What this document is. The constraint topology that makes new mechanisms imaginable. Every mechanism in this program — the serialization thesis, the surprise economy, the control organ, the spectral eye — was imagined into a shaped gap, and the shapes were carved by these constraints. Imagination without this topology produces aspirations; with it, it produces admissible candidates (which the billing then kills or funds). This is the internal companion to the anchor doc: ABI_CONTRA_PARADIGM_ANCHOR.md marks where external paradigms stop; this document is the grammar builders must generate within.
The admission rule (meta-constraint). No constraint enters this document without a receipt — the billed failure, proof, or empirical result that carved it. A constraints doc that admits asserted principles is an aspirations doc. Every entry below cites its receipt. A proposed constraint without one belongs in a design doc as a hypothesis, not here.
How to read an entry. Statement (what it binds), Receipt (what carved it), Instances (where it has been applied or where it caught a violation).
C0. Influence-control, not access-control. Compartmentalization means provable non-influence, not restricted access. Encryption guards who can read the bits; the requirement is that knowledge not retrieved contributes nothing to computation — not less, nothing. This forces knowledge to have addresses, which forces entries to mean something alone, which forces grounding, which forces receptors. Grounding is downstream of compartmentalization. Receipt: the program's pre-sketch driving problem; discharged empirically by the no-transformer result (the fully address-resolved configuration reaches 0.93x and surpasses at 1.11x — the auditable architecture is not a degraded fallback). Instances: mental model as addressable store; append-only experience log; organ federation's three-media coupling; T149 (the same structure makes unbounded imagination affordable).
C1. The lived-only firewall. Imagined outcomes never update certainty. The experience log accepts only provenance-tagged lived entries; certainty derives only from log deltas; generators write to the environment boundary and nowhere else. Prediction proposes; the world alone funds. Receipt: T113 (imagination funding certainty = private ideology, run privately); the paranoia instance (imagined agent behavior rewriting trust); monitored via calibration divergence, certified by deliberate-breach arms (P46c, P50). Instances: NEXT_SURPRISE Sec 7; the control organ's ledger inherits it by construction; identity change must route through the world (T149).
C2. Learnable surprise, never raw surprise. Every scoring use of predicted surprise multiplies by predicted reducibility. The form (rarity × reducibility, strain × yield) is constitutional; only its parameters evolve, with the mixing floor bounded below so selection cannot zero the reduction term. Receipt: the noisy-TV degenerate maximum (irreducible noise is maximally surprising forever); the adversarial extension (T141: an agent can hold structure irreducible — the TV that watches back). Instances: generation score; exploration priors; P45(c)/P46(b) arms.
C3. The probe-rate floor. Inheritance accelerates but never bypasses grounded experience. The floor lives outside the genome and cannot be selected to zero. Receipt: topology bias inheritance design; without it, inherited priors become unbilled beliefs — C1's cross-generational instance.
C4. Provenance and immutability of ground truth. The experience log is append-only; no learned process overwrites history. Legibility, unlearning, and safe failure are structural consequences, not features.
Receipt: C0's direct corollary; every debugging success in the program (anxiety loop traced to entries; token entry_refs addressing the exact failed beliefs) is its ongoing receipt.
C5. Receptor-native. Capabilities enter as firing conditions in the observation vector, never as silent subroutines. Capability without receptor is latent and never gets used (T3); a silent matcher improves behavior, a firing receptor becomes targetable by curiosity, billable by selection, and schedulable by the controller. Receipt: T3 supported directly (anxiety-loop break was receptor-specific — random channels and shuffled receptors do NOT break it); the thinking substrate (MCTS made thinking selectable by externalizing it as channels). Instances: pattern_novelty and pattern_match as receptors, not library internals; thinking channels; surprise fold-back channels.
C6. One currency per organ; the loop closes; no translation layers. Receptors fire in the currency's terms, effectors act in them, the organ's mental model maps between them. Any design where the organ selects outputs that get compiled into another system's vocabulary has reintroduced the gap and will fail credit assignment. Receipt: control organ anti-requirements (motor compilation is many-to-many against influence deltas); visual organ precedent. Instances: allocation modulations ARE the acts; visual words, not visual commands.
C7. Funded and evictable. Every mechanism pays on the organism's own ledger, and the eviction path is real. Applies reflexively: the predictor and generator are funded organs; a topology term that shows no margin dies by its own ablation arm. Receipt: the funding equation; P48's carving-yield claim explicitly NOT billed after multi-seed non-replication — the discipline held against the program's own hopes. Instances: generation VoI ledger; receptor metabolic costs; allocation cost coupling (holding the dial is not free).
C8. Seeds, not specifications. The genome is a search space, not a blueprint. Provide minimal seed receptors; refinement is earned through conflation-driven discovery (T125). The genome is the seed, not the ceiling. Receipt: discovered receptors beyond the genome (contextual_signal_interpretation, epigenetically transmitted); the fractionation discovery pathway. Instances: control organ's two-seed minimum (exertion-effect, shared-fate); visual pattern family as 9 seeds.
C9. Couplings learned or heritable, never hard-wired (the T82 discipline). The designer's guess about what matters must remain overridable. New channels are exposed to the Router/policy/search; mixing weights are heritable and initialized at zero, so selection bills the coupling rather than the designer asserting it. Receipt: T82 confirmed — the organism's evolved sense of what matters differs from the designer's specification (thinking budgets collapsed, value weights shifted). Instances: forecast-informed search prior (init 0); w_topo heritable; the refused hardwire of material-value into MCTS evaluation (the fix that fit: visual-sensitive thought types, so the motor store earns the coupling).
C10. Demand-driven, never top-down. Nothing is built in anticipation of a capability. A mechanism may begin only when something already running demands it — a strain, a falsification, a billing gap. Design happens at the level of conditions (environments, criteria, seeds, disciplines), never outcomes. Receipt: the Epistemic family born from T55's falsification; rank-rarity forced by three deaths; the young-ledger law found, not planned; withdrawal posture emerged, unscripted. The founding sketch's own annotation: "no self reflection" on the top-down box. Instances: the organ-discovery rule (an organ is demanded when strain resists every existing currency — receptor discovery one level up); register-before-building (P48's baseline condition preserved).
C11. Environment-first — and proportionally survival-contingent. The structure must exist in the environment, be survival-contingent at a magnitude selection can see, and be recoverable in principle (oracle checks) before any detector for it is built. Receptors for absent structure are metabolic waste by construction, and the battery must be able to refuse to fund them. Nominal contingency is not contingency: a consequence worth ~0.3% of fitness variance is invisible to selection regardless of how cleanly the signal is recoverable. Oracle checks certify recoverability; only a fitness-share check certifies load-bearingness — they are different properties and both are required. Receipt: T4 (the environment decides what's worth sensing); P1-style falsifier arms passing in both directions (material world funds material words; material-less world correctly funds nothing). The proportionality clause: material-world evolution run 1 (2026-08-08) — organ-level recoverability 8/8, behavioral funding flat across all arms because the material game was ~0.3% of fitness; single-parameter rescale registered as run 2. Instances: hidden-agency environment before the control organ; metamer materials before the spectral claims; every organ build began with an environment build.
C12. Umwelt-relativity of all quantities. Surprisal, rarity, thresholds, magnitudes — measured against the organism's own lived history and its own carvings, never an absolute scale. Rank against self; the lived-history cap and the young-ledger law are consequences, not additions. Receipt: T138 supported — three absolute constructions failed acceptance with identified mechanisms (certainty-clip ceiling, quality-vs-rarity saturation, unsupported parametric tails); the rank formulation is the sole survivor.
C13. Falsifiers stated in advance; acceptance gates between phases. Every claim carries its refutation condition before the experiment runs; every phase has accept criteria; the next layer waits. And a falsification must be EARNED exactly like a confirmation: a negative result bills against the theory only if the instrument was sound — domain-matched, endpoint-independent, phenomenon present at measurable strength (the C20 pre-flight). A negative from a deficient experiment bills against the EXPERIMENT (C14 ledger entry about the instrument); the theory's status stays untested. Three verdicts exist, not two: FALSIFIED (sound instrument, prediction failed), VOID/instrument-failure (the run could not have detected the effect — or measured the intervention itself), UNTESTED (no valid attempt yet). The falsification-friendly culture makes overbilled negatives the easy sin — deficient code producing a "NOT SUPPORTED" feels like honesty and is actually noise in the ledger. Receipt: the registry discipline throughout; T55's productive falsification (the method's best advertisement: being wrong generated a receptor family). Counter-receipt (2026-08-09, the rule's origin): CV-P1 run 1 recorded as "FALSIFIED as designed" when its detector and environment were both unsound, P45(a) as "FAILED" when the predictor was out-of-domain, PC-8 as "NOT SUPPORTED" from an endogenous design with a contaminated metric — all three corrected to instrument-failure/untested.
C14. Failures are ledger entries, not embarrassments. Failed constructions are recorded with their mechanisms in the design docs, adjacent to the survivor. Receipt: the four surprisal estimator iterations, recorded in NEXT_SURPRISE Sec 2.1; the P45(a) confound, recorded in the experiment docstring.
C15. No toys; the full organism. Experiments run all systems — thinking, banks, masking, motor store, online mental model. Reduced organisms produce results about reduced organisms. Receipt: standing user directive after violations; the anxiety-loop results are meaningless below full width — the break WAS the width.
C16. Beat the number or don't change it. Working signals are not changed because a change "sounds right." The burden is a measured improvement. Receipt: standing user directive; the eigen coder earned its place at 1.11x, not by argument.
C17. Integrate at the level the mechanism operates. No wrapping mechanisms as feature extractors for other mechanisms; find the level where the new signal is native and integrate there. Receipt: standing user directive; instance: visual channels into thought-type groups (the motor store's native key) rather than bolted onto action selection.
C18. Controls that can actually refute. Placebo arms controlling for dimensionality; paired seeds with snapshot-identical engines; null calibrations per test type; negative controls with known-absent structure. Receipt: the random-channel control (163 noise channels do NOT break the loop — without it, T3 would be undecided); the P45(a) confound caught and redesigned; the magnetoreception false positive documenting the efference confound.
C19. Engineering hygiene with rationale. No large data through multiprocessing pipes (write temp files). Cache engine calls (embeddings reused, not recomputed). Observation-vector moves are deliberate, all-at-once bookkeeping events with the invariant stated (400 → 408 → 420, each documented). Receipt: standing user directives after real failures; the obs invariant discipline is why three extensions landed without breakage.
C20. The pre-flight check — run before compute, not after. Eight questions, answered in writing in the experiment's docstring before launch; a deficiency found post-hoc that any of them would have caught is a method failure, not bad luck:
- Domain match — is every trained component deployed in the distribution it learned? (A gate-passing predictor is domain-gate-passing, nothing more.)
- Endpoint independence — does the intervention write into the measured quantity? In this architecture measurement and intervention share substrate BY DEFAULT (annealing writes the certainty the stuckness detector reads; shortcuts zero the channels the composition metric reads; removing execution steps raises coordination share by arithmetic) — independence must be engineered every time, the way C1 engineered it for imagination and belief.
- Exogeneity — is the causal variable manipulated, or merely observed co-varying with time? Endogenous growth curves yield correlations; only treatments yield causes.
- Pairing proven, not assumed — a zero-treatment identity check (bit-identical arms) before any paired result is billed; global RNG state counts as part of the pairing surface.
- Phenomenon strength — does the environment contain the phenomenon at measurable strength, with the billing floor stated in advance? A comparison over an absent phenomenon is VOID, not a negative.
- Endpoint sensitivity (added 2026-08-11, F16's receipt) — can at least one pre-registered endpoint MOVE through a legal causal path at the magnitudes the intervention produces? Endpoint independence (check 2) firewalls the intervention out of the measurement; run too hard, it firewalls the benefit out too. E1's demand arms delivered testimony that was 64–72% corroborated-true and registered zero gain, because all three endpoints were structurally inert: closures starved by the no-closure regime, the radius EMA frozen by thousands of prior fits, certainty firewalled by design. Independence and sensitivity are a tension to be engineered jointly: for each endpoint, name the legal path from intervention to measurement and estimate whether the intervention's magnitude can traverse it.
- Genesis endpoints and receipt currency (added 2026-08-11, F24's receipt) — two clauses. (a) Any endpoint gated on a RECEIPTS-GATED GENESIS EVENT (closure, ratification, individuation, phrase birth) must be booked as a RATE over replicates, or explicitly booked conditional — the gate exists to make the event rare, so single-run presence/absence is systematically underpowered for exactly the events the program most wants to see (F18, F22, F24: the same lesson three times). (b) Phenomenon-strength receipts are DISTRIBUTION-BOUND: an existence demonstration under one world source does not transfer to another (closure propensity varied from gen-2 to never across three world sources); when the world source changes, check 5 must be re-earned there. Receipt: v4's fusion run — five endpoints UNTESTED after 41 minutes because the card gated everything on single-run closure in unscreened corpus worlds, one day after F22 impl. 8 billed the law; the ledger held the lesson but the checklist did not, and booking reads the checklist (T154's addressability clause, reflexively).
- Population closure (added 2026-08-13, the fifth-arrival wave's receipt) — is every SET entering a comparison — arms, matched controls, oracle events, endpoint membership, mediator traffic — fixed at treatment start or floored in advance, and is its membership unwritable by the treatment? Three first-iteration cards fell to this in one day: P105 v1's matched-control pool exhausted (12 < 14) with no floor stated; P105's oracle population carried zero strain events in two venues (check 5's clause arriving at the ORACLE, which had been floored only on the detector side); P86 v1's endpoints averaged service over each arm's own slot population while the treatment WROTE the population (69 vs 45 slots — the young composed slots' native miscalibration billed as a treatment effect). The ledger-mass clause is this law's magnitude face; this is its membership face. A comparison whose population the treatment authors is VOID by construction. INSPECTOR CLAUSE (added same day, after the check caught four deficiencies POST-run — a check that fires only in the post-mortem is a coroner): the card must ENUMERATE, in writing before launch, every rule-relevant CLASS DIMENSION its compared populations span — eviction classes and exemptions, EMA clocks and their horizons, caps and strides, immortality privileges — and state for each either 'closed' (populations match on it) or 'floored' (its imbalance is bounded and billed). The fourth catch's receipt: a mortality endpoint paired eviction-exempt inherited controls against mortal composed slots (F41) — mortal-vs-immortal by rule, printed as SUPPORTED 0.95 before the class audit voided it. THE MATURATION CLAUSE (2026-08-14, three receipts): cohorts compared on ACCUMULATION metrics (fits, solvency, offspring, funding) must be age-matched, or the young flagged as right-censored — a newborn structure's low count is age, not property, until it has lived. Receipts: F45 (late nursery cohort's solvency partly censoring), F68 (implication-born slots read as under-trafficked at 834 mean fits), F69 (the same cohort gained ~13,000 fits each in six further generations — the 'quiet half' was the YOUNG half, caught one experiment before it shaped an organ build). Receipt: one session's cost of not having the first five (2026-08-08/09): the P45(a) domain mismatch (rich-trained predictor deployed in the field env — checkable at design time), the movement metric contaminated by annealing, the PC-8 composition metric reading the shortcut mechanism's own zeroed channels (a −0.995 artifact), PC-8's endogenous confound, the rich-stack global-RNG pairing leak, and CV-P1 targeting run against a world with ~0.5 stuck clusters to bite on. Every one was caught by C13/C18 discipline after compute was spent; every one was preventable before. The sixth check's receipt: E1's demand null (2026-08-11).
The ledger-mass clause (added 2026-08-12, F30 impl. 7's receipt). A web at 1e5 receipts is a different physical regime than one at 1e3: accumulated lived mass makes every EMA-mediated quantity conservative in proportion to history, so identical interventions land orders of magnitude apart depending on when they arrive. Three nulls in one session shared this root (E1's demand arms, the split-knob, P76 v1-v3). Every card must state the ledger-mass regime its endpoints live in and size interventions to it; check 6's magnitude estimate is QUANTITATIVE (events x per-event effect vs accumulated state mass) or the card is not ready.
The floor family (pattern note, 2026-08-11). Three constitutional floors now exist, and they are one law three times: receipt-driven economies starve whatever produces no receipts, and the correction is always a floor, never a weight. The probe floor (inherited motor priors must keep being tested — habits re-validated), the expression floor (silenced receptors must periodically be un-silenced — blindnesses re-validated), and the pose floor / curiosity quota (F16: across 3,600 poses, the starvation term never once won attention — struggle signals renew every episode, blindness generates no evidence at all, so evidence-driven inquiry is structurally myopic; a reserved fraction of the pose budget must go to the starved class). Predictive use: any future evidence-driven allocator (attention, generation targeting, imagination budget) will exhibit the same starvation of silent classes and will need the same constitutional correction — check for the missing floor at design time.
When reviewing any proposed mechanism or experiment, check:
- Does it enter as a receptor, or as a silent subroutine? (C5)
- Does its loop close in one currency, or does a translation layer smuggle the gap back in? (C6)
- What is its funding condition, and is the eviction path real? (C7)
- Is it seeded or specified? (C8)
- Is any coupling hard-wired that selection should decide? (C9)
- What existing strain demanded it? If nothing did, why is it being built? (C10)
- Does the environment contain the structure, survival-contingent, oracle-checked? (C11)
- Is any quantity measured on an absolute scale that should be self-relative? (C12)
- Is the falsifier stated before the run? Can the controls actually refute? (C13, C18)
- Is a failure being smoothed over instead of recorded? (C14)
- Is the organism reduced anywhere? (C15)
- Is a working number being changed without beating it? (C16)
- Could imagination be writing to identity anywhere without the world's countersignature? (C1)
- Has the pre-flight check run — domain match, endpoint independence, exogeneity, pairing identity, phenomenon strength with billing floor, endpoint sensitivity, genesis-endpoint rate design with distribution-bound receipts, population closure? (C20)
- If the mechanism is an evidence-driven allocator: where is its floor for the silent class? (the floor family)
This document is the mechanism behind the program's own generativity. Constraints are the imagination's receptor topology: they give the design space a geometry, geometry gives gaps shapes, and imagination fills shaped absences (T127, T133, T137, T3 — applied at the researcher level). The constraints were earned the way receptors are earned — through billed failures — and they have a dependency structure with C0 as trunk. A collaborator who internalizes this document acquires the carvings that make the next mechanisms visible; that is what the document is for (T107: teaching is installing detectors). Publishing it changes the topology it maps (T112) — future mechanisms will be imagined against these constraints, including the ones that eventually revise an entry here. Revision requires what admission requires: a receipt.
The grammar says what may exist. The billing says what may persist. The demand rule says what may begin. Imagination generates inside all three — and that is why it works.
C21. Decompose before mechanizing. A receipt underdetermines its mechanism. When a fix is gated on a receipt, name the 2x2 (or minimal factorial) that decomposes the receipt into its candidate mechanisms; if the missing cells are cheap, run them BEFORE building. When the cells are expensive and the build proceeds on an undecomposed receipt, tier the fix's invasiveness to the decomposition's depth: opt-in organs and sandbox steps are tolerable bets, default rewrites and core-rule changes are not. Receipt: three same-day instances (2026-08-13). F38: the pooled-total verdict blamed endogeny until channel decomposition found a treatment-independent genesis constant. F44: the extinction was diagnosed as threshold-absolutism until the economics decomposed it into congenital insolvency. F47: the completed transfer 2x2 (solvency x protocol) showed the built repair was redundant in its motivating venue — solvency was necessary AND sufficient; the repair survives as a battery-sealed opt-in awaiting its true demand venue (a crossing that changes the activation grain). Cost of the undecomposed build: one hour, because the fix was tiered small (the rule's second clause, followed by instinct, now stated). Instances: the constant-channel cousin clause of C20 check 8 is this law's special case; the F47 method note is its founding statement.
C22. Gate mobility and the powered-replication rule (verdict validity's remaining legs; with C13 they are one tripod). (a) GATE MOBILITY: every locked gate must be PASSABLE AND FAILABLE under the realized distributions, and the conditioning event's BASE RATE is printed next to the verdict. A gate that cannot be passed bills phantom falsifications (VOID-BY-IMPOSSIBLE-GATE — F70's hammer: 2 usable refs can never be a majority of 7); a measure that cannot fail bills phantom nulls (VOID-BY-SATURATED-MEASURE — F76's 42-percent conditioning event pinned the ratio at 1.00; F78's 66-percent extension rate hid a DETERMINISTIC coupling at 0.00 +/- 0.09, the clause's second receipt within a day of proposal, against its own author). (b) POWERED REPLICATION: any interval-based claim whose statistic was chosen after seeing data must replicate on FRESH SEEDS at stated power before billing. Receipt: pw1c v2's C_OWN=+0.26 CI[+0.03,+0.49] — the program's first CAUGHT false positive at statistic grain, significant on the peeked family, dead at 5x power on fresh seeds.
C23. The grain-match clause (the sampling theorem of the Umwelt). Every card states the PHENOMENON'S NATIVE GRAIN and the INSTRUMENT'S GRAIN; a mismatch is UNTESTED BY CONSTRUCTION, before any run. A vocabulary holds only what its grain resolves: to represent a structure spanning k steps, some element must have temporal support <= k. REFLEXIVE LINE (2026-08-15, F81 — third same-day instrument-organ mismatch): the card's GATE grain must match the organ's OUTPUT grain, stated in the same sentence that states the phenomenon's grain. Gates are instruments too; every verdict-validity clause applies to them.
C24. The anchor arm (falsifier tests bound their own meaning). Every card built to falsify carries an arm whose success is PRESUMED — a positive-control anchor; when the anchor also fails, the falsification is RE-SCOPED from the claim to the instrument or venue before any theory-level conclusion bills. Receipt: F79 — without the REFERENCE arm, "arbitrary matches certified" would have billed against reality-grounding itself; with real familiar shapes ALSO priced as furniture, the falsification relocated to the perceiver (identity-blindness), which proved correct (F80-F91: the blindness was the temporal hole, and curing it opened four gates). Corollary (split-bookkeeping, F90): mass-preservation gates computed over PRE-EXISTING populations are void-by-construction when the treatment SPLITS members — traffic moves to funded descendants and reads as damage; the honest form is whole-lineage flow.
C25. Concept identity is geometry, never address (the best-match rule). Every cross-web comparison — invariance, divergence, corroboration, culture — is scored by BEST geometry match, never by slot id: canalized organisms file the same concept at different addresses, and same-id scoring manufactures phantom divergence. Receipt: F98/F99 — same-id invariance read 63→36→22 across three perturbations; best-match read 62→60→57. The "17-slot core" of T169 was an addressing artifact and was withdrawn the same day. Kin: the floor family gains its lifecycle law (F92 impl. 4: never-funded / young-voiced / dormant / other-held each need a floor; curiosity, youth, and otherness quotas built in the bus, defaults 0). Receipts (three, one law): XO-3 (state-typed slots cannot see entry events — billed then as UNTESTED-BY-EFFECT-REPRESENTATION, which this clause subsumes), P78 (scalar-per-generation quotients too lossy for the reflection tower), F77 (generation-grain fit thresholds cannot see step-grain act-response conjunctions: obs diverged at t=0, the ledger noticed at t=19,212 — the encoder rebuild, not the fit pipeline, is the only door event structure currently has).