# Information that keeps the lease: observed semantic information for software agents under a truth-tracking contract

*Working title:* **Information that keeps the lease.** Of everything a coding agent knows about its repository, how much does it need in order to stay there? We take a measure built for living things, point it at a seat in a GitHub repository, and say exactly what would have to be scrambled to find out.

Author: [withheld in the repository copy]. Contact: [withheld].
Status: **DRAFT — not submitted. Luke submits by hand, and only if he chooses to.**
Slate: candidate thirteenth paper from the landscape map. Research note: `research_notes/Information to quantum belief landscape/semantic_information_deployed_agents.md`. Companions: a withheld short paper under double-blind review and "Adapters never promote" (P12) supply the contract; this paper supplies a measure over it.

## Venues

| Venue | Deadline (time zone) | URL | Format / review | Fee or barrier | Checked |
|---|---|---|---|---|---|
| ALIFE 2027, Prague, 2027-07-19→23 | not posted; 2026 pattern: full papers 30 March → 12 April AoE (extended), notification 7 June, camera-ready 21 June | https://alife.vscht.cz/news/260821announcement (no 2027 site yet); 2026 CFP https://2026.alife.org/call-for-papers/ | full paper 3–8 pp excl. refs; summary 2 pp; single-blind ("Submissions are NOT anonymous"); MIT Press open-access proceedings | no publication fee listed; registration (in person or online) required per accepted paper; ISAL membership bundled; amounts not on the page | READ 2026-11-02 (announcement, 2026 CFP and registration); 2027 dates PATTERN |
| Artificial Life (MIT Press, ISAL journal) | rolling | https://alife.org/publications/ ; https://direct.mit.edu/artl (403) | quarterly; OA optional after acceptance | OA fee UNVERIFIED; subscription route presumed free | ISAL page READ 2026-11-02; MIT Press and Sherpa 403 two days running |
| Journal of Physics: Complexity (IOP) | rolling | https://publishingsupport.iopscience.iop.org/?p=3528 | single- or double-anonymous at author's choice | APC £2,110 / €2,505 / $2,900 ex-VAT; no submission charge; 25% IOP-member discount | READ 2026-11-02 |
| PRX Life (APS) | rolling | https://journals.aps.org/prxlife/about ; https://journals.aps.org/authors/open-access-physical-review-physical-review-letters | CC-BY; "Collaborative Review Process"; scope lists "Neural networks, and artificial intelligence" | APC $3,105 (2026 table); waivers only for lower-income countries | READ 2026-11-02 |
| Entropy (MDPI) | rolling | https://doaj.org/api/search/journals/issn:1099-4300 (mdpi.com 403) | single-anonymous; CC BY | APC CHF 2,600; waiver policy exists, terms unread | DOAJ record (2026-09-17) READ 2026-11-02 |
| Interface Focus (Royal Society) | theme issues only | https://royalsociety.org/journals/authors/which-journal/ | "Submission of individual articles is not considered" | propose a theme issue to publishing@royalsociety.org; S2O no-APC in 2026 per the venues note; hybrid fallback £1,995 | READ 2026-11-02; S2O and APC from the venues note 2026-11-01 |
| arXiv cs.AI, cross-list cs.MA (alternative nlin.AO) | rolling | https://info.arxiv.org/help/endorsement.html | preprint | one `cs` endorsement covers cs.AI and cs.MA; nlin.AO sits in the separate `nlin` domain (inference from the page's "high-level areas" rule) | READ 2026-11-02 |

## Abstract (draft)

A coding agent that works in a shared repository reads the repository before it speaks or acts. It reads test results, branch state and issue state, then opens pull requests and makes claims. Some of what it reads it needs; the rest it could lose without consequence. Kolchinsky and Wolpert (2018) call the part it needs *semantic information*: the correlations between a system and its environment that are causally necessary for the system to keep existing. They measure it by scrambling those correlations and watching viability. We transpose that measure to a software agent, a "seat", under a published truth-tracking contract. The system is the seat's statements and actions. The environment is repository truth: referee exit codes, branch state, lease state. Viability is a governance event: at a horizon τ the seat's lease is kept, its pull request is not reverted, and the truth run on its merged commit is green. The intervention coarse-grains what the seat read of the repository and re-scores viability. We give two estimators, one a replay of the event ledger with the claim-to-state pairing scrambled, one a shadow run with the seat's view degraded, and say what each can and cannot measure. The dataset is seven weeks of one repository: 68 pull requests from at least five seat identities, 58 merged, 151 referee runs (121 green, 30 red), with an eight-day CI outage carried as dark. Every computed value in this paper is dark until the replay is run. We pre-empt the objection that semantic information was built for physical systems with free-energy viability. (236 words)

## Contributions

1. **The mapping.** Seat = system X, repository truth = environment Y, viability = lease kept ∧ not reverted ∧ referee green at τ. This is inside the framework's own licence: K–W say viability is "exogenously determined by the scientist analyzing the system" and use negative entropy only as a worked example; Sowinski et al. use expected lifetime; Lekić et al. use dose–response performance. What is lost is the thermodynamic multiplier, which K–W defer for observed information anyway.
2. **The intervention, stated exactly.** Following K–W Sec. V.2, the channel p(X_{t+1} | X_t, Y_t) becomes p(X_{t+1} | X_t, φ_s(Y_t)) with the repository's own dynamics untouched. φ_s at level s ∈ [0, 1] replaces a fraction s of the referee outcomes, branch states and lease states the seat read with the fog law's own `dark` value. s = 0 is identity; s = 1 is K–W's "fully scrambled" case, where transfer entropy vanishes. The viability-optimal intervention minimises summed transfer entropy subject to viability at τ unchanged (K–W Eq. 21); observed semantic information is what survives (Eq. 22).
3. **Two estimators and their honest reach.** E1, replay: code each seat statement by the repository state it asserts (the contract already tags statements `claim`, `wall` or `referee`), estimate transfer entropy Y → X with bias-corrected discrete estimators and surrogates, apply φ_s to Y, re-estimate. E1 yields the syntactic curve TE(s); viability under φ_s cannot be re-observed, so E1's viability curve is dark. E2, shadow: run a seat in shadow-first egress with its view degraded at level s; referee every shadow proposal on a throwaway branch; viability is referee green ∧ no illegal lifecycle transition ∧ no spine refusal at τ. E2 yields V(s) and a threshold s_c: Sowinski's design transposed.
4. **The contract as instrument.** The η ladder separates the seat's speech from the environment's speech, so the channel to scramble is already labelled. The fog law supplies the coarse-graining value (`dark`, never 0). The pinned replay property (event-log replay bit-matches the materialised table) makes E1 exact as a replay, whatever its statistical limits.
5. **A reportable negative.** If the merging seat merges its own green PRs, lease-keeping may depend on almost nothing the seat read, and semantic efficiency may sit near zero. That is a finding about the governance, not a failure of the measure, and the paper commits to reporting it.
6. **Comparison protocol.** Empowerment (Song, Gore & Kleiman-Weiner, ICML 2026) and pragmatic information (Weinberger) on the same ledger, so a seat with high empowerment and near-zero semantic information can be exhibited if it exists. Dark until run.

## Outline (≈5,500 words; ALIFE full paper, 8 pages)

1. **A seat reads before it speaks** (500). One owner, several seats, a public site. A seat reads referee results and branch state, then writes prose and opens PRs. The false-success literature shows prose can be decoupled from state (Advani 2026; Smyth et al. 2026). We ask the complementary question: of what the seat reads, how much does it need to keep its place?
2. **Semantic information in one page** (700). Plain words first: syntactic information is any correlation; semantic information is the correlation whose loss ends you. Stored (initial mutual information) versus observed (transfer entropy over time); the viability-optimal intervention; semantic efficiency. Sowinski's forager: noise on the sensor, expected lifetime, a threshold.
3. **The objection, and the licence** (500). "This was defined for physical systems with free-energy viability." K–W made viability the analyst's choice and ignored energetics; the programme has since used lifetime, control performance and dose–response. A seat's existence in a repository is its lease, and an abandoned strand or a reverted PR is the death event, a boundary that must be maintained in Beer's sense. What is given up: no temperature, no work, no Landauer bill.
4. **The mapping** (700). X_t: the seat's tier-tagged statements, pushes, PR actions, lease. Y_t: referee exit codes, `main` and `live` state, strand status, spine hash. V_τ(strand) = 1 iff lease kept ∧ PR not reverted ∧ truth run green at τ; V is the mean over strands. τ = 7 days, justified by the release cadence.
5. **The intervention** (800). φ_s defined; why s = 0 and s = 1 match K–W's two limits. E1 and E2 as above. Discrete transfer entropy with bias correction and permutation surrogates (Schreiber 2000; Bossomaier et al. 2016; Lizier 2014). The exact-equality constraint in Eq. 21 replaced by a tolerance; non-uniqueness of the optimal intervention acknowledged.
6. **The ledger** (600). 2026-09-15 → 2026-11-01: 68 PRs, 58 merged, by branch prefix; 151 truth runs, 121 green and 30 red; the outage 2026-10-23 → 10-31 carried as dark and excluded from τ windows it overlaps. What the ledger does not record: what a seat read, as opposed to what it said.
7. **Results** (400). Every number here is dark in this draft; the section is the protocol and the table shells: TE(s) for E1; V(s), s_c and semantic efficiency for E2; empowerment and pragmatic information on the same rows.
8. **Related work and the negative space** (800). As below.
9. **Threats, dark, what this is not** (500). As below.

## Related work and the negative space

**The measure.** Kolchinsky & Wolpert (2018), *Interface Focus* 8, 20180041: semantic information as causally necessary syntactic information; stored and observed variants; viability exogenous; energetics ignored; thermodynamics of observed information deferred. Sowinski et al. (2023), *PRX Life* 1, 023003: sensor-noise scrambling, lifetime viability, a semantic threshold. Sowinski, Frank & Ghoshal (2024), *Phys. Rev. Research* 6, 043188. Piñero et al. (2026), *Communications Physics* 9, 120. Bartlett et al. (2025), *PRX Life* 3, 037003. Solé et al. (2024), *Interface Focus* 14(5).

**Beyond biology.** Lekić et al. (2025), arXiv:2506.22137 / *IEEE TMBMC*: viability-based semantic information for a drug-delivery channel. Bartlett (2024), arXiv:2407.07335, and Sowinski et al. (2025), ALIFE 2025, DOI 10.1162/isal.a.871: Causal Leverage Density for "any living, non-living or technological system."

**Viability and regulation.** Beer (2004), *Artificial Life* 10(3). McShaffrey & Beer (2026), arXiv:2605.16753: viability portraits. Conant & Ashby (1970), *Int. J. Systems Science* 1(2). Virgo et al. (2025), ALIFE, arXiv:2508.06326: an observer may read a regulating agent as a belief-holder. Kagan et al. (2026), arXiv:2601.03498.

**Information quantities for LLM agents.** Song, Gore & Kleiman-Weiner (2026), ICML, arXiv:2509.22504: empowerment from multi-turn text traces, correlated with task success. Ellis et al. (2026), ICML / PMLR 306: assistants trained to raise the human's empowerment. Takahashi & Hayashi (2026), AGI-26, LNCS: empowerment per joule. Klyubin, Polani & Nehaniv (2005), ECAL, and Salge, Glackin & Polani (2014): the empowerment hubs. Reid, Hafez & Nazeri (2025), arXiv:2509.10423: mutual information as a health signal for deployed RL agents. Zhang et al. (2026), ICLR MemAgents workshop, arXiv:2602.07885: agent memory as an information bottleneck. Di Felice et al. (2025), arXiv:2505.11509: value of information in multi-scale feedback systems.

**Meaning as decision relevance.** Weinberger (2024, v8 2026), arXiv:2403.12324: pragmatic information as KL between prior and posterior over the decision variable; Weinberger (2026), arXiv:2608.28803. Polani (2009), *HFSP J.* 3. Rovelli (2016), arXiv:1611.02420.

**Estimation.** Schreiber (2000), *PRL* 85, 461. Bossomaier, Barnett, Harré & Lizier (2016), Springer. Lizier (2014), *Frontiers in Robotics and AI* 1. Pearl (2009), *Causality*, 2nd ed. Varley (2025), *Physics Reports* 1148.

**The contract and its neighbours.** Blackwell (1953) and Cover & Thomas: the basis of "adapters never promote" (P12). Advani (2026) and Smyth et al. (2026): false success and overclaiming. Stechly, Valmeekam & Kambhampati (2024): self-critique adds no evidence. Li, Zhang & Hassan (2025), Watanabe et al. (2025), Ehsani et al. (2026): agent PRs at population scale, CI at most a feature.

**The negative space.** No paper computes stored or observed semantic information for a software agent, an LLM agent, a repository, a dashboard or an organisation, and none uses a governance event as the viability function. The frontier has reached engineered physical channels (Lekić 2025) and declared itself open to technological systems (Causal Leverage Density), and information-theoretic quantities have been estimated from LLM-agent traces (empowerment, ICML 2026), so two adjacent claims are closed: "first outside biology" and "first for LLM agents". What remains is the viability conditioning under a deployment contract, with the channel already labelled by evidence kind and the death events already public. The arXiv listings fetched 2026-11-02 (viability ∧ "semantic information": six entries, newest June 2025; the 40 newest "semantic information" ∧ language-model papers: all NLP or 6G sense; empowerment ∧ LLM: none on deployed agents) confirm the 2026-11-01 scan of the 171-paper citation graph.

## Evidence each claim rests on

| Claim | Pointer | Tier |
|---|---|---|
| Three kinds of speech: observation / recommendation / commitment | the-broadcast `docs/papers/01-truth-tracking.md` §2 | PUBLISHED (v1.0, 2026-10-30) |
| η ladder `*_claim` 0.25 < `*_wall` 0.40 < `*_referee` 0.92; only the referee tier full-collapses a belief; an adapter may reshape a value, never raise its tier | §3 | PUBLISHED |
| Lifecycle `proposed → ready → leased → working → (blocked \| review) → completed \| abandoned`; illegal transitions rejected; `evidence_added` never changes status; replay of the event log must bit-match the materialised table (`graph_hash`) | §4; `s3/protocol/strand.py`; `s3/test_truth.py` (6 tests) | PUBLISHED + PINNED |
| Fog law and receipts; `revenue_usd` never typed | §5–6 | PUBLISHED |
| GitHub as reducer: issue = strand, one assignee = lease, merge/label/tag = commitment; `release.yml` fast-forward only from a green SHA with an annotated tag; verify polls `deployed_revision`; `test/workflows.test.js` refuses `continue-on-error` / `if: always()` that paint green | §7; `.github/workflows/release.yml`; AGENTS.md | PUBLISHED + PINNED |
| Spine referee: SHA-256 over the public projection recorded with an authority line or the build fails; dam referee enforces the shape of human-only issues | `scripts/spine.js`, `scripts/dam-check.js`, `dam.yml` | PINNED |
| 2026-09-15 → 2026-11-01: 68 PRs, 58 merged; the head seat's prefix 45/45, six other prefixes 13/23 between them (4/9, 1/4, 3/4, 3/3, 0/1, 2/2); 151 truth runs, 121 success / 30 failure; CI outage 2026-10-23 → 10-31 treated as DARK; hand releases only by recorded owner override with local checks named in each tag | GitHub REST, counted 2026-11-01 | MEASURED |
| Not separated: billing refusals among the 30 red runs; post-merge defect rate; any no-contract baseline; branch protection status: DARK, must be re-checked at submission time | — | DARK / threats to validity |
| "Nothing approves itself; anything may earn approval"; egress act rule with Beta posteriors per class; `egress.money` pinned at p = 0 | `docs/LAW.md`, `docs/law.spec.json` | PROPOSED (not ratified) |
| E2 evidence (η no-op, shadow-first egress, earned promotion, spec-gate re-run): needs public evidence (a public repro or measurement) before it can be claimed | — | DARK |

E2 is DARK until public evidence exists. The intended mechanism: a seat in shadow proposes under a degraded view, is refereed, and dispatches nothing.

## DARK (no number is typed)

- Every computed semantic-information value: transfer entropy Y → X, the curve TE(s), observed semantic information, semantic efficiency, the threshold s_c, the viability curve V(s).
- Whether E1's estimate is meaningful at n = 68 PRs and 151 runs; discrete estimators bias upward at this count, and the surrogate test may fail to reject.
- Whether a seat's reading of the repository is recoverable from the ledger at all. The ledger records what seats said and did, not what they read; E1 infers the channel from statements that reference state.
- The E2 shadow run: not designed in detail, not scheduled, not run.
- Empowerment and pragmatic information on the same ledger.
- How many of the 30 red runs were billing refusals; post-merge defect rate; any baseline without the contract.
- Whether the η ceilings are calibrated; they fix an order, not probabilities.

## Threats to validity

One repository, one owner, seven weeks. Seat identities are branch prefixes under one GitHub account. Branch protection status is DARK and must be re-checked at submission time; if a seat can merge its own green PRs, "lease kept" and "not reverted" are weakly tested, and semantic efficiency may read near zero for a governance reason rather than an informational one. The eight-day outage removes referee observations and shortens the series. Transfer entropy on short discrete series is biased; the paper uses bias correction and permutation surrogates and reports intervals or nothing. X is text and needs a coding scheme; the tier tags reduce but do not remove coder judgement. The horizon τ is a choice; sensitivity over τ is reported or marked dark. K–W's exact-equality constraint on viability becomes a tolerance in practice, and the optimal intervention may not be unique; both are limits named in the source paper. Three quantities share the letter η across the sources (Sowinski's noise, K–W's efficiency, the contract's ceilings); the paper renames two. The paper is itself a claim at tier 0.25 about a measure it has not yet computed.

## What this paper is not

Not thermodynamics: no temperature, no work, no Landauer bill; K–W ignore energetics and defer the thermodynamics of observed information, and so do we. Not a claim that a seat is alive, conscious, or an agent beyond K–W's operational sense. Not a benchmark of agent code quality; defect rates are dark. Not the adapters paper (P12) or the withheld short paper; it uses their contract as an instrument. Not the engine's architecture paper; the engine appears only for the shadow-first and earned-promotion mechanisms E2 runs on. Not a replication of Sowinski; a transposition of their design. No number from any earlier deployment the engine descends from; EVIDENCE.md §9 and the §1 rows used here contain none.

## Luke's steps, if he wants to submit

1. Watch for the ALIFE 2027 call. The 2026 pattern puts the full-paper deadline in late March to mid-April 2027, Anywhere on Earth, 8 pages, single-blind; the 2027 site and CFP do not exist yet, and https://alife.vscht.cz/news/260821announcement is the only official source today. No publication fee was listed in 2026; registration with ISAL membership bundled is required for the proceedings.
2. Decide E1 only or E1 + E2. E1 is a replay over the existing ledger: a coding pass over the 68 PRs plus a transfer-entropy script (JIDT or a Python equivalent). E2 needs one seat run in shadow with its view of referee outcomes and branch state degraded at three or four levels of s, each proposal refereed on a throwaway branch. E2 is the paper's result; E1 alone is a methods note.
3. Before any number is typed: separate billing refusals among the 30 red runs from the Actions history; count reverts and follow-up fixes to merged PRs; run `npm test`, `python3 s3/test_truth.py` and `npm run spec:check` so every PINNED row is current.
4. arXiv: primary **cs.AI**, cross-list **cs.MA**; one `cs` endorsement covers both and is the same endorsement P3 and P12 need. Open a recent cs.AI abstract page in the neighbourhood (Song et al., arXiv:2509.22504, or Virgo et al., arXiv:2508.06326), click "Which authors of this paper are endorsers?", and send the six-character code with the PDF. Choose **nlin.AO** only if the ALife framing wins; it is a separate `nlin` endorsement domain by the page's rule, an inference rather than a listed fact.
5. Journal route after the conference: Artificial Life (open https://direct.mit.edu/artl by hand for the OA fee; the subscription route should carry no charge) or Journal of Physics: Complexity (£2,110 / $2,900; no submission charge). PRX Life is $3,105 and biological in track record. Entropy is CHF 2,600 with a waiver policy unread. Interface Focus takes theme issues only.
6. Rename two of the three η's before anyone reads a draft.
7. Log each submission in the authority record when there is a receipt.