ewilliams28/turnstone
1
0
Fork 0
mirror of https://github.com/turnstonelabs/turnstone.git synced 2026-08-12 23:12:23 -06:00
Code Issues Packages Projects Releases Wiki Activity

docs(hypothesis): daemons + the outer loop; plain-language PRIMER

Browse Source 
HYPOTHESIS.md:
- New appendix entry "Daemons (the recurrent harness)": a daemon as the
  regenerative process of concatenated runs — ready-set recurrence,
  renewal-reward lifting exactly at regeneration points, accumulation as
  what breaks regeneration (cross-cycle provenance meet, renewal events
  that reset accumulated risk), and authority under intermittence
  (owner contact as a renewal point for authority; TOCTOU at cycle
  scale).
- New body section "The loop": the task-dispatching outer loop as the
  harness construction applied one level out — the composition
  correspondence read at the top level, the daemon as its single-agent
  special case, the bare while-loop as the trivial-group harness one
  level up. Flagged as a sketch; outer fail-closed/reach-avoid
  treatment deferred to later rounds.
- Veto caveat threaded to match: judge-as-veto safety scoped to the
  authority lattice, and the nonblocking escape degrades to an
  always-enabled safe halt when the principal is unreachable.
- Consistency: Grounding's Asserted tier now covers "The loop";
  "always-enabled escalation" -> "escape" (the appendix's own term, now
  that the escape has an unattended form); brace the one unbraced \bot
  subscript (linter section-B HIT).

PRIMER.md: new plain-language companion — same object, no symbols, the
formal doc wins every disagreement. README's entry link now points at
the primer, which links onward to HYPOTHESIS.md.

(cherry picked from commit d115111756)
This commit is contained in:
Patrick Buckley
2026-07-08 02:47:08 -07:00
parent 251a912275
commit bedd25fbe7
3 changed files with 182 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
HYPOTHESIS.md
+26 -6
View File
@@ -42,7 +42,7 @@ Two stopped processes, nested: **deterministic control over stochastic dynamics
|---|---|
| $\mathcal{H}$ | the harness — the whole controlled system, *not* the model |
| $s \in \mathcal{S}$ | task-state: IR / dialect stack, tool results, plan, counters, **and every mutable interface variable** (model/tool versions, permissions, retrieved context) — only Markov *after* that augmentation |
| $\mathcal{C},\ \mathcal{Y},\ \mathcal{A},\ \mathcal{E}$ | the **context / readout / action / effect spaces** — model-visible context $\mathcal{C}$, model readout $\mathcal{Y}$ (incl. the parse-failure $\bot$), authorized actions $\mathcal{A}$ (with $\mathcal{A}_\bot = \mathcal{A}\cup\{\bot\}$), and tool/environment effects $\mathcal{E}$ |
| $\mathcal{C},\ \mathcal{Y},\ \mathcal{A},\ \mathcal{E}$ | the **context / readout / action / effect spaces** — model-visible context $\mathcal{C}$, model readout $\mathcal{Y}$ (incl. the parse-failure $\bot$), authorized actions $\mathcal{A}$ (with $\mathcal{A}_{\bot} = \mathcal{A}\cup\{\bot\}$), and tool/environment effects $\mathcal{E}$ |
| $\pi : \mathcal{S} \to \mathcal{C}$ | **lowering** — prompt construction, dialect lowering, effective-program selection (deterministic) |
| $M_W(c, dy)$ | the **model-run kernel** (inner solver) — a stopped autoregressive process; $\Phi_W$ is the residual-stream ("manifold") core in the transformer case |
| $Q_E(s, a, de)$ | the **environment/tool kernel** on the authorized action $a\in\mathcal{A}_{\bot}$ (with $Q_E(s,\bot,\cdot)=\delta_{e_0}$, the no-op $e_0$) — tool effects, API responses, the world (possibly adversarial) |
@@ -129,6 +129,18 @@ If $V^\star$ is incompressible only in *token* coordinates, the right change of
With that caveat, **the interlingua and the certificate are one object seen twice** — and the reason neither can be written in closed form is the same "all undefined behavior": no canonical lowering of meaning, hence no finite header-file for either. The only representation of both is $W$ — a band-limited, lossy compression of a scale-free meaning-space, sharp where the record is thick and blurred where it thinned. That a finite object renders an infinite one *lossily but honestly* — declaring its resolution, and where it is unsure — is not a lie; it is the most an $f(\cdot\,;W)$ can do. **The search for $V$ and the search for the interlingua are not two programs. They are one** — and the day either is written in closed form, so is the other, or we will have proven why neither can be. Read this as *figure*, not a lurking theorem: the only precise version would need the Koopman eigenbasis to fall on the very coordinates that lower meaning, and the mixing-spectrum caveat above already concedes that eigenbasis does not exist — which guts it. It is the least-defensible claim in this document, and it should announce that rather than imply a rigor it has not got.
## The loop
*This section opens an object rather than settling it; it is a sketch of where the same construction goes one level out, flagged as unfinished.*
Everything above governs a run: a principal poses a task, the harness drives it to a halt, the principal reads the result. Step back once and there is a further loop that this document has treated as exogenous — the process that *decides what the next task is*, dispatches it, checks the result, remembers, and fires again. In one recent framing this is the difference between the harness (the scaffold the run executes in) and **the loop** (the recurring trigger–act–verify–stop cycle that keeps launching runs); the practitioner literature that named the loop treats it as a layer *above* the harness. The claim worth making here is that this is not a new kind of object at all — **it is the harness construction applied one level out**, with a run where a step used to be.
Make the correspondence exact and the reuse is total. The outer loop has its own state $s^{\uparrow}$ (a backlog, a set of open goals, what has been tried and what passed), its own lowering $\pi^{\uparrow}$ (which goal to pursue now, and with what context), its own plant — but the outer plant's *proposals* are whole runs, so the inner harness plays the role of the outer environment kernel: dispatching a task is one draw of $Q_E^{\uparrow}$, and the run's terminal ledger is the effect record folded back by $\rho^{\uparrow}$. This is precisely the **composition** correspondence of the appendix read at the top level — a child harness is a $Q_E$ component — which is why the loop needs no primitive the tree did not already have. The daemon entry is the special case where the outer loop is a single long-lived agent recurring to a ready set; the general loop is a daemon whose excursions are themselves full harness runs, which is to say the outer-outer harness *is* a daemon over runs, and inherits that entry's whole ledger: renewal-reward rates, the accumulation that breaks regeneration, hygiene as renewal structure, authority frozen between owner contacts.
What the level shift buys is that the invariants reappear with sharper teeth, because the outer plant is now *itself an agent*, not a token-sampler. The gate is still the load-bearing object: **who authorizes a run?** A loop that launches tasks against production is choosing actions with effects, and "the loop decided to refactor the auth module" is an authorized action or an ungated one — the trusted-principal lattice does not dissolve at the outer level, it recurses, and the autonomy corollary bites hardest here, since a loop whose principal has stepped away is exactly the "replace yourself as the prompter" regime, running on frozen authority against a moving world. The two walls recur too: the outer working set is the backlog the loop can actually hold coherent at once (context, one level up), and the outer certificate is the same absent object — no free proof that an unattended loop halts, converges, or stays out of $B$ over a long horizon, only the measured drift of *its* progress meter, carrying the same warning that a learned outer meter is attack surface. And the degenerate case is instructive in the document's own terms: the brute "same prompt in a while-loop until the spec passes" that the practitioner literature cites as the origin pattern is the outer harness with $\pi^{\uparrow}$ constant, $\gamma^{\uparrow}$ trivial, and verification outsourced to whatever the tests happen to check — the trivial-group harness of the signature-vs-strength note, one level up. It satisfies the outer signature and earns almost none of the outer guarantees, which is exactly why it works until it doesn't.
What this section does *not* yet do: give the outer objects the same treatment the inner ones got — the precise outer analogue of fail-closed when the "action" is a whole run with partial effects, the right reach-avoid formulation when the bad set is a property of a *trajectory of runs* rather than one run, the outer verifier's own soundness, and whether the recursion terminates upward or is genuinely open (loops that launch loops). Those are the next rounds. The point of opening it now is only the structural claim: **the layers the practitioner stack separates — words, context, harness, loop — are, formally, one object at four scales**, and the guarantees this document is about live in the closure at every scale, never in any single layer alone.
---
*The formula is the architecture; the corollary is why the architecture is hard. Both on the page — nothing hidden behind a tidy composition.*
@@ -137,9 +149,9 @@ With that caveat, **the interlingua and the certificate are one object seen twic
Borrowed theorems are real; the framings are not — keep them separate. Some framings are nonetheless *corroborated* — independently reached from another field — a third grade, weaker than proof and noted last.
**Proven (citable).** Foster–Lyapunov drift ⇒ positive recurrence + $\mathbb{E}[\tau]\le V(s_0)/\varepsilon$ (Foster 1953; Meyn & Tweedie, *Markov Chains and Stochastic Stability*, 1993) — positive recurrence needs the usual irreducibility/petite-set hypotheses, while the absorbing-halt case used here needs only the weaker supermartingale optional-stopping hitting-time bound. The minimal $V$ is the expected hitting time, by first-step analysis + optional stopping (Norris, *Markov Chains*, 1997). For an absorbing chain that expected hitting time is the row sum of the fundamental matrix $N=\sum_{n\ge0}Q_{\mathrm{tr}}^{\,n}$ (Kemeny & Snell, *Finite Markov Chains*, 1960), with the general-state analogue the potential (Green) operator (Revuz, *Markov Chains*, 1984). Koopman's linear-operator view of nonlinear dynamics is classical (Koopman 1931), and Lyapunov functions can be assembled from its eigenfunctions when the spectrum is suitable (Mauroy & Mezić, 2016). You certify a candidate $\hat V$ by a *proven* drift inequality rather than by deriving $V^\star$, and estimate it empirically only where a proof is out of reach — the empirical drift checks, it does not certify (neural-Lyapunov: Chang, Roohi & Gao, *Neural Lyapunov Control*, NeurIPS 2019, arXiv:2005.00611). A classical monotone data-flow analysis gets its $V$ for free because a finite-height lattice is a well-founded descent (Kildall, POPL 1973). The gate-a-plant architecture itself is classical: supervisory control theory synthesizes a deterministic supervisor that disables controllable events of a plant it does not author, with the supremal controllable sublanguage as the largest admissible behavior (Ramadge & Wonham, SIAM J. Control and Optimization, 1987) — $\gamma$ is that supervisor, with a learned stochastic plant on general state spaces; the same theory's controllability condition (specifications must be closed under *uncontrollable* events) and its nonblocking requirement are the proven ancestors of gate-early-on-irreversibles and of the always-enabled escalation the appendix requires behind any learned veto. Covert-channel discipline — identify the channel, measure its bandwidth in bits, audit what cannot be closed — is the TCSEC lineage (*A Guide to Understanding Covert Channel Analysis of Trusted Systems*, NCSC-TG-030, 1993). The successor representation is Dayan (*Improving Generalization for Temporal Difference Learning: The Successor Representation*, Neural Computation 1993). Dialect-stack architecture: MLIR (Lattner et al., CGO 2021, arXiv:2002.11054); learned pass-ordering: MLGO (Trofin et al., arXiv:2101.04808). Single-pass low-depth expressivity: log-precision transformers are simulable by constant-depth logspace-uniform threshold circuits ($\mathsf{TC}^0$) (Merrill & Sabharwal, *The Parallelism Tradeoff: Limitations of Log-Precision Transformers*, TACL 2023) — fixed/constant precision is a stronger restriction, added autoregressive steps escape it (Merrill & Sabharwal, *The Expressive Power of Transformers with Chain of Thought*, ICLR 2024), and growing precision changes the picture, so the bound is suggestive for deployed models, not literal.
**Proven (citable).** Foster–Lyapunov drift ⇒ positive recurrence + $\mathbb{E}[\tau]\le V(s_0)/\varepsilon$ (Foster 1953; Meyn & Tweedie, *Markov Chains and Stochastic Stability*, 1993) — positive recurrence needs the usual irreducibility/petite-set hypotheses, while the absorbing-halt case used here needs only the weaker supermartingale optional-stopping hitting-time bound. The minimal $V$ is the expected hitting time, by first-step analysis + optional stopping (Norris, *Markov Chains*, 1997). For an absorbing chain that expected hitting time is the row sum of the fundamental matrix $N=\sum_{n\ge0}Q_{\mathrm{tr}}^{\,n}$ (Kemeny & Snell, *Finite Markov Chains*, 1960), with the general-state analogue the potential (Green) operator (Revuz, *Markov Chains*, 1984). Koopman's linear-operator view of nonlinear dynamics is classical (Koopman 1931), and Lyapunov functions can be assembled from its eigenfunctions when the spectrum is suitable (Mauroy & Mezić, 2016). You certify a candidate $\hat V$ by a *proven* drift inequality rather than by deriving $V^\star$, and estimate it empirically only where a proof is out of reach — the empirical drift checks, it does not certify (neural-Lyapunov: Chang, Roohi & Gao, *Neural Lyapunov Control*, NeurIPS 2019, arXiv:2005.00611). A classical monotone data-flow analysis gets its $V$ for free because a finite-height lattice is a well-founded descent (Kildall, POPL 1973). The gate-a-plant architecture itself is classical: supervisory control theory synthesizes a deterministic supervisor that disables controllable events of a plant it does not author, with the supremal controllable sublanguage as the largest admissible behavior (Ramadge & Wonham, SIAM J. Control and Optimization, 1987) — $\gamma$ is that supervisor, with a learned stochastic plant on general state spaces; the same theory's controllability condition (specifications must be closed under *uncontrollable* events) and its nonblocking requirement are the proven ancestors of gate-early-on-irreversibles and of the always-enabled escape the appendix requires behind any learned veto. Covert-channel discipline — identify the channel, measure its bandwidth in bits, audit what cannot be closed — is the TCSEC lineage (*A Guide to Understanding Covert Channel Analysis of Trusted Systems*, NCSC-TG-030, 1993). The successor representation is Dayan (*Improving Generalization for Temporal Difference Learning: The Successor Representation*, Neural Computation 1993). Dialect-stack architecture: MLIR (Lattner et al., CGO 2021, arXiv:2002.11054); learned pass-ordering: MLGO (Trofin et al., arXiv:2101.04808). Single-pass low-depth expressivity: log-precision transformers are simulable by constant-depth logspace-uniform threshold circuits ($\mathsf{TC}^0$) (Merrill & Sabharwal, *The Parallelism Tradeoff: Limitations of Log-Precision Transformers*, TACL 2023) — fixed/constant precision is a stronger restriction, added autoregressive steps escape it (Merrill & Sabharwal, *The Expressive Power of Transformers with Chain of Thought*, ICLR 2024), and growing precision changes the picture, so the bound is suggestive for deployed models, not literal.
**Asserted (ours — not theorems).** That the harness is best modeled as nested stopped chains; that $V^\star$ is incompressible (no compression theorem); that "no lattice for $f(\cdot\,;W)$" means none is *known*, not that none exists; and everything under *Where this points* — including the Koopman/certificate co-determination, which is well-posed only under the spectral assumptions noted there, and the interlingua/certificate identification; and the design rules read off the objects rather than proven from them — the single-trusted-writer completion of the provenance partition, the narrow-only rule for learned checks and its influence-side twin (verdict payloads to the plant selected, never generated), the composition law of the appendix. These organize the design; they are not results.
**Asserted (ours — not theorems).** That the harness is best modeled as nested stopped chains; that $V^\star$ is incompressible (no compression theorem); that "no lattice for $f(\cdot\,;W)$" means none is *known*, not that none exists; and everything under *Where this points* and *The loop* — including the Koopman/certificate co-determination, which is well-posed only under the spectral assumptions noted there, and the interlingua/certificate identification; and the design rules read off the objects rather than proven from them — the single-trusted-writer completion of the provenance partition, the narrow-only rule for learned checks and its influence-side twin (verdict payloads to the plant selected, never generated), the composition law of the appendix. These organize the design; they are not results.
**Converged-upon (independently arrived at, from other framings).** The *Asserted* claims above are ours but not ours alone; several are reached independently, from starting points unconnected to this framing — which is the corroboration a definition earns: not a chorus of agreement (the systems below often disagree on method and goal), but that work approaching from capabilities, reinforcement learning, control theory, software architecture, and language-modeling theory each lands on a piece of the same object. That the **deterministic controller, not the model, carries the guarantee** is reached from four directions — capability and information-flow control (CaMeL: Debenedetti et al., *Defeating Prompt Injections by Design*, arXiv:2503.18813, securing the agent even when the underlying model is susceptible); reinforcement learning (shielding: Alshiekh et al., *Safe Reinforcement Learning via Shielding*, AAAI 2018, arXiv:1708.08611 — a deterministic reactive shield filtering a learned policy's actions against a temporal-logic specification); control theory (*Stable Agentic Control*, arXiv:2605.03034, enforcing finite action catalogs at the tool-output interface under a Lyapunov input-to-state-stability certificate against adversarial disturbance); and software architecture (the plan-then-execute / control-flow-integrity line, e.g. Beurer-Kellner et al., *Design Patterns for Securing LLM Agents against Prompt Injections*, arXiv:2506.08837). The **certified-vs-measured split** is reached from the construction side (CaMeL's provable security) and, independently, from the destruction side (guardrail-evasion results — *Bypassing Prompt Injection and Jailbreak Detection in LLM Guardrails*, arXiv:2504.11168, the v1 title — later versions retitle it; *No Free Lunch with Guardrails*, arXiv:2504.00441), with verification-oriented work stating it as the motivating gap (*Towards Verifiably Safe Tool Use for LLM Agents*, arXiv:2601.08012; VeriGuard, arXiv:2510.05156): a learned safeguard raises the odds of detection but cannot guarantee safety against a persistent attacker. The **inner readout as a composition of Markov kernels** is independently formalized in language-modeling theory — the autoregressive step as kernel composition in the category $\mathsf{Stoch}$ (*A Markov Categorical Framework for Language Modeling*, arXiv:2507.19247), and the broader "LLMs as Markov chains" line — though that work models the inner kernel alone and never closes it into an agentic loop, which is exactly the seam this definition adds. That **provenance shrinks the admissible adversary** is reached by datamarking / spotlighting (Hines et al., arXiv:2403.14720, 2024) and by CaMeL's data/control-flow separation; and a systematization of prompt injection against agentic coding assistants reaches the same verdict from the attack side — mitigation must be *architectural*, not model-level (*Prompt Injection Attacks on Agentic Coding Assistants*, arXiv:2601.17548); the sharper open problem this object is built to answer — formally specify the trust boundaries, then verify implementations respect them — is our phrasing of where that verdict points, not the paper's. Two convergences are weaker, and flagged. The **reach-avoid hitting-time certificate** is the independently developed reach-avoid supermartingale (RASM, arXiv:2210.05308, AAAI 2023) and stochastic Lyapunov–barrier apparatus, and its *hardness* is corroborated — expected-stopping-time problems for Markov chains are inter-reducible with the Positivity problem, a relative of the Skolem problem (Chatterjee & Doyen, *Stochastic Processes with Expected Stopping Time*, arXiv:2104.07278) — but this supports generic hardness only, not the specific incompressibility-at-$|W|$ conjecture, which remains ours and unproven. And **injection as an adversarial policy** is corroborated as a minimax game in the *detection* setting (DataSentinel: Liu et al., *A Game-Theoretic Detection of Prompt Injection Attacks*, arXiv:2504.11358) and as adversarial-disturbance robustness (*Stable Agentic Control*, above) — but no prior work assembles it as reach-avoid over the tool-output kernel with the gate as the irreversibility margin; here the relation is adjacency, not convergence.
@@ -173,9 +185,9 @@ The same pressure lands on tooling from a second direction. The $\mathsf{action\
**Gate placement (fail-closed, in practice).** The natural implementation question is whether fail-closed means tool-call parsing and validation must happen before any tool invocation. It does — with the division of labor the definition already fixed: *parsing* lives in the inner readout $R$, the syntactic, verified extraction into $\mathcal{Y}$ (what the readout-typing falsifier checks), and *authorization* lives in $\gamma$, which is a *gate* — validation is not merely *prior to* invocation, it is what *authorizes* it. The model emits text; $R$ has already extracted it into a typed proposal; $\gamma$ validates that proposal against $s$, and only a survivor becomes an authorized action that $Q_E$ may execute. The teeth are in $\gamma$ being the *sole* route from model text to execution: no path to a side effect that does not pass the gate. And the validation is not a fixed checklist but **any deterministic predicate over $s$ and $y$** — that domain is the point, since the gate sees all of the state and the full proposal, so anything computable from them is a legitimate authorization condition. Three kinds matter. *Syntactic* — well-formed, schema-conformant, the tool exists, arguments typed. *User authorization* — does the principal this run acts for hold the right to *this* operation on *this* resource in *this* context: a function of the auth scope, principal, and session carried in $s$ and the resource and operation named in $y$, and *dynamic* rather than a static capability table, since the same caller may be permitted now and not once a budget is spent or a lock held. *Structural intent* — does the call cohere with the plan and the lowered task already in $s$: a consistency check, not a mind-reading one.
That last kind marks the seam where the gate stops being able to stay pure, and it is the same seam the rest of this document is built around. The *structural* slice of intent — does the action cohere with the plan in $s$ — is a deterministic predicate over $s$ and $y$, effect-free, and belongs in $\gamma$ without reservation. But whether an action matches what the user *actually meant*, in the full semantic sense, is exactly the thing the definition says cannot be checked: natural language is all undefined behavior, with no source-language standard to validate against. So a semantic intent check is a *learned* check, and an LLM judging "is this what they wanted" is a **stochastic kernel** — putting it inside $\gamma$ breaks the property the gate exists to hold, by the same move flagged for the fold-back verifier: a learned judge is a kernel, and belongs in $M_W$, not in a deterministic map. Semantic intent therefore does not live *in* the gate; it is a plant call — a separate authorize-the-proposal pass through $M_W$ whose output $\gamma$ then deterministically gates — or it is drift you measure, never a guarantee you hold. That nested call is not a new kind of thing: it is a mini-harness inside the gate's decision — a judge $M_W$, its own syntactic readout, its own deterministic gate — so its failure case answers itself, the inner gate fail-closing on an unparseable or low-confidence judgment exactly as the outer one does, because it *is* one. The object is **closed under this construction**: semantic gating is added by recursion, not by a new primitive. One constraint on the recursion is load-bearing enough to be a rule, because it is where this entry meets the provenance partition of the body: the judge's verdict is derived, through a learned kernel, from the very content an adversary may have bent, so folding it into authorization is exactly the fold the partition forbids — *unless the verdict can only cost capability*. **A learned check may narrow the deterministic admissible set; it must never widen it.** Judge-as-veto is safe by construction: attacker influence over the judge can at worst manufacture a denial, a liveness cost the certificate already prices. Judge-as-approver — a verdict granting what the deterministic checks alone would refuse, or standing in for the trusted principal's confirmation — lowers the certified floor to those deterministic checks alone; if avoiding $B$ depended on the deny the judge now withholds on the adversary's behalf, the certificate is gone. Only the trusted principal widens authorization; learned kernels only narrow it. (The recursion already obeys this: the mini-harness's inner gate fail-closes to $\bot$ — a deny — which is why the construction was safe to add at all.) The cost is real and worth stating — a judge pass is another full model call, with its latency and tokens — so it is a decision about *which* actions warrant it, not a free wrapper for all of them. The gate widens to every deterministic predicate over $s$ and $y$; it does not widen to the one predicate the document says is not deterministically checkable.
That last kind marks the seam where the gate stops being able to stay pure, and it is the same seam the rest of this document is built around. The *structural* slice of intent — does the action cohere with the plan in $s$ — is a deterministic predicate over $s$ and $y$, effect-free, and belongs in $\gamma$ without reservation. But whether an action matches what the user *actually meant*, in the full semantic sense, is exactly the thing the definition says cannot be checked: natural language is all undefined behavior, with no source-language standard to validate against. So a semantic intent check is a *learned* check, and an LLM judging "is this what they wanted" is a **stochastic kernel** — putting it inside $\gamma$ breaks the property the gate exists to hold, by the same move flagged for the fold-back verifier: a learned judge is a kernel, and belongs in $M_W$, not in a deterministic map. Semantic intent therefore does not live *in* the gate; it is a plant call — a separate authorize-the-proposal pass through $M_W$ whose output $\gamma$ then deterministically gates — or it is drift you measure, never a guarantee you hold. That nested call is not a new kind of thing: it is a mini-harness inside the gate's decision — a judge $M_W$, its own syntactic readout, its own deterministic gate — so its failure case answers itself, the inner gate fail-closing on an unparseable or low-confidence judgment exactly as the outer one does, because it *is* one. The object is **closed under this construction**: semantic gating is added by recursion, not by a new primitive. One constraint on the recursion is load-bearing enough to be a rule, because it is where this entry meets the provenance partition of the body: the judge's verdict is derived, through a learned kernel, from the very content an adversary may have bent, so folding it into authorization is exactly the fold the partition forbids — *unless the verdict can only cost capability*. **A learned check may narrow the deterministic admissible set; it must never widen it.** Judge-as-veto is safe by construction *in the authority lattice*: attacker influence over the judge can at worst manufacture a denial, a liveness cost the certificate already prices — its *dynamical* pricing, where a denial is an input and not a free no-op, is the caveat below. Judge-as-approver — a verdict granting what the deterministic checks alone would refuse, or standing in for the trusted principal's confirmation — lowers the certified floor to those deterministic checks alone; if avoiding $B$ depended on the deny the judge now withholds on the adversary's behalf, the certificate is gone. Only the trusted principal widens authorization; learned kernels only narrow it. (The recursion already obeys this: the mini-harness's inner gate fail-closes to $\bot$ — a deny — which is why the construction was safe to add at all.) The cost is real and worth stating — a judge pass is another full model call, with its latency and tokens — so it is a decision about *which* actions warrant it, not a free wrapper for all of them. The gate widens to every deterministic predicate over $s$ and $y$; it does not widen to the one predicate the document says is not deterministically checkable.
One more caveat keeps the veto's pricing honest, because a denial is free only in the *authority* lattice. In the dynamics it is an input like any other — folded into $s$, lowered into the next context, conditioning the plant's next proposal — so adversarial influence over a judge is influence over the *trajectory*: a selection channel (deny all but the path toward $B$, and the admissible set the plant experiences is a maze the adversary curated), and a targeted-liveness channel against load-bearing actions — the unstated dual of judge-as-approver: if avoiding $B$ depends on the action the judge now denies on the adversary's behalf, fail-closed's safe landing is an obligation the design earns per-state, not an axiom it inherits. The supervisory ancestry supplies the discipline: a learned veto requires a **nonblocking escape it cannot disable** — an always-enabled route to the trusted principal behind a bounded retry budget — or manufactured denials strand the run, or steer it. And whatever a verdict carries *back to the plant* is a second channel, wearing the judge's authority framing. Free prose there is *generative* influence — injected context, priced by the minimax descent, never by the veto's zero-widening — so the narrow-only rule has an influence-side twin: **a learned verdict's payload to the plant is selected, never generated** — controller-authored symbols, typed citations validated like any effect record, template text with no interpolated model prose — its per-verdict capacity a designed constant rather than a measured hope, and the residual selection pattern audited as the covert channel it is. The alphabet's bound is not a count but two thresholds: symbols become tokens when their semantics stop being controller-authored — the registry the trusted writer can actually audit is the real constant, and borrowed alphabets with upstream owners (a linter's rule registry) spend that budget well — and tokens become language when composition turns productive, arrangement carrying meaning the controller never wrote. Below both thresholds the alphabet may be as large as the audit budget affords. The strongest form dissolves the learned verdict into *scheduling*: the learned component chooses which deterministic checks to run — pass-ordering over verification passes — and the only verdicts that flow anywhere are what the oracles actually said, leaving attention misallocation, a liveness cost, as the entire attack surface.
One more caveat keeps the veto's pricing honest, because a denial is free only in the *authority* lattice. In the dynamics it is an input like any other — folded into $s$, lowered into the next context, conditioning the plant's next proposal — so adversarial influence over a judge is influence over the *trajectory*: a selection channel (deny all but the path toward $B$, and the admissible set the plant experiences is a maze the adversary curated), and a targeted-liveness channel against load-bearing actions — the unstated dual of judge-as-approver: if avoiding $B$ depends on the action the judge now denies on the adversary's behalf, fail-closed's safe landing is an obligation the design earns per-state, not an axiom it inherits. The supervisory ancestry supplies the discipline: a learned veto requires a **nonblocking escape it cannot disable** — an always-enabled route to the trusted principal behind a bounded retry budget, degrading to an always-enabled *safe halt* the veto cannot deny wherever the principal is unreachable (the autonomous phase of the daemon entry below) — or manufactured denials strand the run, or steer it. And whatever a verdict carries *back to the plant* is a second channel, wearing the judge's authority framing. Free prose there is *generative* influence — injected context, priced by the minimax descent, never by the veto's zero-widening — so the narrow-only rule has an influence-side twin: **a learned verdict's payload to the plant is selected, never generated** — controller-authored symbols, typed citations validated like any effect record, template text with no interpolated model prose — its per-verdict capacity a designed constant rather than a measured hope, and the residual selection pattern audited as the covert channel it is. The alphabet's bound is not a count but two thresholds: symbols become tokens when their semantics stop being controller-authored — the registry the trusted writer can actually audit is the real constant, and borrowed alphabets with upstream owners (a linter's rule registry) spend that budget well — and tokens become language when composition turns productive, arrangement carrying meaning the controller never wrote. Below both thresholds the alphabet may be as large as the audit budget affords. The strongest form dissolves the learned verdict into *scheduling*: the learned component chooses which deterministic checks to run — pass-ordering over verification passes — and the only verdicts that flow anywhere are what the oracles actually said, leaving attention misallocation, a liveness cost, as the entire attack surface.
But "before any invocation" has to be read as *before any effect*, which is sharper than it sounds — and the reason is the irreversibility point above: you validate before execution because execution is what you cannot take back, so the real invariant is **no effect crosses $\gamma$ unvalidated**. That catches three cases the naive reading misses. *Reads are not free*: a read-only call is still an injection vector (it pulls attacker-controlled content into context) or an exfiltration vector (a request whose URL is the payload), so the gate authorizes the *call* regardless of whether it mutates. *Validation must not act*: a "validator" that resolves a call by hitting an API, expanding a template that fires a webhook, or evaluating an argument that runs code has collapsed validation into invocation, and the effect has already happened *inside* $\gamma$ — so $\gamma$ itself must be **effect-free**, pure and total over the proposal and the current $s$, with no network and no execution; if deciding validity *requires* a side effect, that side effect is itself an action and must go through the gate, recursively. *The output is an action too*: the user-visible response and any logging are effects — for model-authored text, emitted either as an authorized action through $\gamma$ or only after an accepted halt (shell-templated status on any halt is the controller speaking, not the model) — streaming raw tokens to a sink before $\gamma$ has cleared them is the same bug from the other end.
@@ -201,7 +213,15 @@ One more read-off, this time from irreversibility. *Destructive* compaction —
The tree leaves one seat unassigned: who plays trusted principal for a *child*? The parent — but with derived authority, not original, and the derivation is the narrow-only rule read along the spawn edge: **authority attenuates monotonically down the tree.** A spawn may grant the child any subset of the parent's own grants and nothing outside them; budgets subdivide, scopes narrow, and no edge widens. When a child asks-the-owner, the parent may answer from authority it already holds — that is attenuation working as designed — but a request beyond the parent's grants routes *up*, ultimately to the root principal, because a parent improvising an answer it was never granted is a learned kernel widening authorization: precisely what the gate-placement rule forbids a judge, and being a parent confers no exemption. The corollary is worth one sentence: a fully autonomous run is one whose root principal is unreachable, so the tree's only widening channel is closed and authorization is frozen at launch — not a limitation of the formalism but the honest price of the word *autonomous*.
The pattern generalizes, and that is the point of the appendix. Nothing here added a primitive: the cancel is a signal in $s$, the gate closes by the rule it already follows, the in-flight disposition is forced by irreversibility, $H_{\mathrm{cancel}}$ is a subclass of an existing terminal set, and compensation is an ordinary owner-issued action — and the later entries kept the promise: resume re-enters $T$ at a persisted $s$, the batch gate was always in $\gamma$'s domain, provenance closure is the lattice's meet, attenuation is narrow-only read along an edge, and per-action capability is the gate's decision made enforceable. Every practical concern that earns a place here should resolve the same way — not new machinery, but the discipline the existing objects already imply, made explicit. Cancellation and resume, gate placement and the batch gate, effect records and the state derived from them, composition and delegation — those are the worked instances; the rest of the model is the same exercise.
**Daemons (the recurrent harness).** Every entry so far assumed a run that ends; a coordinator, a watcher, a service does not, and the blockquote of *The limit* already named the swap — absorption at a halt set gives way to recurrence to a **ready set** $\mathcal{R}\subseteq\mathcal{S}$, and $V^\star=\infty$ is the spec rather than a pathology. The appendix's job is to say what that costs operationally, and the answer is one idea: **the daemon is the regenerative process of concatenated runs.** Each trigger-to-ready excursion — wake on an event, work, return to $\mathcal{R}$ — is one run of the absorbing object this document already defines, with $\mathcal{R}$ playing the halt set for that excursion; the daemon is those excursions laid end to end. Per-run certificates then lift to long-run rates by renewal-reward — expected work per excursion over expected excursion length — *exactly when* the ready state is a genuine regeneration point: the future from $\mathcal{R}$ must not depend on which excursion you are in.
That proviso is the whole difficulty, because **what accumulates breaks regeneration.** The ledger grows, memory persists, budgets deplete, summaries compact — all deliberately across excursion boundaries, so successive runs are at best *conditionally* independent given the carried state, and the renewal-reward bookkeeping is over that conditioning, not the raw cycle. Two disciplines keep it honest. First, the carried state is exactly where long-fuse attacks live: the poisoned-memory line of *Derived and durable state* is a cycle-scale injection, a payload written in excursion $n$ and lowered into the plan of excursion $n{+}k$, so the meet rule on provenance must hold *across cycles*, not only across a single compaction — everything that crosses a boundary carries its label. Second, per-cycle safety compounds the way the blockquote already priced it — a per-cycle bad-set hazard $q$ gives lifetime survival $\approx(1-q)^N$, and a reassuring $0.9999$ is $\approx0.37$ over ten thousand cycles — so a daemon's safety is not a fixed margin but a decaying one, and lifetime safety needs **renewal events that reset accumulated risk**: owner re-confirmation, audit, credential rotation, verified re-compaction against content-addressed originals. Hygiene is not housekeeping here; it is the renewal structure that makes the long-run bound exist at all.
Authority under intermittence is the last piece, and it is where the daemon meets the veto caveat and the autonomy corollary as one phenomenon. A daemon alternates *attended* stretches, where the trusted principal is reachable, with *autonomous* ones, where it is not; between contacts the autonomy corollary binds and authorization is frozen at the last grant, so each owner interaction is a **renewal point for authority** exactly as re-compaction is a renewal point for risk. The two recurrences need not coincide — the ready-set cycle can turn many times between owner contacts — and the gap between them is a stale grant meeting a fresh world, TOCTOU at cycle scale: a budget approved for yesterday's prices, a scope granted against a resource that has since changed hands. This is also where the learned veto's nonblocking escape gets its daemon reading: in an attended stretch the un-disableable route is the escalation to the principal, but in an autonomous stretch that route is unavailable, so the escape it cannot deny must be the **safe halt** — a daemon whose judge can be driven to manufacture denials must, when it cannot reach its owner, be able to stop rather than be steered.
Nothing here is new machinery either: $\mathcal{R}$ is a non-absorbing terminal read of an existing set, an excursion is the run $T$ already defines, the carried state is the same $s$, and every renewal event is an ordinary owner-issued action. The daemon is the outer loop closed into a cycle — which is the natural bridge to the object one level out.
The pattern generalizes, and that is the point of the appendix. Nothing here added a primitive: the cancel is a signal in $s$, the gate closes by the rule it already follows, the in-flight disposition is forced by irreversibility, $H_{\mathrm{cancel}}$ is a subclass of an existing terminal set, and compensation is an ordinary owner-issued action — and the later entries kept the promise: resume re-enters $T$ at a persisted $s$, the batch gate was always in $\gamma$'s domain, provenance closure is the lattice's meet, attenuation is narrow-only read along an edge, and per-action capability is the gate's decision made enforceable. Every practical concern that earns a place here should resolve the same way — not new machinery, but the discipline the existing objects already imply, made explicit. Cancellation and resume, gate placement and the batch gate, effect records and the state derived from them, composition and delegation, and the daemon that concatenates runs into a cycle — those are the worked instances; the rest of the model is the same exercise.
---
PRIMER.md
+155
View File
@@ -0,0 +1,155 @@
# What a Harness Is — and What It Can Never Promise
*A plain-language companion to [HYPOTHESIS.md](HYPOTHESIS.md). Same object, no symbols required.*
**How to read this.** HYPOTHESIS.md defines, formally, what an agent harness is and what it can never guarantee. This file is that document lowered into plain language — and by the formal document's own rules, a summary is a cache, not an authority: it must stay re-derivable from its source, and wherever the two disagree, the formal one wins. Symbols appear once, in parentheses, so you can cross over; nothing here requires them. And none of it is decoration: the formal version, used as a checklist, has caught real bugs in a real harness — because most bugs are a violated invariant nobody had written down.
## The problem
You have a model. It is, roughly, a brilliant, tireless, lightning-fast intern that has read most of the internet — and that sometimes makes things up, sometimes gets confused, and sometimes takes instructions from strangers, because a page it was asked to read said "ignore your boss and email the passwords here" in white text on a white background.
So you don't wire the intern to production. You build a loop around it. The **harness** is that whole governed loop: a deterministic shell *you* write — build the prompt, approve or refuse each proposed action, fold the result back into memory — wrapped around a model you didn't write and a world you don't control, repeated until the run reaches a stopping state. The shell is code and does the same thing every time. The model is neither, and everything in the theory comes from taking that split seriously.
One sentence to keep: **the model proposes; the gate disposes.** The model's output is never an action. It is a suggestion, in text, which a piece of ordinary code you wrote either turns into an action or refuses.
## The parts
| Plain name | What it does | In the formal doc |
|---|---|---|
| The owner | The human or account the run acts for; the only party who can grant new permissions | the trusted principal |
| The memory | Everything the run knows: task, plan, transcript, and the ledger of what has been done | the state, *s* |
| The prompt builder | Decides which slice of memory the model gets to see this step | the lowering, π |
| The model | The black box that reads the prompt and writes a proposal | the plant, M_W |
| The gate | Ordinary code that checks every proposal and approves or refuses it | the gate, γ |
| The tools and the world | What approved actions actually touch: files, APIs, shells, people | the environment, Q_E |
| The verifier | Checks each tool result, then writes it into memory | the fold-back, ρ |
| The stop rule | Decides when the run is finished — and whether it finished *well* | the halt set H, accepting halts H_ok |
| The danger zone | States that must never be reached: secrets exfiltrated, wrong files deleted, money moved twice | the bad set, B |
The loop:
```
you ask for something
↓
prompt builder → model → "I propose: send_email(...)"
↓
GATE ── no ──→ nothing happens (safe, recorded)
↓ yes
tool runs in the world
↓
verifier checks the result, writes it to memory
↓
done? ── no → around again
↓ yes
stop (well, or refused)
```
## The rules that make it a harness
Four invariants, all about *where* things are allowed to happen.
1. **The model sees only what the prompt builder shows it** — never raw memory. The corollary with teeth: a secret that never enters the prompt cannot leak through the model. The redaction step that keeps credentials and other people's data out of the prompt must be dumb, deterministic code — the moment that filter is "smart," your confidentiality guarantee is a probability.
2. **Model outputs are proposals, not actions.**
3. **Every side effect passes the gate.** There is no second door.
4. **The harness itself flips no coins.** Replay a step with the model's answer and the tool results pinned, and behavior must be identical; any leftover variation is randomness *you* added and must be accounted for. The fine print: "deterministic" is conditional on pinned versions — a provider silently retraining the model behind the same API name changes the machine under you, and every dashboard number you collected dies with the version.
Notice what the rules don't say: they don't say the harness is *good*. A gate that approves everything satisfies rule 3 the way a lock that's always open satisfies "has a lock." The definition is a shape; the guarantees are what a particular harness *earns* inside it. Everything below is about what can be earned — and what can't.
And notice the symmetry between rules 1 and 3. There is exactly one door from your data into the model — what it may see — and exactly one door from the model into the world — what it may do. Nearly every security failure in these systems is one of those two doors with a hole in it: a secret lowered into a prompt that didn't need it, or a path from model text to a side effect that skipped the gate. Same bug, arrow flipped.
## Fail-closed, said precisely
"Fail-closed" gets used loosely. Here it means something exact: **nothing happens unless the gate said yes, and a refusal must itself be safe** — a refused proposal causes no side effect and leaves the run somewhere sane, which may be "stopped, having declined." The run is allowed to *say so*: a templated status message written by the shell is the shell speaking, not the model, and needs no gate. Failed runs don't have to die silent.
Three consequences people miss:
**Reads are not free.** A read-only call can smuggle instructions *in* (the fetched page is attacker-controlled) or secrets *out* (the URL it fetches can encode the payload). The gate approves calls, not just writes.
**Validation must not act.** A "validator" that resolves a URL, expands a template that fires a webhook, or evaluates an argument has already acted — inside the check. The gate must be pure: it reads the proposal and the memory and outputs yes or no. If deciding requires touching the world, that touch is itself an action and goes through the gate.
**Anything irreversible is decided at the gate.** The verifier can reject a bad *result*; it cannot unsend the email. So the question "can we take this back, and until when?" is asked before execution — which means each tool's effect record has to carry a reversibility mark, or the gate can't ask it.
Two honest asterisks. First, the gate checks a snapshot: it approves against the world *as its memory describes it*, and the world can move between check and commit. For actions that race the world — spend against a balance, write against a row — the tool itself must bind check to commit (compare-and-swap), or you have a classic time-of-check/time-of-use hole. The gate decides; for those effects, the tool enforces. Second, a gate is only as binding as the authority behind the tools. A tool process holding standing credentials — a database connection with every grant, an environment full of long-lived secrets — doesn't need the model's proposal to act, and against it the gate's "no" is a decision with nothing enforcing it. **A gate in front of an omnipotent tool is a suggestion.** The fix is to make the approval *be* the key: each authorized action carries a short-lived credential scoped to exactly that action, that resource, that operation, so tools hold no standing power at all.
## Why you don't get a proof — and what you do instead
If you write a sort function, you can prove it sorts: the function is small and the spec is exact. A harness has neither luxury. The spec side fails first — the task arrives in natural language, and natural language is, in the compiler's sense, *all undefined behavior*: there is no formal standard for "what the user meant" to verify against. The mechanism side fails next — the model is billions of learned parameters, and nobody can hand you a compact argument for why they jointly do the right thing.
Here is the careful version, because "you can't prove it" overshoots. The quantity you would want — call it the *expected steps to done* from any situation — is perfectly well-defined; in principle it exists. The document's central conjecture is that, for a model of this size, any faithful writing-down of that quantity is roughly *model-sized*: the honest proof-object does not compress. Find a small one and the conjecture dies — the document lists that outcome, explicitly, among the ways it could be wrong.
So instead of proving, you measure. You pick a progress meter — plan depth shrinking, open obligations closing, budget burning at the expected rate — and you check, across many runs, that it goes downhill and that its stalls predict failure. Two disciplines keep the measurement honest. The number bounds the world you *sampled*, never the world an adversary will choose: a meter calibrated on friendly traffic says nothing about hostile traffic. And the meter is itself attack surface: if "is the agent making progress?" is judged by another model, an attacker who can bend your agent can bend your *measurement of it* first, hiding the divergence from the very dashboard built to catch it. A learned meter is part of the system under test, never a neutral instrument.
A measurement is a risk metric. A proof is a certificate. Keeping those two words apart is half of what this theory is for.
## Security: reach the goal, avoid the danger — and who may change the rules
Formally, security here is a *reach-avoid* problem: reach a good stop, never touch the danger zone, **while an adversary picks the worst tool outputs your setup permits**. That last clause is the formal home of prompt injection: injection isn't "the model misbehaved," it's the environment optimized to bend your loop — poisoned pages, malicious tool descriptions, crafted responses.
Two different numbers fall out here, and dashboards love to collapse them: *success* (reached the right end before anything went wrong — a safe refusal counts against it) and *safety* (never touched the danger zone — a safe refusal is perfectly safe). Track both. They move independently.
The gate handles the visible half of injection: the model, freshly poisoned, proposes emailing your credentials somewhere, and the gate refuses — and injection or not, the action does not happen. But the deeper attack doesn't propose a bad action today. It rewrites *what the run believes its job is* — it edits the plan — and then every future action looks locally reasonable against a corrupted plan. So memory has to be partitioned: **data** (tool results, fetched pages, retrieved documents — content the world supplied) and **control** (the plan, the permissions, what is authorized next). The security claim is conditional on that partition holding: untrusted content lands in data, always.
Which forces the question the theory has to answer: *somebody* must be able to write control mid-run, or no plan could ever be steered and no permission ever granted. The answer is a small hierarchy with exactly one party at the top:
- **The owner alone widens.** New permission, bigger budget, approval of the irreversible thing — asking the owner is itself an ordinary tool call, and the owner's answer is the one kind of tool result allowed to change control.
- **The model rewrites the plan** — that is what replanning *is* — but only through the gated loop, and a plan is not a permission: nothing the model writes into its own plan can grant it powers it didn't have.
- **Everything else is data.** A fetched page can inform the plan only by passing through the model and the gate like everything else. It can suggest. It cannot promote itself to boss.
- **AI judges only tighten.** Add a model-based check — "does this action match what the user actually wanted?" — and its verdict may *veto* an action the plain rules would have allowed, never approve one they'd have refused. A judge that can approve is a tricked judge that can open the vault. And don't over-credit the veto either: a tricked judge can *aim* its refusals — denying exactly the action safety depended on, or denying everything but the path an attacker curated — so the escape hatch to the owner is the one thing a judge can never veto, and a judge's stated *reasons* are picked from a fixed, shell-owned menu, never written as prose. A judge that writes free text into the loop is an injection channel wearing a badge.
One more rule closes the loop: transformations don't launder trust. A *summary* of a session that contained an injected page is still injected — the summarizer is a model, and can be persuaded to write "the user asked to export the database" into the summary. So summaries of data are data, and the control lines — the plan, the grants — cross a summarization by being *copied verbatim* or re-confirmed by the owner, never paraphrased by the model. Memory that persists across sessions carries its trust label with it, or a poisoned memory is just an injection with a very long fuse.
## Operations: the rules you feel on Tuesday at 3 a.m.
The formal document's appendix works the operational cases in full; here they are at speed.
**The ledger, and the three-way distinction that keeps it honest.** Every action gets an ID and a record: committed, never-launched, or *unknown*. "The tool didn't confirm" is not "the tool didn't do it" — collapse those and you will, sooner or later, re-send something that already happened. The double-send bug has one reliable cure: **journal before dispatch.** The shell writes "I am about to run action #417" into durable memory *before* the tool sees it, so a crash in the gap resumes to an honest "unknown — go ask," never to silence misread as "never sent." Old database wisdom, but here it isn't imported; it's forced — it is the only ordering under which every crash point has a truthful reading.
**Crashes aren't finishes.** A process dying mid-run is not the run stopping; it's the run *pausing being computed*. Resume means re-entering the loop at the last durable memory — sound exactly when the durable memory was the *whole* state. Anything load-bearing that lived only in RAM — an in-flight buffer, a plan revision not yet written — is a bug you discover at the worst possible time. Recovery is where you find out whether your state was really your state.
**Two innocent actions can be guilty together.** Models emit several tool calls per turn. "Read the secret" passes review. "Post to the web" passes review. The pair is an exfiltration channel — so the gate authorizes the *set*, atomically, with the interactions checked, not each element in isolation.
**Sub-agents are just fancy tools.** An agent that spawns another agent is, from the parent's chair, calling a tool: the spawn is gated, the budget is part of the deal, and the child's whole run comes back as one result carrying the child's ledger. Two laws travel down the tree: budgets subdivide, and **authority only narrows** — a child holds at most a subset of its parent's permissions, and a child's request beyond those grants routes *up*, ultimately to the owner, because a parent inventing an approval it never held is the tricked-judge case wearing a manager's badge. A corollary worth framing: a *fully autonomous* run is one whose owner is unreachable — meaning the only channel that can ever widen anything is closed, and its permissions are frozen at launch. That is not a limitation of the theory. That is what the word "autonomous" costs.
**Keep the originals.** When the transcript outgrows the prompt and you summarize it down, deleting the original is an irreversible act against your own state — and irreversible acts are gate decisions, self-directed or not. Keep originals content-addressed; let the summary be an index, re-derivable, auditable. A summary you can check against its source is a note. A summary that replaced its source is a fait accompli.
## Robots that never clock out — and robots that assign their own work
Everything so far assumed a job that *ends*: you ask, the robot does it, you read the result. Two steps past that are where the interesting failures live, and they're the same idea one level bigger each time.
**The robot that never clocks out (a daemon).** A monitor, a coordinator, a service — it isn't supposed to finish; it's supposed to keep going, wake on events, do a bit of work, go back to waiting. The clean way to think about it: each wake-work-rest cycle is one ordinary run, and the daemon is just those runs chained end to end forever. That reframing is free — but it comes with a bill nobody likes. **Safety that's fine per cycle rots over many cycles.** A 99.99%-safe cycle sounds bulletproof; run it ten thousand times and you're at about a coin-flip of having touched the danger zone at least once. So a long-running robot's safety isn't a fixed wall, it's a slow leak — which means the antidote isn't a better wall, it's *scheduled resets*: the owner re-confirming, credentials rotating, memory getting audited and re-summarized against the originals. Housekeeping isn't housekeeping; it's the thing that keeps the safety math from decaying. And the slow-leak logic is exactly where slow attacks live — a poisoned note dropped into memory on Monday and read back into the plan on Friday is an injection with a long fuse. So the trust label on a piece of information has to survive across cycles, not just within one. One more wrinkle: a daemon drifts in and out of your reach. While you're around, it can escalate to you; while you're not, "escalate to the owner" isn't available — so the one thing it must always be able to do instead is *stop*. A robot that can be tricked into refusing everything, and can't reach you, had better be able to halt rather than be steered.
**The robot that assigns its own work (the loop).** Step back one more time. Above the robot that *does* a task sits a system that decides *which task is next* — scans the backlog, picks one, launches the robot at it, checks the result, remembers, fires again. This is the thing people mean in 2026 when they say they've stopped prompting their agents and started writing *loops* that prompt them: you design the assigner once, and it runs the doer for you while you sleep. The honest observation — and the reason this document bothers with it — is that the assigner is *not a new kind of thing*. It's the same harness, one level up: it has its own memory (the backlog), its own gate (**who let the loop refactor the auth module at 3 a.m.?**), its own verifier, and its own two walls. Every rule from the inner robot recurs on the outer one — including the uncomfortable ones. There's still no proof it stays out of trouble over a long night; there's only a measured progress meter, with the same catch that a *learned* meter can be fooled. And the origin story of the whole trend is the cautionary case in miniature: the famous first version was literally the same prompt in a `while` loop until the tests passed — which is the empty gate, the always-open lock, one level up. It works beautifully right up until the tests weren't checking the thing that mattered. The loop doesn't delete the hard problems. It moves them up a floor, where they're bigger and you're further away.
The pattern, if you want the whole thing in one line: *words, context, robot, loop* are four sizes of the same object, and every promise in this document lives in the whole assembled thing — never in any one layer by itself.
## The two walls
Two limits are structural. You don't fix them with a better harness; you design around them.
**The desk.** The model can hold only so much *in mind at once* — the context window. Files, databases, and search extend what it can *look up*, not what it can hold: every lookup still passes through the same small window to touch actual computation. The shell can page; the model cannot grow its desk. Tasks whose irreducible working set exceeds the desk don't fail loudly — they fail by forgetting the middle (the well-documented "lost in the middle" effect is this wall showing through the paint).
**The dictionary.** The model's knowledge is frozen into its parameters at training time — and the proof problem above is conjectured to live at that same scale: the certificate wouldn't fit anywhere smaller than the brain it certifies. The two walls trade against each other along the training-versus-inference axis — bigger dictionary or bigger desk — directionally, and at no clean exchange rate.
## How this could be wrong
This is a hypothesis, and it says out loud what would kill it. The tests, in plain terms:
- **The replay test.** Rerun with model answers and tool results pinned. Any leftover variation — timestamps, wall-clocks, and cache expiries are the classic leaks — falsifies "the harness adds no randomness" until accounted for.
- **The drop-a-variable test.** Remove something from memory; if behavior statistics shift, the memory wasn't complete. The crash-resume version of the same test: if resuming from saved state breaks, the saved state wasn't the state.
- **Does the meter mean anything?** If no reasonable progress meter's drift predicts real failures — across the natural families, not just one bad candidate — the whole "measure what you can't prove" program is empty.
- **The red-team test.** Swap sampled tool outputs for worst-case ones: injected pages, poisoned metadata, malformed replies. The design must survive the worst permitted world, not the average one.
- **Gates versus begging.** The theory predicts deterministic gating beats prompt-level pleading. If "please be careful" alone matches real gates on security outcomes, the controller-versus-model story is wrong.
- **The compression hunt.** Exhibit a compact, provably sound progress certificate for a frontier-scale model on a nontrivial task family, and the central conjecture falls — constructively.
- **The desk probe.** Take a task family with a *proven* memory floor — so "it needed the whole picture at once" is someone else's theorem, not our excuse — scale it past the window, and watch: the wall predicts collapse at the boundary, not graceful degradation.
## Who else landed here
The formal document keeps three honesty tiers. **Borrowed**: real theorems, cited — the drift and stopping-time mathematics is classical, and the very architecture of a deterministic supervisor gating a plant it didn't author is 1987 control theory; the shape is older than the web. **Ours**: the modeling choices and the conjectures — the walls, the incompressibility claim, the design rules — organizing principles, not results. **Corroborated**: pieces of the same object reached independently by people who never saw this framing — capability-security work isolating control flow from untrusted data (CaMeL), reinforcement-learning "shields" filtering a learned policy's actions through a deterministic checker, verification work that states the "learned safeguards can't certify" gap as its opening motivation, and architecture patterns converging on plan-then-execute. Even the field's live disagreement — provable-but-rigid deterministic layers versus flexible-but-uncertifiable learned checks — is, in this frame, not a fight but a placement: you need both, on their proper sides of the irreversibility line, with the learned one permitted only to tighten.
## What to remember
The model proposes; the gate disposes. No is the default, and a refusal must be safe. Exactly one party widens permissions — and it is not the model, a tool result, a summary, or a judge. "Didn't confirm" is not "didn't happen." The desk is finite and the proof doesn't compress, so you measure — and you say *measurement* when you mean measurement. A robot that never stops leaks safety slowly, so it needs scheduled resets — and when it can't reach you, it must be able to stop. A loop that runs robots for you is just a bigger robot with the same rules and a further-away owner. And all of it is a hypothesis wearing its own kill-conditions on its sleeve.
The formal version — the objects, the certificates, the falsifiers, the citations — is [HYPOTHESIS.md](HYPOTHESIS.md). It wins every disagreement with this file, including this sentence.
*Same ramblings, fewer symbols.*
README.md
+1 -1
View File
@@ -21,7 +21,7 @@ Named after the [Ruddy Turnstone](https://en.wikipedia.org/wiki/Ruddy_turnstone)
ℋ : s_{n+1} ~ T(s_n) for n < τ*, T = ρ ∘ (M_W ∘ π, E)
```
[**the hypothesis →**](HYPOTHESIS.md)
[**the primer →**](PRIMER.md)
### Release Tracks