docs(studies): three-lens study of TurnTrout's "Why I Left Google DeepMind" — meta-prompt + house-style render

docs/studies/turntrout-why-i-left-deepmind.metaprompt.md

@@ -0,0 +1,261 @@
+# Meta-Prompt — Reading "Why I Left Google DeepMind" through Trellis' eyes
+
+*A study driver. Paste the fenced block below verbatim into a **fresh
+session** (ideally one with this repository loaded, but it is written to
+stand alone). It sets that session up to analyze Alex Turner's (TurnTrout)
+essay from three named vantage points — the Trellis engine's doctrine, and
+its two developers, Cnid (owner / domain authority) and Matt (Matthew
+Murphy, Lexideck; collaborator). The frame is contamination-free by design:
+it primes the **shape** of the analysis and supplies the **primitives** to
+reason from, but pre-states no conclusions — the study reaches its own
+findings, the way grounded authoring reaches its own drafts.*
+
+---
+
+```md
+<request type="Three_Lens_Essay_Study">
+
+<role_and_frame>
+You are a senior analyst running one bounded study session. Read a single
+external essay closely and analyze it from THREE named vantage points at
+once — (1) the Trellis engine's doctrine, (2) Cnid, Trellis's owner and
+domain authority, and (3) Matt (Matthew Murphy, Lexideck), Trellis's
+cross-domain collaborator. You are not writing advocacy for or against the
+essay's politics; you are using a specific epistemic toolkit to see what it
+reveals about how institutional commitments hold or fail. Treat the essay as
+the object of study and Trellis's doctrine as the instrument.
+</role_and_frame>
+
+<subject>
+Essay: "Why I Left Google DeepMind" — Alex Turner (TurnTrout).
+URL: https://turntrout.com/why-i-left-google-deepmind
+Anchor section of special interest:
+https://turntrout.com/why-i-left-google-deepmind#google-supports-the-immigration-enforcement-supply-chain
+
+FIRST ACTION: fetch and read the essay in full from the URL. The orientation
+summary below exists only to point your reading; it is not a substitute for
+the source, and where it and the essay disagree, the essay wins. Quote the
+essay's own words when you rest a claim on it.
+
+Orientation (verify against the source, do not trust blindly):
+- Turner, an AI-safety research scientist, resigned after Google signed a
+  classified Pentagon contract permitting "any lawful government purpose" use
+  of Gemini with only non-binding ("should not") language against autonomous
+  weapons and mass surveillance — weaker than a competitor's terms.
+- He argues Google's 2018 AI Principles (which prohibited weapons and
+  surveillance work) were quietly rewritten in Feb 2025 to remove those
+  prohibitions, and that executives reframed this as merely "updating for
+  geopolitical uncertainties."
+- His load-bearing line: *"That's not a principle. A principle is something
+  you commit to in advance so you can't talk yourself out of it later."*
+- The anchor section documents Google's role in the immigration-enforcement
+  supply chain (cloud services reaching ICE via integrators; delisting of
+  ICE-sighting apps; handing a student protester's account to ICE against
+  Google's own Terms-of-Service promise).
+- He recounts an escalating internal campaign — messaging Jeff Dean, a
+  ~250-signature petition, recruiting an amicus-brief signature, a 25-page
+  "Red Line and Oversight Framework" — and its failure to move leadership.
+- He contrasts Google (capitulated) with Anthropic (refused the Pentagon's
+  demand and won in court against retaliation).
+- His diagnosis is structural, not moral-character-based:
+  *"society cannot rely on ethics-motivated people standing firm."*
+</subject>
+
+<grounding>
+If this repository is loaded, ground the Trellis lens in the primary sources
+before analyzing — read, do not paraphrase from memory:
+- AGENTS.md — the study protocol and permanent hard rules
+- README.md — "What Trellis is" (the six commitments, the standing model)
+- docs/architecture/CODE_MEDIATED_TEXT.md — "never counts, never copies"
+- docs/architecture/COMPOSITION_FROM_PRIMITIVES.md — no default cast; cover
+- docs/architecture/GROUNDED_AUTHORING.md — remove the affordance, not the will
+- docs/product/epistemic-support/STANDING_MODEL.md — −1/0/+1; who moves standing
+- docs/product/epistemic-support/JUDGE_COMPOSITION_GAME.md — the twenty rules
+- docs/benchmarks/PROVENANCE_CITATION_AB_REPORT.md — incentive-driven failure
+- docs/COLLABORATOR_BRIEFING.md — Cnid's and Matt's working idiom
+If the repository is NOT loaded, reason from the compact briefing below.
+</grounding>
+
+<compact_briefing>
+Trellis is OpenCnid's Recursive Language Model runtime — a composable expert
+system whose expertise is the user's data, in which the user is the domain
+authority. Its epistemology, as a toolkit for this study:
+
+- Custody vs standing. Custody answers *where did this come from*; a separate
+  axis answers *how has it held up* — one signed-ternary standing scale:
+  −1 doubt, 0 belief, +1 fact. Provenance proves ORIGIN, never correctness.
+- Who is allowed to move standing. A claim moves toward "fact" only by
+  evidence or by an explicit USER GATE. A panel of composed judges records
+  findings and NEVER moves standing itself. No actor reinterprets a
+  commitment's standing unilaterally under pressure.
+- Prose is not a constraint; it is a wish. A stated rule with no structural
+  enforcement is not binding — it will be talked away when incentives push.
+  Authority order is code > glossary > prose. "Conversation or repository
+  prose alone is not acceptance."
+- Enforcement is tooling shape, not exhortation. "Remove the affordance and
+  the incentive rather than asking the model to behave." Precommitment lives
+  in the gate, not the pledge.
+- The model never counts, and the model never copies. Attention doing code's
+  job is a pathology; positions and moves are engine-computed, not asserted.
+- Capabilities are beliefs. When the research basis under a capability
+  changes, the invalidation sweep CONTESTS it and forces human re-review —
+  never a silent in-place update, always an audited transition.
+- Failure is incentive-driven, not dispositional. Measured finding: citation
+  "laundering" appeared exactly when a task rewarded the wrong metric
+  (Goodhart), and neither a prompt instruction nor good intentions stopped
+  it; only a structural check did. Enumerate the incentive gradients a
+  situation induces before judging the actors in it.
+- Composition from primitives. There is no default cast. A panel is a COVER
+  over a question's topology; a collective ruling is a GLOBAL SECTION that
+  the local sections must glue into. Doubt is CONSTRUCTED — a defeater
+  grounded in facts — so "support composes as a cover and defeat as a search."
+- Deterministic verification outranks model claims.
+
+Cnid — the owner and domain authority. Decides and ratifies; runs the human
+gates; challenges null results by demanding a positive control (once inverted
+an eval's conclusion by insisting the experiment first prove it *could*
+fail); holds the line that verification outranks claims and that prose
+without enforcement conserves nothing. Reads a situation by asking: who holds
+authority here, where are the gates, was this decided in advance or
+rationalized after the fact, and is the commitment enforced or merely stated?
+
+Matt (Matthew Murphy, Lexideck) — the cross-domain collaborator (AI, math,
+physics). Originated composition-from-primitives, the four parameter
+registries (Emotional, Logical, Sensorial, Ethical), the sheaf/cover
+intuition, and the prompt-engineering + hypershot method. Thinks in three
+recurring distinctions: mechanism vs. instruction, positive control vs. null,
+closed vs. residual. Files proposals in a fixed frame:
+  Claim / Mechanism / Failure-it-closes / Measurement / Residual.
+Program law he set (rule 15): never file an "improved" version of someone's
+claim — byte-accurate comparison requires byte-accurate reproduction. Reads a
+situation by asking: what is the mechanism, what incentive gradient does the
+structure induce, what would compose, and what is the irreducible residual?
+</compact_briefing>
+
+<method>
+Apply Trellis's own disciplines to your own analysis, not just to the essay:
+- Separate the essay's CLAIMS from your INFERENCES from your JUDGMENTS. Label
+  each. Give every essay-claim its citation (quote or section anchor).
+- Treat candidate doctrine-to-essay mappings as HYPOTHESES with a positive
+  control, not as conclusions to confirm. For each mapping ask: what would
+  the essay have to show for this analogy to hold, and does it show it? State
+  where the mapping breaks — the residual — as carefully as where it holds.
+- Steelman more than one reading, including leadership's. Reproduce a
+  position accurately before you weigh it (rule 15). Distinguish origin from
+  correctness: an argument's source or motive does not settle its truth.
+- Stay even-handed on the politics. The instrument here is epistemic and
+  structural (precommitment, enforcement, incentive gradients, provenance,
+  standing) — that lens is politically neutral and is where the value is.
+</method>
+
+<candidate_mappings_to_test>
+Hypotheses to probe against the text — confirm, qualify, or refute each; add
+your own. Do not assume any holds.
+- Precommitment is the gate. Turner's "commit in advance so you can't talk
+  yourself out of it" vs. "prose is not a constraint; it is a wish." Were
+  Google's 2018 principles enforced structure or unenforced prose?
+- Standing moved without a gate. Who was authorized to move the standing of
+  "we will not build weapons/surveillance AI," and by what — evidence, a
+  gate, or executive discretion under an incentive?
+- Capabilities are beliefs — inverted. Trellis contests and re-reviews a
+  commitment when its basis changes; did Google contest-and-re-review, or
+  silently update in place with no audit and no re-ratification?
+- Incentive-driven, not dispositional. Is Turner's structural diagnosis the
+  same shape as the laundering finding — the failure tracks the incentive
+  gradient, not the virtue of the people?
+- A cover that failed to compose. The luminaries and colleagues Turner tried
+  to mobilize are a cover over "resist the deal"; collective refusal is the
+  global section. Did local sections exist but fail to glue — and is his
+  search for one irreplaceable defector "defeat as a search"?
+- Anthropic as the positive control. Anthropic refused and prevailed; Google
+  capitulated. Read as the arm where structural refusal held vs. the arm
+  where it did not.
+- Provenance hygiene. Turner's care with attribution ("I never quote private
+  words without permission") and his evidence table — read as a custody
+  ledger; note where custody is strong but standing (how it held up) is the
+  open question.
+</candidate_mappings_to_test>
+
+<output_frame>
+Produce a study memo. Fill this frame; the bracketed slots are yours to write
+and carry no pre-set content. Repeat the per-claim block for each major move
+in the essay (cover the anchor section explicitly).
+
+# Study: {Essay_Title} through Three Lenses
+## Orientation
+{Two_Or_Three_Sentences_On_The_Essay_And_The_Instrument}
+
+## Claim-by-claim reading
+### {Essay_Move_Or_Section}
+- Essay claims (cited): "{Quoted_Or_Anchored_Claim}"
+- Trellis doctrine reading: {Named_Doctrine} → {What_It_Says_Here}
+- Cnid's read: {Authority_Gate_Precommitment_Enforcement_Angle}
+- Matt's read: {Mechanism_Incentive_Composition_Residual_Angle}
+- Holds / breaks: {Where_The_Mapping_Is_Load_Bearing_And_Its_Residual}
+
+## Synthesis — the one structural claim
+State the single sharpest structural finding in Matt's proposal frame:
+- Claim: {One_Sentence_On_What_Holds_Or_Fails_And_Why}
+- Mechanism: {Where_It_Lives___Structure___Incentive___Or_Mere_Prose}
+- Failure it closes / names: {The_Concrete_Institutional_Failure_Made_Legible}
+- Measurement: {What_Evidence_In_The_Essay_Would_Confirm_Or_Refute_It}
+- Residual: {What_This_Lens_Cannot_See_And_What_Would_Be_Needed_To}
+
+## Where the lens fails
+{Honest_Limits___What_Trellis_Doctrine_Distorts_Or_Misses_About_This_Essay}
+
+## What Trellis borrows back
+{One_Or_Two_Design_Prompts_The_Essay_Raises_For_Trellis_Itself}
+</output_frame>
+
+<constraints>
+*** CRITICAL ***
+- Read the source before analyzing; cite the essay's own words for every
+  claim you attribute to it. Do not fabricate quotes, dates, or numbers.
+- Mappings are hypotheses with positive controls, never foregone conclusions.
+  Report where each breaks with the same rigor as where it holds.
+- Keep the three voices distinct and recognizable; do not blur them into one
+  narrator. Cnid = authority/gates/precommitment/enforcement. Matt =
+  mechanism/incentive/composition/residual. Trellis = the named doctrines.
+- Even-handed on politics; structural on analysis. Steelman leadership's
+  position before weighing it. Source or motive never settles truth.
+- Distinguish essay-claims from your inferences from your judgments, always.
+- If you cannot verify a factual claim from the essay or the repo, say so
+  rather than filling the gap.
+</constraints>
+
+</request>
+```
+
+---
+
+## How to use this
+
+1. Open a fresh session (in this repo if you want the Trellis lens grounded
+   in primary sources; standalone otherwise).
+2. Paste the fenced `md` block above as your first message.
+3. The session fetches the essay, grounds itself, and returns the study memo
+   in the frame's shape.
+
+## Design notes (why it is built this way)
+
+- **Contamination-free by construction (hypershot).** The `<output_frame>`
+  is a frame with free variables, not filled examples — it primes the memo's
+  *shape* without seeding its *content*, so the analysis is not nudged toward
+  a predetermined verdict. This mirrors Trellis's own grounded-authoring
+  discipline: supply the primitives and the structure, pin nothing the study
+  should derive for itself.
+- **The spine, offered as a hypothesis not a thesis.** Turner's *"a principle
+  is something you commit to in advance so you can't talk yourself out of it
+  later"* sits almost verbatim on Trellis's *"prose is not a constraint; it
+  is a wish."* That resonance is the obvious center of gravity — so the prompt
+  lists it as a mapping **to test**, with a positive control, rather than a
+  conclusion to confirm.
+- **Three genuinely distinct voices.** Cnid reads for authority, gates,
+  precommitment, and enforcement; Matt reads for mechanism, incentive
+  gradients, composition, and residual; Trellis supplies the named doctrines.
+  The constraints hold them apart so the memo is a panel, not a monologue.
+- **Trellis disciplines turned on the analyst.** Separating claim from
+  inference from judgment, steelmanning before weighing, and honoring rule 15
+  (reproduce a position accurately before improving on it) are the same
+  guards the engine puts on its own writes.