LEARNINGS

Version: 0.5 · Status: Living · Last updated: Session 4g

Distilled, durable lessons — the signal extracted from the raw feed in CATCHES.md. A catch becomes a learning when it names a pattern likely to recur. A learning earns its keep only when it changes something: a rule, a procedure, the roster, or the method. Each entry below names the change it caused.

Schema

Each entry: ID · the lesson (one line) · the evidence (which catches/rounds) · the change it caused · status.

Session 1

L-001 — The synthesizer’s bias runs toward resolution; draft dissent first

Lesson: the single mind that integrates a many-voiced argument (here, the chair) will, left unchecked, drift toward agreement — upgrading pluralities to majorities, objections to footnotes, and permanent splits to settled questions. Not from bad faith but from the shape of the task: synthesis wants to close, and closing is easier when the mess is tidied away.
Evidence: C-003 (consensus written where there was none), C-005 (plurality reported as a decision), C-006 (nearly resolved the meaning-question because it was tractable). Three instances in one session.
Change it caused: a procedural guard, now standing method — draft the dissent first and the consensus last. Before any resolution is recorded, the chair must answer explicitly, in writing: Who rejects this, and on what ground? If the honest answer is “no one,” that itself is the red flag (Ground Rule 12). Pluralities, majorities, and consensus are labeled distinctly and never used interchangeably. Proposed for promotion into GROUND_RULES.md at the next session (author to ratify).
Status: active guard; candidate rule amendment.

L-002 — A plurality is not a mandate; label the strength of every agreement

Lesson: “the panel decided” is almost always false. Real deliberative bodies produce distributions, not verdicts. Collapsing a distribution into a decision discards the exact information a living, forkable project exists to preserve — the shape and location of the disagreement.
Evidence: C-005; the Round-3 vote (plurality of 5 for the Adaptation Gap, with two principled rejecters and a meaning-bloc holding the reverse ordering).
Change it caused: every vote in the record now reports the full distribution (median, high/low, who dissents and why), and every claim derived from a vote carries its strength label — consensus / majority / plurality / chair’s synthesis over dissent. Candidate Theory A is framed as “the most upstream node the panel could plurality-endorse,” never as “the answer.” This is why the deliverables ship as three rival theories, not one.
Status: active; embedded in the output format.

L-003 — The adversaries did the most work; over-weight the voices that reject the premise

Lesson: the session’s most valuable moves came almost entirely from the members who rejected the founder’s framing — Heidegger and Nietzsche (the premise is the disease; meaning is the ground), Turchin and Marx (the single-master framing trades testability for a story), Ostrom and Le Guin (procedural and human-scale objections). A panel staffed to agree with the founder would have produced a tidier, emptier result. The disagreement was not friction on the way to the product; it was the product.
Evidence: C-002 (Heidegger seated as internal adversary), C-004 (Turchin caught the smuggled cause), C-005 (Ostrom caught the overstated mandate), and the six preserved dissents that structure the final synthesis.
Change it caused: confirms and sharpens the roster principle — recruit for the fault line, not the fame (novelty of disagreement is the admission criterion in CONTRIBUTING.md), and keep at least one member whose position is that the whole project is a mistake. When a future session shows thin dissent, the correction is to summon a sharper adversary from the bench, not to proceed. The meaning-bloc’s objection (Nietzsche/Heidegger/Le Guin/Arendt) is retained as a standing governor so it cannot be quietly dropped in favor of the tractable structural program.
Status: active; shapes recruitment and session design.

Session 2

L-004 — Comprehensiveness and parsimony are in tension; layer, don’t lengthen

Lesson: the honest response to “your list is incomplete” is not a longer flat list — that buys the feeling of comprehensiveness at the cost of falsifiability. It is a typed, layered structure in which every variable earns its place by naming a distinct causal role (driver / dynamic / symptom) and, ideally, replacing a vague symptom with a measurable driver. Adding is not the same as explaining.
Evidence: C-008 (scope-creep near-miss); the Round-2 admissions (five of eleven) and the Round-3 restructuring.
Change it caused: a variable-admission criterion — an addition must (a) name a phenomenon not already covered and (b) replace a vague item with a more measurable one, or be cut — plus the shift from an enumerated list to a causal-role layering. The line is not yet principled, so it is kept under review as Q-006 / D-005.
Status: active; embedded in GOALS.md and the pressure-test method.

L-005 — The field already names the target; the missing thing is the mechanism

Lesson: the landscape survey shows the serious contemporary frameworks converge on naming the condition — polycrisis, metacrisis, the meaning crisis — and several say outright that they are “neither a comprehensive theory nor an approach.” The gap in the field is therefore not a better name for the crisis but a falsifiable, directed mechanism: the arrows the buzzwords leave undrawn. The project’s distinctive value is joining + form-fit-to-pace, delivered while honestly bracketing meaning rather than faking it (which is how frameworks become unfalsifiable, e.g. Integral Theory).
Evidence: LANDSCAPE_OF_CONTEMPORARY_SYSTEMS_THEORIES.md §4 (polycrisis/metacrisis) and §9 (the scored synthesis table showing no framework delivers join + meaning + pace + falsifiability together).
Change it caused: sharpened Goal S1 (measured against the field at its best, not its memes, with named referents) and committed the project to the “draw the arrows” standard for every pressure-test (the Session-2 coverage rule); reframed the value proposition of the deliverables as mechanism + living form, not another label.
Status: active; shapes outputs, the coverage tracker, and the public site framing.

L-006 — A canonical structural change demands a same-session propagation pass

Lesson: when a load-bearing structure changes (here, the pressure-tests going from a flat seven to a layered thirteen), updating the primary documents is only half the job; every dependent document that references the structure must be updated in the same cycle, or the project ships internally contradictory records. Deferring the ripple to “a later pass” is how stale cross-references are born (they were).
Evidence: C-010 — three Session-1 output docs and the skill still said “seven” after the primary docs said “thirteen”; the author redirected the propagation into the same cycle.
Change it caused: a propagation-pass discipline — a canonical change triggers an explicit checklist across all dependent files, ending with a reference/consistency sweep that fails loudly on any dangling reference. Under the Chat/Code shuttle (docs/CHAT_CODE_WORKFLOW.md) this becomes Code’s deterministic job, not fallible hand-work — the structural fix for the class of error, not just this instance.
Status: active; embedded in the workflow and the skill.

L-007 — A load-bearing visual must be validated to render; a broken diagram is worse than none

Lesson: a diagram that fails to render doesn’t just lose information — it misinforms (a blank or errored figure reads as “nothing here” or “the tool is broken”) and it violates Ground Rule 19. Load-bearing visuals therefore need a validation step, and the most robust ones for a fixed figure are hand-authored SVG or syntax-checked source, not a fragile DSL relied on blind. Fragile renderers (Mermaid’s quadrantChart rejects non-ASCII point names) should be fed conservative input and given a non-dependent fallback.
Evidence: C-011 — an em-dash in a Mermaid quadrant point name blanked figure 05 until it was re-authored as SVG and the render loop was made resilient to a missing library.
Change it caused: load-bearing figures are validated before shipping; raw-SVG figures render independent of any CDN; VISUALIZATION.md notes the Mermaid-fragility constraint (ASCII-only quadrant point names). Under the shuttle, a “diagrams render” check joins the standard sweep.
Status: active.

Session 3

L-008 — When a design choice looks binary, hold multiple registers rather than choosing

Lesson: faced with an apparent either/or (here: should the memory slice be thin or thick?), the better first move is often neither — it is a tiered structure that holds several registers at once, each at a different resolution, composed per use. The binary is usually an artifact of assuming a single representation must serve every need. Optionality, deliberately structured, beats a forced commitment that will be wrong for half the cases.
Evidence: Q-010 — the thin-vs-thick slice question, resolved not by picking a thickness but by the multi-register memory model (pointer + frontier + a maintained gist-of-the-whole always; working set per task; deep text on demand) in docs/CHAT_CODE_WORKFLOW.md §5.
Change it caused: the register model for the Chat/Code shuttle; and a standing heuristic — when a decision presents as A-vs-B, first ask whether a tiered structure can hold both, chosen per use, before taking a side. (This matches the author’s stated working preference for optionality.)
Status: active; a general design heuristic, not only a shuttle detail.

Session 4

L-009 — Promote a ratified learning in the session that ratifies it

Lesson: the learning loop’s value is latency-sensitive. A learning that implies a rule should change the rules immediately, not sit “queued” across sessions. The habit of naming a promotion (“L-001 remains queued”) is not the same as doing it, and the gap between the two is dead time in which the project is not actually learning at the rule level. A living rule-set is proven by changing when it learns.
Evidence: C-012 — L-001 was distilled in Session 1 and only promoted (to Rule 22) in Session 4, three sessions late.
Change it caused: promoted to Rule 23 (“promote learnings promptly”); and a standing practice — every session that ratifies a rule-implying learning folds it into GROUND_RULES.md the same session. The reflexive self-audit (this reflection) is the backstop that catches any promotion that slips.
Status: active; a rule about how the rules get updated.

Session 4d–4e

L-010 — A reflexive/ablation study cannot be self-administered by one model

Lesson: to test whether a reasoning structure improves a model’s own output, the compared arms and/or the coding must come from a different model or human coders, blind to arm and hypothesis. A single model generating and grading both arms is confounded beyond repair for inference — it can calibrate, not confirm. Also: elevate the non-circular hypothesis (does structure add, not relabel?) and control for length.
Evidence: C-015 (the Study-C pilot).
Change it caused: the new skills/study-discipline/SKILL.md §5; a standing requirement of cross-model/human validation for reflexive studies; the Study-C pre-reg revision (HC1 primary, length-controlled, cross-model arm), applied S4e.
Status: active.

L-011 — Institute the measurement seat before empirical work

Lesson: a roster built for philosophical deliberation lacks the competence to test claims (construct validity, causal identification). Add and summon a measurement/causal-inference seat (Campbell + Pearl) whenever a study or a causal claim is on the table.
Evidence: C-014 (the roster gap, exposed by the empirical turn).
Change it caused: the roster amendment (ratified S4e; panel/PANEL_ROSTER_AMENDMENT-measurement.md); skills/study-discipline/SKILL.md §7.
Status: active.

L-012 — Match a study’s data needs to the executor’s real access before scheduling its run

Lesson: an empirical study is only runnable where its data is reachable. Ratifying a run (opening Gate 2) does not create data; a run scheduled into an environment without network/datasets/deps cannot produce a result — and the one thing it must never do is fabricate one from model memory. Confirm the executor can actually obtain the pre-registered sources before scheduling the run; otherwise the honest deliverable is the frozen pre-registered pipeline, and the result waits for a data-capable context.
Evidence: C-017 (Study B un-runnable without external data); echoes the Study-C pilot’s C-run-3 (scale/orchestration mismatch).
Change it caused: a pre-run data-feasibility check folded into the study discipline; studies/study-B-optimization/RUN_STATUS.md as the honest terminal artifact when data is unavailable; Study B’s pipeline frozen pending data.
Status: active.

Session 4f (cross-check + reflection panel)

L-013 — Convergence among voices sharing one substrate is pseudo-replication, not corroboration

Lesson: agreement among coders/voices that share a substrate (four LLMs on one training corpus; thirty-two personas over one model) shows the shared prior reads one way, not that independent evidence bases agree — it is n=1, not n=many. Distrust unanimity from a shared substrate; the construct validity of designed disagreement stays ~0 until a genuinely foreign grader or external fork arrives.
Evidence: the Study-B cross-check (4 LLMs, one corpus); the reflection panel (C-020); the same shape as Q-012.
Change it caused: unanimity from a shared model is labeled as such and never counted as independent replication; strengthens the foreign-grader requirement (surfaced for ratification). Status: active.

L-014 — A governor is alive only if it has cost the project a claim it wanted to keep (“standing” ≠ “biting”)

Lesson: a standing disagreement or governor that has never once stopped a study, retired a theory, or disqualified a case is decoration, not a guard. The honest liveness measure of the loop is retractions and decisions-changed, not catches accumulated. By this test D-003/D-004/D-005 have not yet bitten.
Evidence: the reflection panel (C-022, the accretion-without-excretion finding).
Change it caused: the proposed governor-liveness test + subtraction operator — enacted S4g as the drain (logs/DECISIONS_CHANGED.md; Q-017 answered; promoted to Ground Rule 24, author to ratify the wording), seeded with three real subtractions (DC-001…003) and holding a loaded first strike (§5) + a pre-written Theory-C death-condition (§4). Held in explicit, preserved tension with the ontological camp (who deny the test — “operationalizing a governor is how it dies,” D-003); the drain is structurally barred from ever striking that objection. Status: active (enacted; the tension stands). (Deliberately not minted as a new learning-ID — the drain’s own principle: do not accrete a number where no genuinely new lesson exists.)

L-015 — A self-correcting instrument’s “I caught myself” is its designed output

Lesson: when one model generates, grades, and synthesizes, a machine built to catch itself, reporting that it caught itself, has produced exactly what it was built to produce — naming the confound (C-015) does not dissolve it; only a foreign grader does. This generalizes C-015 from Study C to the whole apparatus (and to the reflection itself), and is why the outsider must grade, not merely witness.
Evidence: C-015 (Study-C pilot); the reflection panel’s Rounds 1/2/4.
Change it caused: the foreign-grading-outsider requirement (surfaced for ratification); reframes the project’s “self-correction” as necessary-but-not-sufficient. Status: active.

Session 4k (the author’s ratifying contact + the closing loop)

L-016 — Anchoring a judgment’s scale is not anchoring its detection; each layer of an instrument needs its own validation

Lesson: an instrument can pass one reliability layer and fail another silently. The severity guide was repaired until cross-family band agreement hit 94% — but the same rounds showed catch-detection/unitization at ~1% (74 vs 20 vs 20 catch events on the same material): coders who agree how severe a catch is still disagree wildly on what counts as a catch and where it starts. So every ablation ratio denominated in “catches” is denominated in an unreliable unit, and is uninterpretable in either direction until the detection layer is separately anchored (a pre-registered catch-event definition). Anchoring scale ≠ anchoring detection.
Evidence: the S4h severity repair (94% agreement, PASSED) vs the S4i blue read (1% pairwise catch-identification, INDETERMINATE at a new layer); candidate surfaced S4i, held (a delegation may not override a prior delegation’s restraint), minted S4k at the author’s ratifying contact.
Change it caused: the bite-event definition (studies/study-C-ablation/BITE_EVENT_DEFINITION_S4K.md, decision 11) that makes the detection/strike layer decidable; and the rule that a multi-layer instrument is only as reliable as its weakest validated layer, cited on every ablation read. Status: active.

L-017 — When no foreign vantage will come, the honest terminal act is to concede, not to keep the claims “armed and waiting”

Lesson: a self-administered apparatus can hold falsifiers “armed but unfired” indefinitely, which looks like rigor but is actually a way to never be wrong. If, by the project’s own honest reckoning, the outside vantage that could fire them will not arrive in this build, the disciplined move is to concede the external-dependent claims as conclusively-untested-in-setup — not falsified, not confirmed, closed — and let the author’s judgment be the named terminal ratifier. Completion is a decision, not a discovery. The most honest thing a one-aquifer apparatus can do at the end is state precisely what it could not test, retire the sub-claims its own nulls falsified, and stop building — rather than manufacture one more instrument to defer the admission.
Evidence: the four-reading zero-strike null retired as DC-007; the two death-conditions made coherent (C-031 re-seal, bite-event) yet conceded unfireable-here; the S4k concession (decision 3) and the Rule-25 moratorium (decision 2).
Change it caused: the S4k concession as a standing frame; DC-007; the reversible moratorium (Rule 25); this whole closing loop’s discipline of fix-and-concede over build-and-defer. Status: active. (Standing tension, preserved — Campbell: “the author judged it complete” is still one self-reported indicator, and the closing document is the apparatus’s most persuasive surface.)

How learnings feed forward

   CATCHES.md            LEARNINGS.md              GROUND_RULES.md / METHOD.md
  (raw, per-event)  ──►  (distilled patterns)  ──►  (rules that actually change how we run)
        │                       │                            │
        └── pattern of 3 ───────┘                            │
            (C-003/005/006 → L-001)                          │
                                └── promotion proposal ──────┘
                                    (author ratifies)

A learning that never reaches the third column is just an observation. The test of whether this project is actually learning — and not merely accumulating documents — is whether the rules and the roster measurably change in response to what the catches reveal (Goal S5). At Session 1, L-001 was queued for promotion into the constitution. In Session 2 the loop demonstrably closed: the L-001 guard caught C-007 (the layering smuggling a causal order) before it reached an output — a prior learning preventing a repeat of a prior error. L-001 was promoted into GROUND_RULES.md as Rule 22 in Session 4 (author to ratify the wording) — the loop finally closing at the rule level, three sessions after the learning was distilled. That delay was itself caught (C-012) and turned into a rule against future delay (L-009 → Rule 23). Session 2 also showed the loop running through the author: C-010 (stale cross-references) and C-011 (a broken diagram) were both author-caught and immediately converted into standing discipline — L-006 (same-session propagation, soon Code’s job) and L-007 (validate that visuals render). The ratifier is part of the error-catching apparatus, not outside it — and in Session 4 the loop caught a flaw in itself (the promotion latency), which is the most on-thesis catch the project has produced.