skill-tree-feedback-for-diagram-maker.md",
        "founderMatrix": "work-agent/deliverables/omni-skill-matrix.mp4",
        "proposal": "matrix-2-deltas",
        "adoptedWard": [
          {
            "delta": "routing policy in DIRECT",
            "why": "DIRECT routes informally today; a written policy (fit, prerequisites, cost, evidence quality, risk; hard constraints before speed) plus a route trace in every receipt attacks the serial-deliberation speed tax without touching verify or record (feedback memo, SLOW item 1)."
          },
          {
            "delta": "evidence expiration",
            "why": "Receipts in this house never expire; stale 'already verified' evidence is the exact shape of the v17 failure - a frame fix believed because a pack-bytes fix was once true. Assumption-naming plus an invalidation list is an accuracy mechanism nothing current provides."
          },
          {
            "delta": "self-benchmarking against the incumbent",
            "why": "The articulation's own step 7 demands the challenger beat the incumbent on filed numbers - completion rate, founder actions (the supervision ratio), latency, regression rate. A structural change that cannot beat the incumbent should not land."
          }
        ],
        "refused": [
          {
            "item": "optional-stage adaptive loop replacing the phase-law trunk",
            "why": "The trunk is the kernel's state law; optional stages dissolve the enforceable sequence and the fresh-eyes verify the receipts seal through."
          },
          {
            "item": "shared memory bus",
            "why": "Capsule discipline scopes working state per project; only evidence is global, and it is global because it is append-only and hash-chained. The shared-bus shape was already refused in the N8 Link verdict."
          },
          {
            "item": "EVOLVE-style self-promotion",
            "why": "Promotion is the Gate's act; the proposals module refuses the agent's own id mechanically. Thresholds the agent sets for itself are self-approval with extra steps."
          },
          {
            "item": "any variant dropping the kernel, the Gate, the students, or the less-hands invariant",
            "why": "Those are the constitution, not features; a version without them is a different agent wearing the name."
          }
        ],
        "dissent": [
          {
            "claim": "evidence-expiration",
            "by": "measurement",
            "objection": "Expiration needs a clock the ledger does not keep. Today's receipts carry no assumption sets, so the invalidation list is enforceable only from the day methods start naming their assumptions; it cannot be applied retroactively to the existing chain, and the delta should never be quoted as if the old evidence were covered. Recorded as dissent so the limit survives with the adoption, if the Gate adopts."
          },
          {
            "claim": "tree-from-data",
            "by": "curator",
            "objection": "A generator stops garbling, not overclaiming. The tree's states are still assigned by a builder reading records at build time; what keeps them honest is that every state cites evidence the verify phase re-reads, and that the suite re-checks the citations. The citations are the discipline - keep them mandatory, or this tree becomes the next plausible picture."
          }
        ]
      },
      "at": "2026-11-07T13:12:01.392Z"
    },
    {
      "method": "Examining an external improvement package against a live tool. (1) Reproduce the defect against your own original before reading the patch: the substring-trust defect class — a parser that greps narrative text for verdict-shaped substrings will believe a PASS printed anywhere, including before a crash and inside another program's log line; the cure is anchored, complete-line evidence with uniqueness requirements, and the proof is the original convicting itself on the forged transcript while the hardened parser refuses it. (2) Fail closed over destructive defaults: a runner that opens by deleting whatever sits at its work path makes every mislabeled run a small disaster; refusal is the correct default posture, and reclamation — where the workflow genuinely needs it — must be earned by a marker the tool itself wrote, never assumed from a path name. (3) Evidence sidecars as a receipt pattern: a receipt that travels with the SHA-256 of its transcript, its runner and itself lets any later reader detect tampering with the evidence a verdict rests on; name the limit in the same breath — hashes beside the files are tamper-evidence, not authenticity; authenticity wants a store or a signature, which is separate, queued work. (4) Classify before you attribute: name the failure's category from its evidence before writing the sentence that explains it, and let 'cause not established' stand as a first-class verdict — the original door attributed every failure to the GL path because the GL wall was the failure it knew; attribution without evidence is how a wall becomes an alibi. Recorded from task 0006-bench-improvements-examination; the examined code is live in tools/bench-session.cjs, tools/diagnostics.cjs, tools/bench-session.test.cjs and the hardened patched bench runner.",
      "provenance": "task 0006-bench-improvements-examination — SI core deliberation over ChatGPT's two bench-challenge packages (~/workspace/challenge/returned/, ~/workspace/challenge/returned2/), commissioned by the founder through the bench challenge 2026-11-07; examined and adopted on the founder's directive to deeply examine the improvements, learn from them, and integrate (2026-11-07 08:22 EST)",
      "question": "Which of the external bench improvements survive verification against the live door, what general lessons do they carry, and what — if anything — must be adapted or refused, with reasons?",
      "kind": "examination-knowledge",
      "digested": {
        "source": "~/workspace/challenge/returned/ + ~/workspace/challenge/returned2/",
        "challenge": "~/workspace/challenge/CHALLENGE.md",
        "verdictSource": "~/MEMORY.md (parent verification 2026-11-07) — read first, then re-derived independently in this tree; the parent's word was not taken as evidence",
        "task": "0006-bench-improvements-examination",
        "adopted": [
          {
            "piece": "round-1 parser hardening (anchored full-line matching; PASS requires exactly one result line, one exit-0 line, stages complete, no contradictory evidence)",
            "why": "Her 12 tests pass in this tree; the forged PASS-then-abort transcript returns FAIL under the live service while the preserved original returns PASS on the same bytes — the hole is demonstrated, not described."
          },
          {
            "piece": "evidence-bound attribution + pack staged agreement (verifyPackLine)",
            "why": "A non-GL failure no longer inherits the GL sentence; staged=true now requires the transcript's src/staged/digest to agree with the service's own record (probe: filed line verifies true, digest-swapped line verifies false)."
          },
          {
            "piece": "diagnostics module (seven evidence-linked fault categories; recovery recommendation-only)",
            "why": "Classification runs on the filed transcript (graphics-implementation-unavailable, high confidence, the GL line as its evidence) and the deep tests prove the recovery plan is never auto-executed (attempted: false, every step automatic: false)."
          },
          {
            "piece": "bounded execution window (exec-sec 1..600, refused at policy)",
            "why": "A bench run's window is now a validated input, not a string passed through to a timeout; the refusal rides the existing validation stack order."
          },
          {
            "piece": "per-run SHA-256 evidence sidecar (.evidence.json beside each live receipt)",
            "why": "Tamper-evidence for the exact artifacts a verdict rests on (transcript, receipt, runner), with its limit stated inside the manifest itself; tamper detection proven by the incorporated deep check."
          },
          {
            "piece": "gl-preflight.sh (read-only guest GLES/EGL discovery)",
            "why": "It does one honest thing — inventories candidate libraries — and prints its own limit on its output; ran clean against a probe rootfs in this examination."
          }
        ],
        "adapted": [
          {
            "piece": "runner fail-closed view handling",
            "from": "refuse any existing view, unconditionally",
            "to": "refuse any existing view except one carrying the door's own completion marker (.omni-run-complete, written at the end of a completed run), which the runner reclaims",
            "why": "The two filed live sessions ran back-to-back under one label because the view is reclaimed between runs; unconditional refusal deadlocks the loop and re-creates manual deletion — the destructive habit in a new place. Probes: unmarked view refused (exit 2, view untouched); marked view reclaimed (run proceeds)."
          }
        ],
        "refused": [],
        "refusalNote": "Nothing in the two packages was refused outright: every piece survived verification in this tree. The reservations ride with the adoption as one recorded adaptation and three preserved dissents.",
        "decisions": {
          "runnerFailClosed": "adapt — adopt the fail-closed view rule with completion-marker reclamation: the runner reclaims only a view carrying its own .omni-run-complete marker and refuses every other existing view; the rm -rf default is removed"
        },
        "dissent": [
          {
            "claim": "anchored-parsing",
            "by": "curator",
            "objection": "Anchoring convicts the forged PASS; it also narrows what the parser can hear. Every pattern is now a whole-line contract with the runner's exact wording — a future rewording of a transcript line (a core upgrade changing a stage line, say) makes the parser fail closed: verdicts degrade toward FAIL and 'cause not established', never toward a false PASS. That is the safe direction, and the coupling is still real: the runner's transcript lines are now part of this tool's interface, and a wording change on either side is a breaking change on both. Recorded so the coupling is chosen, not discovered."
          },
          {
            "claim": "diagnostics-sidecar",
            "by": "measurement",
            "objection": "The sidecar is tamper-evidence, not authenticity, and the distinction must survive the adoption. Its hashes travel beside the files they describe: a writer who can rewrite a receipt can rewrite its sidecar in the same motion, and the manifest itself says so ('local-untrusted-hashes; no authenticity signature'). What it genuinely buys is detection of after-the-fact alteration against a retained digest — worth having, and never to be quoted as proof a receipt was honest when written. The cure for the stronger claim is the content-addressed store already queued on the environment backlog, not this patch."
          },
          {
            "claim": "verbatim-adoption",
            "by": "build",
            "objection": "Verbatim adoption is right for this code and wrong as a blanket rule. Her files arrived carrying no record of where they came from; the integration adds a provenance header to the service and files this examination precisely so the tree never holds code whose reasons live outside it. If the next external patch is adopted verbatim without its own examination, the tree gains machinery nobody here can defend line by line. The dissent is against the habit, not this adoption: verbatim is earned per package, by examination, or not at all."
          }
        ]
      },
      "at": "2026-11-07T13:36:33.497Z"
    }
  ],
  "note": "Proven methods live here - pulled from the fused charters and the program library at direct time, and joined by graduated students' proven methods (with provenance). The memory grows; the roster never does.",
  "schema": "work-agent.method-memory.v1",
  "seededFrom": [
    "workshop/fusion.json",
    "workshop/roles/*.md",
    "~/workspace/crew/fused/ charters",
    "workshop/library/ (12 programs, LIBRARY-RECEIPT.md)"
  ],
  "designDigests": [
    {
      "task": "0005-skill-tree-digest",
      "proposal": "matrix-2-deltas",
      "source": "/home/hatch/workspace/user/files/OMNI_SKILL_MATRIX_2_0_Architecture_48_l33j.md",
      "kept": [
        "routing-policy",
        "evidence-expiration",
        "self-benchmarking",
        "tree-from-data"
      ],
      "corrected": [
        "claim-binding"
      ],
      "struck": [
        "trunk-replacement",
        "shared-memory",
        "evolve-promotion"
      ],
      "dissent": [
        {
          "claim": "evidence-expiration",
          "by": "measurement",
          "objection": "Expiration needs a clock the ledger does not keep. Today's receipts carry no assumption sets, so the invalidation list is enforceable only from the day methods start naming their assumptions; it cannot be applied retroactively to the existing chain, and the delta should never be quoted as if the old evidence were covered. Recorded as dissent so the limit survives with the adoption, if the Gate adopts."
        },
        {
          "claim": "tree-from-data",
          "by": "curator",
          "objection": "A generator stops garbling, not overclaiming. The tree's states are still assigned by a builder reading records at build time; what keeps them honest is that every state cites evidence the verify phase re-reads, and that the suite re-checks the citations. The citations are the discipline - keep them mandatory, or this tree becomes the next plausible picture."
        }
      ],
      "at": "2026-11-07T13:12:01.392Z"
    },
    {
      "task": "0006-bench-improvements-examination",
      "source": "~/workspace/challenge/returned/ + ~/workspace/challenge/returned2/",
      "kept": [
        "anchored-parsing",
        "evidence-bound-attribution",
        "diagnostics-sidecar",
        "gl-preflight",
        "verbatim-adoption",
        "regression-tests-incorporated"
      ],
      "corrected": [
        "runner-hardening"
      ],
      "struck": [],
      "decisions": {
        "runnerFailClosed": "adapt — adopt the fail-closed view rule with completion-marker reclamation: the runner reclaims only a view carrying its own .omni-run-complete marker and refuses every other existing view; the rm -rf default is removed"
      },
      "dissent": [
        {
          "claim": "anchored-parsing",
          "by": "curator",
          "objection": "Anchoring convicts the forged PASS; it also narrows what the parser can hear. Every pattern is now a whole-line contract with the runner's exact wording — a future rewording of a transcript line (a core upgrade changing a stage line, say) makes the parser fail closed: verdicts degrade toward FAIL and 'cause not established', never toward a false PASS. That is the safe direction, and the coupling is still real: the runner's transcript lines are now part of this tool's interface, and a wording change on either side is a breaking change on both. Recorded so the coupling is chosen, not discovered."
        },
        {
          "claim": "diagnostics-sidecar",
          "by": "measurement",
          "objection": "The sidecar is tamper-evidence, not authenticity, and the distinction must survive the adoption. Its hashes travel beside the files they describe: a writer who can rewrite a receipt can rewrite its sidecar in the same motion, and the manifest itself says so ('local-untrusted-hashes; no authenticity signature'). What it genuinely buys is detection of after-the-fact alteration against a retained digest — worth having, and never to be quoted as proof a receipt was honest when written. The cure for the stronger claim is the content-addressed store already queued on the environment backlog, not this patch."
        },
        {
          "claim": "verbatim-adoption",
          "by": "build",
          "objection": "Verbatim adoption is right for this code and wrong as a blanket rule. Her files arrived carrying no record of where they came from; the integration adds a provenance header to the service and files this examination precisely so the tree never holds code whose reasons live outside it. If the next external patch is adopted verbatim without its own examination, the tree gains machinery nobody here can defend line by line. The dissent is against the habit, not this adoption: verbatim is earned per package, by examination, or not at all."
        }
      ],
      "at": "2026-11-07T13:36:33.497Z"
    }
  ]
}
