diff --git a/README.md b/README.md index f44d364..e422632 100644 --- a/README.md +++ b/README.md @@ -1,175 +1,101 @@ +

+ KairoTensor. Evidence changes. Decisions should respond. An abstract diagram of evidence paths branching through time. +

+ # KairoTensor -Kairotensor is a Rust framework for versioned evidence memory. Its first component is a local SQLite store that records which observations an agent used and checks whether a saved result is still eligible to read or resume after evidence changes. +**Evidence-grounded revision for agent memory.** A Rust and SQLite research prototype for deciding when an agent's stored work remains eligible after its declared evidence changes. -## AI x Epistemics sprint submission +[Read the paper](docs/submission/kairotensor-report.pdf) · [Inspect the evidence](docs/submission/EVIDENCE.md) · [Explore the contracts](docs/README.md) · [Run the examples](#run-it-locally) -Author: **Alejandro Castellanos**, **Independent Research**. +[![Contract CI](https://github.com/k2jac9/kairotensor/actions/workflows/check.yml/badge.svg?branch=main)](https://github.com/k2jac9/kairotensor/actions/workflows/check.yml) -The author confirms that no KairoTensor-specific ideas, code or experiments -predate the 13-15 November 2026 sprint. The implementation and reported experiments -were developed during the sprint. This is an author declaration, not a conclusion -inferred solely from Git timestamps. Third-party dependencies and cited prior work -remain identified in their manifests and the report. +**Alejandro Castellanos · Independent Research**
+AI × Epistemics Research Sprint, November 2026 · Trust Infrastructure -The [submission guide](docs/submission/README.md) maps the report's claims to raw -evidence and offline reproduction commands. It includes a real Qwen reuse example -and the failed frozen recovery evaluation; neither establishes general improvement -in agent accuracy. Measured source is pinned to -`20fab84a6baa525452191ecf80eeee339f0b5ac7`; submission documentation is additional. +> **Research question**
+> When evidence changes, can an agent retire conclusions that have lost their declared support, preserve work that remains eligible, and account for the cost of correction? -## Run the withdrawal example +## The moment a decision goes stale -With stable Rust, Cargo, and a C compiler for bundled SQLite, run these commands from the repository root. +A deployment agent reads a test report, judges the result, and chooses a release plan under a policy. Then the policy changes. -```sh -cargo run --example withdrawal -cargo run --example selective_revision -cargo run --example corrected_handoff -cargo run --example scoped_model_reuse -cargo test --all-targets -``` +The plan needs revision. The test-report judgment may still be reusable under an explicit contract. If the **report itself** is replaced, that judgment needs fresh computation. If support disappears, the appropriate outcome can be **unsupported**; withdrawal does not establish the opposite proposition. -The example counts 8 items on one shelf and 4 on another. It starts a second count, then withdraws the second shelf report. The previous total of 12 becomes unavailable for both current reads and resume. The late count still costs two additions. A fresh attempt over the same candidate list sees only the first shelf and publishes 8. The ledger records five additions across all three attempts. +KairoTensor records that distinction at the execution boundary. It captures ordered evidence occurrences, freezes the adapter request, checks publication and subsequent access, and retains usage when attempted work becomes stale or invalid. -## Bulk setup +

+ Five stages: evidence occurrences, admission in an actor and scope, frozen attempt and request, claim and settlement, and current eligibility. Evidence changes trigger new checks; incurred usage and unresolved holds remain recorded. +

-The Rust store provides three ordered batch operations. Each call commits in one SQLite transaction. An error in any item rolls back every write in that call, including index updates and usage records. +*Conceptual lifecycle, not a performance plot. Timed recovery has additional reserve, claim and runtime checks; a diagram or stored receipt grants no dispatch authority.* -| Operation | Use it when | -| --- | --- | -| `observe_and_admit_many(actor, scope, observations)` | Import new source/payload pairs and grant access to their occurrences. The returned IDs follow input order. Repeating the import creates new occurrences, even for identical bytes. | -| `admit_many(actor, scope, evidence_ids)` | Grant access to observations already in storage. Each effective grant advances the view epoch once. Repeated IDs and already-admitted IDs add no events. | -| `publish_support_many(actor, scope, rules)` | Register support rules and publish their initial verdicts together. Every declaration has its own accepted attempt, parent and charge. Repeating a branch name creates successive versions in input order. | +## What the system contributes -Evidence ingestion and rule publication are separate commit boundaries. A caller can reuse IDs from a successful import if rule registration fails. For an existing catalogue, bulk grants update witness counts immediately. Use `revise_support` for an affected head whose verdict needs a new publication. The selective revision example exercises bulk setup followed by an atomic withdrawal. +| Layer | Implemented behavior | Contract | +| --- | --- | --- | +| **Evidence and eligibility** | Distinct acquisition occurrences, scoped admissions, current-read checks and explanations without payloads. | [Current eligibility](docs/current-eligibility.md) | +| **Atomic correction** | Expected-epoch admission batches; changed support verdicts publish once against the final view. Unchanged verdicts retain their attempt and charge. | [Atomic correction](docs/atomic-correction.md) | +| **Explicit reuse** | Same-branch reuse of an accepted model response with matching versioned contract, ordered candidates/inputs, request bytes and admission history. | [Scoped model reuse](docs/scoped-model-reuse.md) | +| **Execution and accounting** | Frozen requests, token reservations, durable send claims and deadline checks. Potentially sent work retains its hold until reconciliation permits settlement. | [Runtime boundary](docs/recovery-runtime.md) | +| **Replay and policy** | Authorized decision export, offline recorded validation, and evaluated policy activation with revision checks. | [Decision replay](docs/decision-replay.md) · [Policy promotion](docs/policy-promotion.md) | -## Current eligibility +These are application contracts over **declared dependencies**. Eligibility does not establish truth, authenticate the source, discover omitted inputs, or reveal what a model internally attended to. The embedding runtime authenticates actors and diagnostic access; the caller validates outputs and reports usage. A historical accepted receipt does not grant current access. -`Store::explain_current(actor, scope, branch)` explains why a head is ready, -empty, or needs review without loading its payload or changing historical -receipts and usage. It shares the authoritative gate with current reads and -resume. [Reason codes and snapshot contract](docs/current-eligibility.md) +## What we measured -## Selective revision +The paper reports the frozen source [`20fab84`](https://github.com/k2jac9/kairotensor/tree/20fab84a6baa525452191ecf80eeee339f0b5ac7). Later reliability features are integrated into the current repository and remain separate from those measurements. -The second example asks whether a test run has a reproduction record. An original report **and** its input manifest qualify together. A self-contained record from a second run also qualifies on its own. Withdrawing the original report preserves the accepted reproduction result through the second run. In the same transaction, the separate check that requires the original report publishes `unsupported`. +| Experiment | Observation | Interpretation and limit | +| --- | --- | --- | +| **Atomic revision** | **13.46×** median paired edit-time speedup for 32 dense edits across 128 rules, versus per-branch revision; six paired trials. | The lighter full scan had the lowest **total** time in all 30 matched trials. History contracts differ. [Protocol and samples](docs/revision-benchmark.md) | +| **One model correction** | **193 vs 404 tokens**, one new call versus two; both routes produced the same checked plan. Replacing the report forced fresh judgment. | One synthetic observation per route after a prompt revision. The earlier prompt produced a rejected error. No estimate of typical savings. [Prompts and traces](docs/scoped-model-reuse.md) | +| **Frozen recovery** | **0/6 held-out successes** for each of three policies. All responses violated the strict JSON contract; all retry receipts were late. | Policies replayed shared trajectories. These are failures of this model/validation/deadline arrangement, not independent policy arms. [Recorded evaluation](docs/evaluations/recovery-policy-v2-2026-11-14.md) | -| Operation | Contract | -| --- | --- | -| `publish_support(actor, scope, branch, alternatives)` | Publish a fixed rule over observation IDs. Every ID in one group must be admitted. Any complete group suffices. | -| `set_admitted_and_revise(actor, scope, evidence, admitted)` | Apply one membership edit and publish changed support verdicts in one transaction. Return the changed branch names, accepted attempt IDs, and verdicts. | -| `support_consumers(actor, scope, evidence)` | List current support-rule branches that name this occurrence in this actor's scope. A listed branch may retain another witness. | -| `revise_support(actor, scope, branch)` | Return the same accepted attempt ID if its verdict still holds. Publish a new attempt if the verdict changed. | - -The built-in rule checks **availability of declared witnesses**. The caller defines which observations qualify for the question. `unsupported` means every declared alternative lacks at least one admitted observation. Missing and withheld IDs have the same effect. The store never needs the payload bytes to answer this rule. - -SQLite updates the affected clause counts in the same transaction as admission or withdrawal. Reads and resumes check the verdict against the current view. In-flight opaque work still goes through the original epoch check. Opening a v1 through v9 database migrates it to schema v10. Existing heads and usage records retain their identities. - -Use `set_admitted_and_revise` when readers should see the membership update and all affected support heads in the same database snapshot. Any failure rolls back the entire edit. A witness that keeps the same verdict retains its head and cost. Repeating the desired membership while it is already current returns an empty change list. An intervening reversal makes a subsequent call a new edit. The original `set_admitted` and `revise_support` remain available for deferred revision. - -The store charges one unit per premise checked during rule registration and one for a changed verdict published by revision. Reusing a head adds no delivery or charge. These logical units exclude index maintenance and read checks. The integration suite compares the incremental answers with fresh databases and an independent evaluator across every admission subset of four observations. - -## Corrected handoff - -The deployment planner example captures exact dispatch bytes and their source IDs, retires an accepted staging continuation, then asks for a canary plan. A clean rebuild provides a second answer against the same release checklist. Run with the deterministic adapter or the pinned local CPU model. [Example, commands and evaluation](docs/corrected-handoff.md) - -## Reusing a model result - -A caller may reuse a previously validated model response for a closed, versioned computation on a branch. For example, a deployment policy can change while an independent test-report judgment remains eligible. In a live Qwen fixture, repairing the plan used one call and 193 tokens; a clean rebuild used two calls and 404 tokens. Replacing the report forced a fresh judgment and the release was held. [Contract and live evaluation](docs/scoped-model-reuse.md) - -The [live evaluation prompt catalogue](docs/evaluation-prompts.md) includes the original and revised instructions, exact user payloads, checkpoint history, and the request used at each step. - -## Token preflight and budgets - -The local Qwen worker can tokenize a frozen chat before generation and return its exact prompt count. Kairotensor can reserve that count plus the enforced output cap in a token budget. An affordable request uses the prepared inputs; an unaffordable request can discard them without running the model. On delivery, the ledger charges actual prompt and completion tokens and releases the unused reservation. Unresolved dispatched work retains its reservation. - -The `token_cap_eval` example compares four output caps on synthetic reports and replays the archived 403/404-token boundary. [Experiment protocol](docs/token-cap-protocol.md) · [Results and reproduction commands](docs/token-cap-results.md) - -## Evidence-backed deterministic verdict - -The [deterministic verdict route](docs/deterministic-verdict.md) derives a complete -CI verdict from one admitted report without a model call. It refuses incomplete -candidate manifests and checks evidence freshness again at publication. Historical -commit receipts are separate from current read eligibility. Run -`cargo run --example deterministic_verdict` for the zero-model-call demonstration. - -## Frozen semantic workload (offline) - -The [incident-routing smoke workload](docs/incident-routing-workload.md) prepares -versioned evidence-grounded routing requests and audits recorded responses against -a strict frozen contract. Its 12 authored cases separate calibration from heldout -scoring. Preparation performs no inference; synthetic fixture scores are not model -performance measurements, and declared evidence eligibility is not Store authentication. -The [prepared-request bridge](docs/incident-prepared-bridge.md) connects these -prompts to the existing worker protocol with offline lifecycle tests and emits a -bounded, unexecuted calibration proposal. It does not grant Store send authority. -The [offline calibration controller](docs/incident-calibration-controller.md) -tests durable send claims, retained uncertain usage and independent process -termination against fixed fake workers. It has no live model launch path. -The [versioned Transformers backend](docs/incident-transformers-backend.md) -connects the bounded controller to pinned local-model APIs, with execution denied -by default. Its automated checks use fake dependencies only; real calibration -still requires provisioned assets and separately approved execution. -The [proposed GitHub/NixOS comparison](docs/calibration-host-matrix.md) lists -matching requirements, readiness gates and unapproved per-host bounds. -The [offline Adaption adapter](docs/incident-adaption-backend.md) exercises a -distinct hosted-response contract and synthetic accounting, without credentials, -network requests, or live execution authority. -The [one-request hosted runner](docs/incident-hosted-runner.md) adds fixed-origin -HTTPS code and a bounded child process, tested only with mocked HTTP and canned -receipts. Live use requires separate approval and a verified billing bound. - -## Implemented contracts - -The [durable recovery APIs and example](docs/budgeted-recovery.md) connect -validation failure to one retry or abstention. Recovery IDs, retry slots, -reservations and single-use send claims survive restart and serialize competing -workers. Only proven-unsent work can be cancelled. Run -`cargo run --example budgeted_recovery` for the deterministic fixture, or add -`--live` for the pinned Qwen worker. The fixture is not a live performance result. - -The opt-in [deadline-aware recovery gate](docs/deadline-aware-recovery.md) binds -the original clock domain and deadline to a runtime/calibration profile, checks -remaining time after acquiring the writer lock, and refuses uncovered requests. -Its offline example uses synthetic timing, not measured inference performance: -`cargo run --example deadline_aware_recovery`. The [runtime send boundary](docs/recovery-runtime.md) -adds a concrete Linux boot clock and a final post-claim check before one backend -invocation, without treating uncertain sends as free or retryable. The -[offline end-to-end runner](docs/timed-recovery-runner.md) joins this path to the -existing fixture Adapter with durable audit files and fault scenarios; live mode -is explicitly refused. - -The [recovery-policy evaluation protocol](docs/recovery-policy-protocol.md) defines -an offline paired comparison with explicit accounting and deadline checks. Run -`python3 scripts/recovery_policy_eval.py --fixture` for synthetic contract cases, -not model-performance measurements. The [Store-backed collector](docs/recovery-policy-runner.md) -adds a versioned corpus, raw traces and a durable ledger; live collection is opt-in. - -| Operation | Behavior | -| --- | --- | -| `observe`, `set_admitted` | Record a distinct observation occurrence and change visibility for one actor and scope. Repeating an admission leaves the view epoch unchanged. | -| `begin_from_candidates` | Select admitted inputs from an ordered list and freeze their bytes, the view epoch and the expected branch head in one transaction. This also covers a selection that returns zero inputs. | -| `begin_attempt` | Require every explicit input to be admitted. Record the ordered IDs and start fresh, with no inherited output. | -| `read_current`, `begin_resume` | Check the accepted head against the current view before returning bytes. A resume records which accepted output it inherited. | -| `record_dispatch` | Freeze the exact serialized adapter request for a still-eligible pending attempt. The request stays linked to the ordered source occurrences. | -| `create_token_budget`, `token_budget_state` | Set a ceiling in input-plus-output tokens and read the amount spent and still reserved. | -| `reserve_dispatch` | Atomically reserve the measured input plus the configured output cap and authorize dispatch. A denial records the decision and consumes no capacity. | -| `create_recovery`, `begin_recovery_attempt` | Bind a stable operation key and allocate at most the original attempt plus one retry after invalid output. | -| `reserve_recovery_attempt`, `recovery_reservation` | Stage exact request/quote and read an unsent reservation for verified re-preparation. | -| `claim_recovery_dispatch`, `cancel_recovery_attempt` | Serialize send permission against cancellation; only one claimant can send, and only unclaimed work releases capacity. | -| `settle` | Validate the captured branch and view, record the delivery and usage, and advance the head only on acceptance. Identical retries return the first disposition. Conflicting retries fail. | -| `reject_invalid` | Record a dispatched response that failed validation, including the versioned validator, reason, exact returned bytes, and final usage. Keep the current branch head. | +A [later controlled correction pilot](docs/controlled-correction-results.md) reports four synthetic changes: post-change repair used four calls and 751 reported tokens; fresh recomputation used six calls and 1,167 tokens. It uses a different model/runtime, is outside the paper, and does not establish population accuracy or latency savings. + +**The evidence supports a working mechanism, measured implementation tradeoffs, and bounded examples of reuse. It does not establish generally more accurate or trustworthy agent decisions.** + +## Run it locally -The caller authenticates actors for every operation. For opaque computations the caller also validates outputs and measures usage. Settlement distinguishes acceptance, stale views, stale branch heads and validator rejection. A rejected answer also records whether its captured view and branch were current at settlement. Each branch belongs to one actor and scope. An effective admission or revocation in that scope conservatively invalidates an opaque head. A fresh computation can reuse raw evidence that is still admitted. Withdrawal governs future reads from this store. Previously delivered bytes may still exist in a caller. +Use the pinned Rust toolchain in `rust-toolchain.toml`, Cargo, and a C compiler for bundled SQLite. These examples use deterministic fixtures and make no model or provider calls: + +```sh +cargo run --locked --example withdrawal +cargo run --locked --example selective_revision +cargo run --locked --example incident_store_bridge +``` -## Reference and next milestone +In `withdrawal`, two shelf reports initially support a total of 12. Withdrawing the second report makes the old total unavailable for current reads and resume. A fresh attempt returns 8. Work that finished too late remains in the ledger: five additions across the three attempts. -[The v0 design](docs/memory-v0.md) describes the larger research program. The Python reference under `reference/` covers additional proposed behavior, including alternative derivations and exact categorical belief revision. Its original fixtures run with the Python standard library. +For the full local contract suite and the independent reference/oracle checks: ```sh +cargo test --locked --all-targets python3 -m unittest discover -s reference -p test_memory.py python3 reference/belief_oracle.py ``` -On the three-way benchmark, atomic revision reduced median edit time by 13.46× for 32 dense edits on 128 rules, compared with the per-branch API. For 128 sparse edits on 1,024 rules, the corresponding gain was 1.40×. The batched full scan still had the lowest total time in all 30 matched trials. On a subsequent matched run with 1,024 rules, bulk setup took a median 214 ms versus 3,690 ms for single-item setup. Both production paths retained per-occurrence admission events, branch lineage and logical charges. [Benchmark, measurements, and reproduction commands](docs/revision-benchmark.md) +The [CI workflow](.github/workflows/check.yml) also runs the Python scripts and fixture integration stages across pinned stable, upcoming and previous runtime lanes. The badge links to current results; fixture passes are not live model-performance evidence. On this owner's workstation, use the [qualified launcher and storage workflow](docs/runtime-adoption.md) before heavy builds. + +## Why this is an epistemics problem + +An answer can remain in memory after its justification changes. Durable provenance helps identify the supplied evidence; revision rules determine which saved work may still be used. KairoTensor studies their integration with exact model requests, conservative continuation checks and accounting for failed or superseded work. + +This builds on established ideas in [truth maintenance](https://doi.org/10.1016/0004-3702(79)90008-0) and [assumption-based truth maintenance](https://doi.org/10.1016/0004-3702(86)90080-9). The contribution under study is an execution-boundary integration, not a new truth-maintenance formalism. See the paper's [related work and limitations](docs/submission/report.md#2-related-work). + +## Research map + +| If you want to… | Start here | +| --- | --- | +| Review the sprint submission | [Paper](docs/submission/kairotensor-report.pdf) · [Editable manuscript](docs/submission/report.md) · [Reviewer guide](docs/submission/README.md) | +| Trace a claim to raw evidence | [Evidence index](docs/submission/EVIDENCE.md) · [Benchmark CSVs](docs/benchmarks/) · [Evaluation records](docs/evaluations/) | +| Understand or embed the Store | [Contract guide](docs/store-guide.md) · [Documentation index](docs/README.md) | +| Inspect the semantic integration | [Store-backed incident bridge](docs/incident-store-bridge.md) · [Frozen workload](docs/incident-routing-workload.md) | +| Audit later reliability work | [Review guide](docs/reliability-review-guide.md) · [Dated submission status](docs/submission/CURRENT_STATUS.md) | + +Next research steps are controlled comparisons after evidence changes, equally durable baselines, broader task families, and independently measured model execution. Hosted adapters and AutoScientist work do not become paper results merely by existing or finishing. + +## Authorship and disclosure + +Alejandro Castellanos confirms that no KairoTensor-specific ideas, code or experiments predate the 13–15 November 2026 sprint. This is the author's declaration; timestamps alone cannot establish when work began. Third-party dependencies and prior work remain identified in their manifests and citations. AI assistance supported development and writing; the author is responsible for the claims and submission. See the paper's [provenance, ethics and LLM usage statements](docs/submission/report.md#appendix-limitations-and-dual-use--ethical-considerations). diff --git a/docs/README.md b/docs/README.md new file mode 100644 index 0000000..fb3890b --- /dev/null +++ b/docs/README.md @@ -0,0 +1,38 @@ +# A field guide to KairoTensor + +[Project overview](../README.md) · [Paper](submission/kairotensor-report.pdf) · [Evidence index](submission/EVIDENCE.md) + +The project studies a practical epistemic boundary: when a stored conclusion may still be used after its declared evidence changes. Read the paper for the measured research question; use the contract pages to understand the current implementation. + +## Follow the evidence + +| Stage | Question | Read | +| --- | --- | --- | +| **01 · Capture** | Which occurrences and exact bytes did this attempt receive? | [Store and execution guide](store-guide.md) · [Corrected handoff](corrected-handoff.md) | +| **02 · Qualify** | Is the head eligible now, and why? | [Current eligibility](current-eligibility.md) · [Atomic correction](atomic-correction.md) | +| **03 · Execute** | Can this request be sent within the declared resource and timing contract? | [Token preflight](token-cap-protocol.md) · [Budgeted recovery](budgeted-recovery.md) · [Runtime boundary](recovery-runtime.md) | +| **04 · Reuse or replay** | Can a prior result be reused, or a recorded request reconstructed? | [Scoped reuse](scoped-model-reuse.md) · [Decision replay](decision-replay.md) | +| **05 · Evaluate** | What changed in the decision, cost and failure behavior? | [Revision benchmark](revision-benchmark.md) · [Controlled correction pilot](controlled-correction-results.md) · [Policy promotion](policy-promotion.md) | + +## Three kinds of evidence + +**Measured research.** The sprint paper freezes source `20fab84`. Its [evidence index](submission/EVIDENCE.md) maps every principal result to samples, prompts or archived responses. The [later correction pilot](controlled-correction-pilot.md) has a separate source/model identity and is not silently added to the paper. + +**Implemented contracts.** Rust tests and offline fixtures cover eligibility, atomic correction, the [Store-backed incident bridge](incident-store-bridge.md), replay and policy activation. A passing contract test establishes the tested behavior, not semantic model accuracy. + +**Prepared experiments.** The [incident workload](incident-routing-workload.md), [Transformers backend](incident-transformers-backend.md), [Adaption contract](incident-adaption-backend.md) and [hosted runner](incident-hosted-runner.md) specify execution boundaries. Their offline tests do not establish live provider compatibility, billing reconciliation or comparative performance. Read each page's execution limits before running it. + +## For reviewers + +1. Read the [submission guide](submission/README.md) and the paper's contribution and limitations sections. +2. Follow a claim through the [evidence index](submission/EVIDENCE.md) to the pinned raw artifact. +3. Recompute archived summaries with the documented commands. Keep this distinct from re-executing or independently authenticating a model run. +4. Review [later reliability changes](reliability-review-guide.md) separately from the measured snapshot. + +## For implementers + +The [Store guide](store-guide.md) retains the detailed API and accounting contracts. [Deadline-aware recovery](deadline-aware-recovery.md), the [timed runner](timed-recovery-runner.md), [decision bundles](decision-replay.md), [policy transitions](policy-promotion.md) and [offline ablations](decision-ablation.md) describe distinct boundaries. None replaces current admission or grants an independent send permission. + +The [original v0 design](memory-v0.md) records research direction. Its historical implementation-status passages are not an inventory of today's code. The Python reference covers additional proposed semantics; do not assume every reference behavior exists in the Rust Store. + +For this workstation, consult the [runtime launcher](runtime-adoption.md) and [runtime qualification](runtime-adoption.md). For hosted checks, use the exact commit's [Actions results](https://github.com/k2jac9/kairotensor/actions), including integration stages and any reported skips. diff --git a/docs/assets/public-presentation/README.md b/docs/assets/public-presentation/README.md new file mode 100644 index 0000000..f840ba5 --- /dev/null +++ b/docs/assets/public-presentation/README.md @@ -0,0 +1,27 @@ +# Public presentation assets + +Original vector artwork for KairoTensor's repository landing page. The diagrams +are conceptual illustrations, not measurements, causal proofs, or evidence of +model execution. The lifecycle is deliberately simplified; linked contract pages +remain the authority for reserve/claim/settlement and reuse behavior. + +| Token | Color | Use | +| --- | --- | --- | +| Midnight | `#0B1023` | Diagram background | +| Paper | `#F4F6FC` | Main lettering | +| Mist | `#C1CBE0` | Supporting lettering | +| Cyan | `#66E4EA` | Evidence and eligibility | +| Violet | `#C4AFFB` | Captured attempts and change | +| Amber | `#F5CE7A` | Accounting | +| Print teal | `#12566A` | Paper heading accent on white | + +`hero.svg` and `evidence-lifecycle.svg` are editable sources with embedded titles +and descriptions, explicit view boxes, no scripts, no remote assets and no font +fetches. Labels accompany all color distinctions. README alternative text and +surrounding prose carry the same information for readers who cannot see the art. +GitHub embeds the SVGs as images; both include their own dark background for +consistent contrast in light and dark themes. DejaVu Sans/Arial are fallback fonts, +not bundled dependencies. No photographic or generated-image source is used. + +The report retains its academic template. Its restrained teal headings and pale +blue table headers share this palette without importing the dark cover artwork. diff --git a/docs/assets/public-presentation/evidence-lifecycle.svg b/docs/assets/public-presentation/evidence-lifecycle.svg new file mode 100644 index 0000000..febcc3e --- /dev/null +++ b/docs/assets/public-presentation/evidence-lifecycle.svg @@ -0,0 +1,18 @@ + + From evidence occurrence to current eligibility + Evidence occurrence, scoped admission, frozen attempt, claim and settlement, and current eligibility. A change in admitted evidence triggers current checks. Incurred usage and unresolved holds remain recorded. This conceptual sequence simplifies the full contracts. + + + + FOLLOW THE DECLARED EVIDENCE + + + 01 / OCCURRENCE02 / ADMISSION03 / ATTEMPT04 / EXECUTION05 / ELIGIBILITY + IdentifyQualifyFreezeAccountRecheck + Distinct event IDsExact payload bytesActor and scopeAdmit or withdrawOrdered inputsExact requestClaim and settleUsage and holdsCurrent read / resumeExplicit reuse gate + + Evidence changes trigger new checks. + + Incurred usage and unresolved holds remain recorded.Eligibility ≠ truth + + diff --git a/docs/assets/public-presentation/hero.svg b/docs/assets/public-presentation/hero.svg new file mode 100644 index 0000000..ea1d054 --- /dev/null +++ b/docs/assets/public-presentation/hero.svg @@ -0,0 +1,35 @@ + + KairoTensor — Evidence changes. Decisions should respond. + An original abstract illustration of evidence paths branching through time. Cyan, violet and amber paths converge around a bright node on a dark navy field. This is conceptual artwork, not measured data. + + + + + + + + + + + + + + + + + + + + + + + + AGENT MEMORY · EVIDENCE REVISION + KairoTensor + Evidence changes. + Decisions should respond. + + A research prototype for justified reuse and accountable revision. + RUST + SQLITE / LOCAL EVIDENCE / EXPLICIT CONTRACTS + + diff --git a/docs/store-guide.md b/docs/store-guide.md new file mode 100644 index 0000000..4524716 --- /dev/null +++ b/docs/store-guide.md @@ -0,0 +1,148 @@ +# Store and execution guide + +[Project overview](../README.md) · [Documentation map](README.md) + +Detailed contracts and examples for the current Rust/SQLite implementation. +Research measurements and their limits are summarized in the project overview. + +## Bulk setup + +The Rust store provides three ordered batch operations. Each call commits in one SQLite transaction. An error in any item rolls back every write in that call, including index updates and usage records. + +| Operation | Use it when | +| --- | --- | +| `observe_and_admit_many(actor, scope, observations)` | Import new source/payload pairs and grant access to their occurrences. The returned IDs follow input order. Repeating the import creates new occurrences, even for identical bytes. | +| `admit_many(actor, scope, evidence_ids)` | Grant access to observations already in storage. Each effective grant advances the view epoch once. Repeated IDs and already-admitted IDs add no events. | +| `publish_support_many(actor, scope, rules)` | Register support rules and publish their initial verdicts together. Every declaration has its own accepted attempt, parent and charge. Repeating a branch name creates successive versions in input order. | + +Evidence ingestion and rule publication are separate commit boundaries. A caller can reuse IDs from a successful import if rule registration fails. For an existing catalogue, bulk grants update witness counts immediately. Use `revise_support` for an affected head whose verdict needs a new publication. The selective revision example exercises bulk setup followed by an atomic withdrawal. + +## Current eligibility + +`Store::explain_current(actor, scope, branch)` explains why a head is ready, +empty, or needs review without loading its payload or changing historical +receipts and usage. It shares the authoritative gate with current reads and +resume. [Reason codes and snapshot contract](current-eligibility.md) + +## Selective revision + +The second example asks whether a test run has a reproduction record. An original report **and** its input manifest qualify together. A self-contained record from a second run also qualifies on its own. Withdrawing the original report preserves the accepted reproduction result through the second run. In the same transaction, the separate check that requires the original report publishes `unsupported`. + +| Operation | Contract | +| --- | --- | +| `publish_support(actor, scope, branch, alternatives)` | Publish a fixed rule over observation IDs. Every ID in one group must be admitted. Any complete group suffices. | +| `set_admitted_and_revise(actor, scope, evidence, admitted)` | Apply one membership edit and publish changed support verdicts in one transaction. Return the changed branch names, accepted attempt IDs, and verdicts. | +| `support_consumers(actor, scope, evidence)` | List current support-rule branches that name this occurrence in this actor's scope. A listed branch may retain another witness. | +| `revise_support(actor, scope, branch)` | Return the same accepted attempt ID if its verdict still holds. Publish a new attempt if the verdict changed. | + +The built-in rule checks **availability of declared witnesses**. The caller defines which observations qualify for the question. `unsupported` means every declared alternative lacks at least one admitted observation. Missing and withheld IDs have the same effect. The store never needs the payload bytes to answer this rule. + +SQLite updates the affected clause counts in the same transaction as admission or withdrawal. Reads and resumes check the verdict against the current view. In-flight opaque work still goes through the original epoch check. Opening a v1 through v9 database migrates it to schema v10. Existing heads and usage records retain their identities. + +Use `set_admitted_and_revise` when readers should see the membership update and all affected support heads in the same database snapshot. Any failure rolls back the entire edit. A witness that keeps the same verdict retains its head and cost. Repeating the desired membership while it is already current returns an empty change list. An intervening reversal makes a subsequent call a new edit. The original `set_admitted` and `revise_support` remain available for deferred revision. + +The store charges one unit per premise checked during rule registration and one for a changed verdict published by revision. Reusing a head adds no delivery or charge. These logical units exclude index maintenance and read checks. The integration suite compares the incremental answers with fresh databases and an independent evaluator across every admission subset of four observations. + +## Corrected handoff + +The deployment planner example captures exact dispatch bytes and their source IDs, retires an accepted staging continuation, then asks for a canary plan. A clean rebuild provides a second answer against the same release checklist. Run with the deterministic adapter or the pinned local CPU model. [Example, commands and evaluation](corrected-handoff.md) + +## Reusing a model result + +A caller may reuse a previously validated model response for a closed, versioned computation on a branch. For example, a deployment policy can change while an independent test-report judgment remains eligible. In a live Qwen fixture, repairing the plan used one call and 193 tokens; a clean rebuild used two calls and 404 tokens. Replacing the report forced a fresh judgment and the release was held. [Contract and live evaluation](scoped-model-reuse.md) + +The [live evaluation prompt catalogue](evaluation-prompts.md) includes the original and revised instructions, exact user payloads, checkpoint history, and the request used at each step. + +## Token preflight and budgets + +The local Qwen worker can tokenize a frozen chat before generation and return its exact prompt count. Kairotensor can reserve that count plus the enforced output cap in a token budget. An affordable request uses the prepared inputs; an unaffordable request can discard them without running the model. On delivery, the ledger charges actual prompt and completion tokens and releases the unused reservation. Unresolved dispatched work retains its reservation. + +The `token_cap_eval` example compares four output caps on synthetic reports and replays the archived 403/404-token boundary. [Experiment protocol](token-cap-protocol.md) · [Results and reproduction commands](token-cap-results.md) + +## Evidence-backed deterministic verdict + +The [deterministic verdict route](deterministic-verdict.md) derives a complete +CI verdict from one admitted report without a model call. It refuses incomplete +candidate manifests and checks evidence freshness again at publication. Historical +commit receipts are separate from current read eligibility. Run +`cargo run --example deterministic_verdict` for the zero-model-call demonstration. + +## Frozen semantic workload (offline) + +The [incident-routing smoke workload](incident-routing-workload.md) prepares +versioned evidence-grounded routing requests and audits recorded responses against +a strict frozen contract. Its 12 authored cases separate calibration from heldout +scoring. Preparation performs no inference; synthetic fixture scores are not model +performance measurements, and declared evidence eligibility is not Store authentication. +The [prepared-request bridge](incident-prepared-bridge.md) connects these +prompts to the existing worker protocol with offline lifecycle tests and emits a +bounded, unexecuted calibration proposal. It does not grant Store send authority. +The [offline calibration controller](incident-calibration-controller.md) +tests durable send claims, retained uncertain usage and independent process +termination against fixed fake workers. It has no live model launch path. +The [versioned Transformers backend](incident-transformers-backend.md) +connects the bounded controller to pinned local-model APIs, with execution denied +by default. Its automated checks use fake dependencies only; real calibration +still requires provisioned assets and separately approved execution. +The [proposed GitHub/NixOS comparison](calibration-host-matrix.md) lists +matching requirements, readiness gates and unapproved per-host bounds. +The [offline Adaption adapter](incident-adaption-backend.md) exercises a +distinct hosted-response contract and synthetic accounting, without credentials, +network requests, or live execution authority. +The [one-request hosted runner](incident-hosted-runner.md) adds fixed-origin +HTTPS code and a bounded child process, tested only with mocked HTTP and canned +receipts. Live use requires separate approval and a verified billing bound. + +## Implemented contracts + +The [durable recovery APIs and example](budgeted-recovery.md) connect +validation failure to one retry or abstention. Recovery IDs, retry slots, +reservations and single-use send claims survive restart and serialize competing +workers. Only proven-unsent work can be cancelled. Run +`cargo run --example budgeted_recovery` for the deterministic fixture, or add +`--live` for the pinned Qwen worker. The fixture is not a live performance result. + +The opt-in [deadline-aware recovery gate](deadline-aware-recovery.md) binds +the original clock domain and deadline to a runtime/calibration profile, checks +remaining time after acquiring the writer lock, and refuses uncovered requests. +Its offline example uses synthetic timing, not measured inference performance: +`cargo run --example deadline_aware_recovery`. The [runtime send boundary](recovery-runtime.md) +adds a concrete Linux boot clock and a final post-claim check before one backend +invocation, without treating uncertain sends as free or retryable. The +[offline end-to-end runner](timed-recovery-runner.md) joins this path to the +existing fixture Adapter with durable audit files and fault scenarios; live mode +is explicitly refused. + +The [recovery-policy evaluation protocol](recovery-policy-protocol.md) defines +an offline paired comparison with explicit accounting and deadline checks. Run +`python3 scripts/recovery_policy_eval.py --fixture` for synthetic contract cases, +not model-performance measurements. The [Store-backed collector](recovery-policy-runner.md) +adds a versioned corpus, raw traces and a durable ledger; live collection is opt-in. + +| Operation | Behavior | +| --- | --- | +| `observe`, `set_admitted` | Record a distinct observation occurrence and change visibility for one actor and scope. Repeating an admission leaves the view epoch unchanged. | +| `begin_from_candidates` | Select admitted inputs from an ordered list and freeze their bytes, the view epoch and the expected branch head in one transaction. This also covers a selection that returns zero inputs. | +| `begin_attempt` | Require every explicit input to be admitted. Record the ordered IDs and start fresh, with no inherited output. | +| `read_current`, `begin_resume` | Check the accepted head against the current view before returning bytes. A resume records which accepted output it inherited. | +| `record_dispatch` | Freeze the exact serialized adapter request for a still-eligible pending attempt. The request stays linked to the ordered source occurrences. | +| `create_token_budget`, `token_budget_state` | Set a ceiling in input-plus-output tokens and read the amount spent and still reserved. | +| `reserve_dispatch` | Atomically reserve the measured input plus the configured output cap and authorize dispatch. A denial records the decision and consumes no capacity. | +| `create_recovery`, `begin_recovery_attempt` | Bind a stable operation key and allocate at most the original attempt plus one retry after invalid output. | +| `reserve_recovery_attempt`, `recovery_reservation` | Stage exact request/quote and read an unsent reservation for verified re-preparation. | +| `claim_recovery_dispatch`, `cancel_recovery_attempt` | Serialize send permission against cancellation; only one claimant can send, and only unclaimed work releases capacity. | +| `settle` | Validate the captured branch and view, record the delivery and usage, and advance the head only on acceptance. Identical retries return the first disposition. Conflicting retries fail. | +| `reject_invalid` | Record a dispatched response that failed validation, including the versioned validator, reason, exact returned bytes, and final usage. Keep the current branch head. | + +The caller authenticates actors for every operation. For opaque computations the caller also validates outputs and measures usage. Settlement distinguishes acceptance, stale views, stale branch heads and validator rejection. A rejected answer also records whether its captured view and branch were current at settlement. Each branch belongs to one actor and scope. An effective admission or revocation in that scope conservatively invalidates an opaque head. A fresh computation can reuse raw evidence that is still admitted. Withdrawal governs future reads from this store. Previously delivered bytes may still exist in a caller. + +## Reference and next milestone + +[The v0 design](memory-v0.md) describes the larger research program. The Python reference under `reference/` covers additional proposed behavior, including alternative derivations and exact categorical belief revision. Its original fixtures run with the Python standard library. + +```sh +python3 -m unittest discover -s reference -p test_memory.py +python3 reference/belief_oracle.py +``` + +On the three-way benchmark, atomic revision reduced median edit time by 13.46× for 32 dense edits on 128 rules, compared with the per-branch API. For 128 sparse edits on 1,024 rules, the corresponding gain was 1.40×. The batched full scan still had the lowest total time in all 30 matched trials. On a subsequent matched run with 1,024 rules, bulk setup took a median 214 ms versus 3,690 ms for single-item setup. Both production paths retained per-occurrence admission events, branch lineage and logical charges. [Benchmark, measurements, and reproduction commands](revision-benchmark.md) diff --git a/docs/submission/CURRENT_STATUS.md b/docs/submission/CURRENT_STATUS.md new file mode 100644 index 0000000..068f1c0 --- /dev/null +++ b/docs/submission/CURRENT_STATUS.md @@ -0,0 +1,129 @@ +# KairoTensor submission handoff: 16 November 2026 + +This is a dated supplement to the frozen report, with live status observations +recorded at approximately 07:39 UTC after PR integration and branch cleanup. +The paper and PDF still report source +`20fab84a6baa525452191ecf80eeee339f0b5ac7`. The subsequent template review replaces +stale branch-status wording with an explicit measured-scope boundary, adds three +contribution bullets and a numbered results table, and links the reviewer package. +The abstract and original result paragraphs are unchanged. No result below is +silently added to the paper. + +The reference-inspired editorial successor adds a concrete opening, Figure 1, +finding labels and an appendix reproducibility table. The author provenance +declaration moves intact to that appendix. Consistent typography and tighter +spacing reduce the report to seven pages: five main, one references, one appendix. +Numbering is applied to the PDF export; the native editable Doc retains the same +content and diagram. See the new editorial verification receipt for exact hashes. + +## Integrated implementation + +PRs [20](https://github.com/k2jac9/kairotensor/pull/20), +[19](https://github.com/k2jac9/kairotensor/pull/19), +[18](https://github.com/k2jac9/kairotensor/pull/18) and +[21](https://github.com/k2jac9/kairotensor/pull/21) are merged. The integrated source +is `109798d776b93d6730b2d6bac74d0fd328c19815`. +[Main CI run 43821777107](https://github.com/k2jac9/kairotensor/actions/runs/43821777107) +passed all six jobs on that exact source. The five contract lanes are: + +| Lane | Rust | Python | +| --- | --- | --- | +| Stable | 1.100.0 | 3.15.2 | +| Upcoming Rust | nightly-2026-11-16 / 1.102.0-nightly | 3.15.2 | +| Previous and retained Rust | 1.99.1 | 3.15.2 | +| Upcoming Python | 1.100.0 | 3.16.0a2 | +| Previous Python | 1.100.0 | 3.14.10 | + +Each contract lane passed 365 Rust tests, 462 Python script tests, 18 reference +tests and eight oracle tests, plus the explicit integration/example stages. The +retained logs contain no compiler/Python warnings or test skips. The sixth job +checks previous-checkout compatibility. These are overlapping runtime checks, +not five independent model experiments. The separate scheduled recovery workflow +was not dispatched as part of this integration. + +The Store now includes shared current-eligibility explanations, atomic admission +batches, the fixture-backed incident bridge, authorized decision export/replay, +and evaluated-policy activation. See the +[reliability review guide](../reliability-review-guide.md). The hosted adapter from +PR #18 remains qualified through offline transport/process fixtures; provider +compatibility and complete live billing reconciliation are not established. + +## Separate controlled-correction pilot + +The [observed results](../controlled-correction-results.md) and +[protocol](../controlled-correction-pilot.md) describe a later local experiment at +`f6694110e3df5b1a13f71dbf9fae2bde89e07c44`. Its cached Ornith/Ollama execution is +separate from the report's pinned Qwen measurements and the incident fixtures. + +All four planned changes completed. Post-change repair used four calls and 751 +reported tokens; fresh recomputation used six calls and 1,167 tokens. Including +the shared three-call bootstrap once, collection recorded 13 calls and 2,515 +tokens. Corresponding generated requests and checked plans matched; withdrawal +refused before dispatch in both arms. The sample is one correlated synthetic +example per condition. It establishes neither population accuracy nor latency +savings, and its configured token envelopes are not proven hard billing bounds. + +The reviewer handoff includes the corrected pilot supplement with SHA-256 +`53061caf647e72a51cd5410b545a7f774346646e31feb28ecbdf91e177eb54e2`. +This successor fixes three guide links; raw evidence remains unchanged. Its +original cloud copy has SHA-256 +`d1e6160e0f7e19777cefb7cd3005e5666447bc87edf7b213c83629a0da66487b` +and is not the corrected successor. + +## Training and delivery boundaries + +The existing AutoScientist run `fb16665b-f02c-4cdc-bbde-22cacfd81830` was still +running when checked on 16 November: one of three iterations completed, reported +best win rate 0.1689 against target 0.8, no error and no downloadable model. +That platform metric has not been independently evaluated and is not a KairoTensor +research result. No dataset expansion, restart or new training run is part of this +handoff. A later observation must recheck the run rather than assume this status. + +The repository is now public, verified on 16 November at 08:36 UTC through +unauthenticated requests. GitHub reported `private: false` and `visibility: public`; +the main README and the measured `20fab84` README returned HTTP 200, with the +latter matching the local Git snapshot byte for byte. The +[verification receipt](verification/public-access-20261116/anonymous-check.json) +records the exact URLs and hashes. This closes the source-access gap. + +At the 08:36 UTC check, the public PDF on main was an earlier export (SHA-256 +`a755007e453b633cae4f19d38fd1c1f557c365a04d07ddfffb52d5753bf10c5a`). +The earlier editorial successor had SHA-256 +`86bf058fcaf0cf5ed886af7b198a7a6aff9978bac5e2994a5a217e9f8d4c692c`. +The current presentation successor in this revision has SHA-256 +`5350a885ee9be61300b2743ef35f78bed674e283578a5c4aa487edd4f0b9b476` and seven pages. It adds restrained teal +accents and directs readers to the public measured source. Abstract, results and +author declaration are unchanged. The repository overview now links the research +narrative, a documentation index and the preserved detailed Store guide. See +[presentation verification](verification/presentation-20261116/README.md). + +The owner authorized read-only link access to the final Drive report PDF and +combined reviewer bundle. The last Drive checks at 08:29 UTC still showed +owner-only access and anonymous reads reached sign-in. The connector supports +named-user and Workspace-domain sharing but does not expose "Anyone with the +link." GitHub visibility does not change these Drive permissions. Earlier PDF +and package statements about repository privacy are dated and superseded by this +note and the current paper; old verified bytes and historical receipts remain unchanged. + +The refreshed combined reviewer bundle contains the current report package, +corrected pilot package, this status note, the exact integrated source and +source-bound CI receipts. Its current report package retains the measured-source +archive and raw evidence unchanged. Prior packages remain retained locally as +historical evidence. Private submission-form contact details and account-access +observations are omitted from the reviewer distribution. Its manifest verifies +byte consistency; it does not authenticate model execution. + +The [official sprint page](https://apartresearch.com/sprints/ai-x-epistemics-research-sprint-2026-11-13-to-2026-11-15) +lists the submission deadline as 16 November at 11:59 UTC, which is 06:59 in +America/Toronto. The report PDF is the primary upload; the form also accepts an +additional-material URL. A local ZIP or owner-only link is not an accessible URL +for judges. The public repository can now serve as a source/evidence URL; public +access to the latest Drive PDF and combined bundle is still unverified. +Event submission remains a separate action; this preparation establishes no +event receipt. + +Ready form values remain: title **KairoTensor: Evidence-Grounded Revision for Agent +Memory**; team **KairoTensor**; location **Online**; author **Alejandro Castellanos**; +affiliation **Independent Research**; publishing interest **Yes**; primary track +**Track 2: Trust Infrastructure**. The abstract is the unchanged 192-word text in +the [report](report.md). Contact details are retained in the private handoff. diff --git a/docs/submission/EVIDENCE.md b/docs/submission/EVIDENCE.md index 043fdb2..beb2503 100644 --- a/docs/submission/EVIDENCE.md +++ b/docs/submission/EVIDENCE.md @@ -3,8 +3,24 @@ Paths below are relative to the measured source root. The accompanying source archive is an exact snapshot of `20fab84a6baa525452191ecf80eeee339f0b5ac7`. CI collection success and model-task success are separate outcomes. Later -reliability commits, unmerged PR #18 and active AutoScientist runs are outside -this measured snapshot and are excluded from the report's outcomes. +reliability commits, the hosted adapter from PR #18, the separate correction pilot +and AutoScientist runs are outside this measured snapshot and are excluded from +the report's outcomes. [Current status](CURRENT_STATUS.md) describes the later +integration and points to the separately reported pilot. + +The final report's Table 1 summarizes the atomic-revision, scoped-reuse and +recovery rows below. Its values are existing observations, not new experiments. +Figure 1 illustrates the two scoped-reuse correction paths in row three; its +caption retains the same-branch, explicit-contract and admission-history checks. +Appendix Table 2 distinguishes retained artifacts, offline recomputation and +claims requiring fresh evidence. Neither addition supplies a new model result. +The [combined reviewer package](https://drive.google.com/file/d/1Nc2X8NwHdDpKTSU2kknw0V_V5cqOEBOP/view) +contains a current report package with the unchanged measured-source archive and +raw evidence, plus the separate pilot and integrated-source supplements. Drive access +must be checked independently of the package's byte manifest. The public +[measured source](https://github.com/k2jac9/kairotensor/tree/20fab84a6baa525452191ecf80eeee339f0b5ac7) +also contains the original artifacts. The current presentation changes headings, +layout accents and source links; it adds no empirical result. | Report claim | Raw evidence and protocol | Result and limit | | --- | --- | --- | diff --git a/docs/submission/README.md b/docs/submission/README.md index 5debb5c..f92dfe7 100644 --- a/docs/submission/README.md +++ b/docs/submission/README.md @@ -1,5 +1,9 @@ # KairoTensor submission guide +[Read the PDF](kairotensor-report.pdf) · [Read the manuscript](report.md) · [Inspect the evidence](EVIDENCE.md) · [Project overview](../../README.md) + +> **Evidence-grounded revision for agent memory.** A study of which saved work remains eligible after its declared support changes, with measured costs, one bounded reuse example, and preserved negative results. + **KairoTensor: Evidence-Grounded Revision for Agent Memory** Alejandro Castellanos, Independent Research @@ -8,6 +12,10 @@ Apart Research AI x Epistemics Research Sprint, 13-15 November 2026 Recommended track: **Track 2: Trust Infrastructure** +For the integrated repository, separate correction-pilot supplement and delivery +status checked on 16 November, see [Current status](CURRENT_STATUS.md). The paper +and historical verification receipts below retain their original source boundary. + KairoTensor records the evidence supplied to an agent computation and checks whether a saved result remains eligible after a correction or withdrawal. The report evaluates atomic revision, scoped reuse and a failed recovery experiment. @@ -24,8 +32,10 @@ are separately attributed in manifests and the report. The measured main-branch snapshot is `20fab84a6baa525452191ecf80eeee339f0b5ac7`. Historical experiments use the source identities in [EVIDENCE.md](EVIDENCE.md). These submission documents were prepared -after that snapshot. Later reliability commits, unmerged PR #18 and active -AutoScientist runs are excluded from the reported results. +after that snapshot. Later reliability commits, the hosted adapter introduced by +PR #18, the separate correction pilot and AutoScientist runs are excluded from +the reported results. PR #18 has since merged; that integration does not change +which experiments the paper reports. ## Review @@ -44,16 +54,52 @@ The 16 November finalization adds Doyle's truth maintenance system and de Kleer' assumption-based TMS to the related work. The contribution is the durable integration of declared evidence dependence at the model execution boundary; dependency tracking itself is established prior work. Measured outcomes are -unchanged. The report retains a 192-word abstract, five pages through code and -data, one reference page, and one page with the ethics appendix and LLM statement. +unchanged. The final template review adds three explicit contribution bullets, +Table 1 with observed outcomes and limits, an evidence-package link, and scope +wording that remains accurate after PR #18 merged. The following editorial pass +opens with the concrete correction example, adds Figure 1 and a reproducibility +table, labels the four findings, and uses consistent serif typography and compact +spacing. The unchanged author provenance declaration is now in the appendix. +The report retains a 192-word abstract and every original results paragraph; it +has five pages through code and data, one reference page, and one page with the +ethics appendix, provenance, Table 2 and LLM statement. The five main pages are +below the eight-page limit and above the four-page recommendation. New receipts are in [verification/finalization-20261116](verification/finalization-20261116/README.md). The original 15 November receipts remain unchanged and describe the earlier -six-page export. They are historical records, not checks of the final PDF. - -The repository was private when the package was prepared. The owner plans to -make it public; this preparation does not change visibility or submit an entry. -Reviewers need the ZIP or repository access to inspect the evidence. +six-page export. Those receipts and the first 16 November seven-page finalization +are historical records. [Template-review verification](verification/improvements-20261116/README.md) +checks the previous eight-page PDF. [Editorial verification](verification/editorial-20261116/README.md) +checks the current seven-page successor. These dated records are retained. + +The native Google Doc remains editable. Numbered PDF delivery adds a footer to +the native export using the [pinned export helper](assets/README.md); every body +pixel, text passage and URI link is checked against that export. A direct future +Docs export will need that numbering step again. Figure source and rendering +assets are included in the reviewer package. + +## Public source and delivery + +The [repository](https://github.com/k2jac9/kairotensor) and +[measured snapshot](https://github.com/k2jac9/kairotensor/tree/20fab84a6baa525452191ecf80eeee339f0b5ac7) +are public. Anonymous source reads succeeded on 16 November at 08:36 UTC; +[the receipt](verification/public-access-20261116/anonymous-check.json) records +that observation. Use the [PDF in this revision](kairotensor-report.pdf) and the +[evidence index](EVIDENCE.md) as the primary reviewer entry points. + +The presentation successor uses original navy/cyan/violet/amber repository +artwork and restrained teal accents in the seven-page paper. It changes public +source links and presentation, preserving the abstract, results, author declaration +and measured source. [Presentation verification](verification/presentation-20261116/README.md) +records the export and checks. Earlier PDFs and verification receipts remain +historical; use the hashes in the current receipt when comparing copies. + +The optional [Drive PDF](https://drive.google.com/file/d/1dTGN3-QSCzFGtf2vwILFT7V1CRl_CmIQ/view) +and [combined evidence package](https://drive.google.com/file/d/1Nc2X8NwHdDpKTSU2kknw0V_V5cqOEBOP/view) +are stable delivery targets. Their permissions are separate from GitHub; public +Drive access requires its own anonymous read check. GitHub source and the paper's +pinned evidence links do not depend on Drive sharing. No event entry has been +submitted by this preparation. ## Verification scope @@ -78,7 +124,10 @@ tokens and zero unresolved reservations. Its bytes were unchanged after auditing No new inference was run in preparing this package. A successful audit is not independent authentication of model or hardware execution. -The newer incident-routing calibration remains unexecuted. The current CI does -not implement the complete stable/preview/previous-stable compatibility matrix; -no such matrix is claimed here. Historical model/runtime pins are retained for -reproduction, not presented as current release recommendations. +The incident-routing model-quality evaluation remains unexecuted. The measured +snapshot's CI predates the complete stable/upcoming/previous compatibility matrix. +The integrated Rust/Python matrix subsequently passed at `109798d`; its exact +source and coverage are recorded in [Current status](CURRENT_STATUS.md). Those +checks do not qualify the separate ML runtime or establish model performance. +Historical model/runtime pins are retained for reproduction, not presented as +current release recommendations. diff --git a/docs/submission/assets/README.md b/docs/submission/assets/README.md new file mode 100644 index 0000000..81aabfa --- /dev/null +++ b/docs/submission/assets/README.md @@ -0,0 +1,33 @@ +# Report assets and export + +`correction-paths.dot` is the editable source for Figure 1. Its SVG and PNG are +presentation artifacts, not new empirical evidence. Render with Graphviz: + +```sh +dot -Tsvg -Gdpi=200 correction-paths.dot -o correction-paths.svg +dot -Tpng -Gdpi=200 correction-paths.dot -o correction-paths.png +``` + +The accepted render used installed Graphviz 15.1.1. The native document embeds +the PNG at 420 by 116.1 points; the caption states the full explicit reuse boundary. + +Google Docs was edited in place. Its public batch-update request catalogue does +not expose insertion of a page-number AutoText field: +https://developers.google.com/workspace/docs/api/reference/rest/v1/documents/request +The final delivery therefore adds numbered footers after exporting the native PDF. + +In an isolated environment with the hash-pinned wheel in `pdf-requirements.txt`: + +```sh +python number_pdf_pages.py native-export.pdf numbered-report.pdf +``` + +The helper preserves the input, refuses an existing output, uses an artifact-marked +footer and isolates the original graphics state. The seven body renders, extracted +text and all URI annotations were compared before/after numbering. This helper +is an export step, not an application dependency or a claim of universal PDF +compatibility. The accepted input is a non-encrypted, unrotated Google Docs export. + +The native exporter repeats one title destination key at two positions on the +same title page. pypdf emits a normalization warning; the original export and +warning are retained. Body pixels, text and URI links are unchanged by the step. diff --git a/docs/submission/assets/correction-paths.dot b/docs/submission/assets/correction-paths.dot new file mode 100644 index 0000000..abe3455 --- /dev/null +++ b/docs/submission/assets/correction-paths.dot @@ -0,0 +1,22 @@ +digraph correction_paths { + graph [rankdir=LR, bgcolor="white", pad="0.12", nodesep="0.34", ranksep="0.38", splines=ortho, outputorder=edgesfirst]; + node [shape=box, style="rounded,filled", fontname="Liberation Serif", fontsize=13, + color="#526475", fillcolor="#F4F6F8", penwidth=0.8, margin="0.14,0.12"]; + edge [color="#526475", arrowsize=0.65, penwidth=0.9]; + + policy [label="Policy corrected\nReport unchanged", group="top"]; + reuse [label="Reuse report judgment\nwhen contract checks pass", group="top", fillcolor="#EDF3F6"]; + plan1 [label="New planner call\nusing corrected policy", group="top"]; + report [label="Report replaced", group="bottom"]; + fresh [label="Fresh report judgment", group="bottom", fillcolor="#F8F2E9"]; + plan2 [label="New planner call\nusing new judgment", group="bottom"]; + + {rank=same; policy; report;} + {rank=same; reuse; fresh;} + {rank=same; plan1; plan2;} + policy -> reuse -> plan1; + report -> fresh -> plan2; + policy -> report [style=invis]; + reuse -> fresh [style=invis]; + plan1 -> plan2 [style=invis]; +} diff --git a/docs/submission/assets/correction-paths.png b/docs/submission/assets/correction-paths.png new file mode 100644 index 0000000..8eaa35b Binary files /dev/null and b/docs/submission/assets/correction-paths.png differ diff --git a/docs/submission/assets/correction-paths.svg b/docs/submission/assets/correction-paths.svg new file mode 100644 index 0000000..5c1d3f1 --- /dev/null +++ b/docs/submission/assets/correction-paths.svg @@ -0,0 +1,80 @@ + + + + + + +correction_paths + + + +policy->reuse + + + + + + +reuse->plan1 + + + + + + + +report->fresh + + + + + +fresh->plan2 + + + + + +policy + +Policy corrected +Report unchanged + + + +reuse + +Reuse report judgment +when contract checks pass + + + +plan1 + +New planner call +using corrected policy + + + +report + +Report replaced + + + +fresh + +Fresh report judgment + + + +plan2 + +New planner call +using new judgment + + + diff --git a/docs/submission/assets/number_pdf_pages.py b/docs/submission/assets/number_pdf_pages.py new file mode 100644 index 0000000..22b52d1 --- /dev/null +++ b/docs/submission/assets/number_pdf_pages.py @@ -0,0 +1,68 @@ +#!/usr/bin/env python3 +"""Add visible page numbers to a native export without reflowing its contents. + +Use pypdf 7.1.0 from pdf-requirements.txt. Retain the original export and verify +text, links and rendered body pixels separately before distributing the result. +""" +import argparse +from pathlib import Path + +from pypdf import PdfReader, PdfWriter +from pypdf.generic import ArrayObject, DecodedStreamObject, DictionaryObject, NameObject + + +def main(): + parser = argparse.ArgumentParser() + parser.add_argument('source', type=Path) + parser.add_argument('destination', type=Path) + args = parser.parse_args() + if args.source.resolve() == args.destination.resolve() or args.destination.exists(): + parser.error('retain the source and choose a new output path') + reader = PdfReader(args.source) + if reader.is_encrypted: + parser.error('encrypted documents are outside this export helper') + writer = PdfWriter(clone_from=reader) + font = DictionaryObject({ + NameObject('/Type'): NameObject('/Font'), + NameObject('/Subtype'): NameObject('/Type1'), + NameObject('/BaseFont'): NameObject('/Times-Roman'), + }) + font_ref = writer._add_object(font) + for number, page in enumerate(writer.pages, 1): + if page.rotation != 0: + parser.error('rotated pages require an explicitly reviewed layout') + resources = DictionaryObject(page['/Resources'].get_object()) + fonts = DictionaryObject(resources.get('/Font', DictionaryObject()).get_object()) + if '/KTPageNumber' in fonts: + parser.error('page-number font resource already exists') + fonts[NameObject('/KTPageNumber')] = font_ref + resources[NameObject('/Font')] = fonts + page[NameObject('/Resources')] = resources + label = str(number) + # Times-Roman digits have 500-unit advance; 9 pt -> 4.5 pt per digit. + x = float(page.mediabox.left) + (float(page.mediabox.width) - 4.5 * len(label)) / 2 + y = float(page.mediabox.bottom) + 32 + stamp = DecodedStreamObject() + stamp.set_data((f'q /Artifact BMC 0.35 g BT /KTPageNumber 9 Tf ' + f'1 0 0 1 {x:.3f} {y:.3f} Tm ({label}) Tj ET EMC Q\n').encode('ascii')) + original = page.get('/Contents') + if isinstance(original, ArrayObject): + streams = ArrayObject(original) + else: + streams = ArrayObject([original]) if original is not None else ArrayObject() + # Native exports may leave a transformed coordinate system active. + # Isolate the original page's graphics state before drawing the footer. + save_state = DecodedStreamObject() + save_state.set_data(b'q\n') + restore_state = DecodedStreamObject() + restore_state.set_data(b'Q\n') + page[NameObject('/Contents')] = ArrayObject([ + writer._add_object(save_state), *streams, + writer._add_object(restore_state), writer._add_object(stamp), + ]) + with args.destination.open('xb') as output: + writer.write(output) + + +if __name__ == '__main__': + main() diff --git a/docs/submission/assets/pdf-requirements.txt b/docs/submission/assets/pdf-requirements.txt new file mode 100644 index 0000000..e6b4ded --- /dev/null +++ b/docs/submission/assets/pdf-requirements.txt @@ -0,0 +1,2 @@ +# Export helper only; not an application dependency. +pypdf==7.1.0 --hash=sha256:47cca6b24bdcf4c4d65b5f7b8b972535c61710a061261e737688777cfd9a7bfb diff --git a/docs/submission/kairotensor-report.pdf b/docs/submission/kairotensor-report.pdf index 013bebe..57af97e 100644 Binary files a/docs/submission/kairotensor-report.pdf and b/docs/submission/kairotensor-report.pdf differ diff --git a/docs/submission/report.md b/docs/submission/report.md index 0e8eb34..f8db86e 100644 --- a/docs/submission/report.md +++ b/docs/submission/report.md @@ -8,11 +8,17 @@ Agents often retain conclusions after the evidence underlying them changes. We s ## 1. Introduction -An agent may make a decision from a set of observations and later receive a correction. Saving the answer alone loses a practical question: which evidence did the computation actually receive, and is the result still eligible when a source is withdrawn? Reusing an obsolete continuation can carry a superseded claim into subsequent actions even when the new evidence is available. +The problem. A deployment agent judges a test report and then chooses a release plan under a policy. If the policy changes, the plan needs revision, but the report judgment may still be usable. If the report itself is replaced, the old judgment must not silently carry over. The useful question is therefore not just what the agent answered, but which supplied evidence still permits reusing its work. -KairoTensor investigates a narrower, operational question: can a local store keep a durable record of an agent computation's declared evidence and refuse stale reads or continuations, while permitting explicitly bounded reuse? The prototype provides an occurrence-level evidence ledger, revision-checked branches, and a deterministic alternative-witness mechanism. It does not authenticate evidence or infer what a model internally attended to. The caller must authenticate actors, supply all relevant dependencies, validate outputs, and correctly report usage. +Our approach. KairoTensor records the evidence supplied to an agent computation and checks whether its saved result remains eligible after a correction or withdrawal. Its Rust/SQLite store preserves evidence occurrences, branch histories, frozen requests and usage. Reuse is an explicit operation under a closed, versioned computation contract; an ordinary read does not automatically revalidate a cached model result. Eligibility does not authenticate a source, establish truth or reveal what the model internally attended to. The caller must authenticate actors, declare all relevant inputs, validate outputs and report usage. -Alejandro Castellanos confirms that no KairoTensor-specific ideas, code or experiments predate the sprint. The repository history begins on 13 November 2026, and the measured revision and model experiments were committed on 13-14 November. This provenance statement is the author's declaration; commit timestamps alone cannot establish when work began. This report covers the measured snapshot 20fab84a6baa525452191ecf80eeee339f0b5ac7. Later reliability work, the unmerged hosted-integration branch and active AutoScientist runs are excluded from the reported results. +Our contributions are: + +- A durable Rust/SQLite execution contract that records declared evidence and refuses stale results while retaining incurred usage. +- A reproducible synthetic comparison of revision strategies that distinguishes equivalent publication histories from a lighter baseline. +- A bounded model-reuse demonstration with a changed-input control, alongside negative recovery results that expose output-format and deadline limitations. + +This report covers the measured snapshot 20fab84a6baa525452191ecf80eeee339f0b5ac7. Subsequent reliability and hosted-integration work, the separate correction pilot, and AutoScientist training are outside this report's measured scope. ## 2. Related work @@ -36,21 +42,37 @@ The revision benchmark uses file-backed SQLite, synthetic rules with two three-p The model reuse example separates a test-report judgment from a deployment planner. A policy correction changes the planner's input but not the report judgment's closed, versioned inputs. Reuse requires the same request bytes, candidate list, contract and relevant admissions; changing the report is a negative control. The live example uses Qwen2.5-1.5B-Instruct, model and tokenizer revision 989aa7980e4cf806f80c7fef2b1adb7bc71aa306, CPU, two threads, PyTorch 2.6.0+cpu and Transformers 4.48.3. Route timings begin after evidence-view preparation and model loading. One revised-prompt run compares repair and fresh rebuild; it is not a randomized population study [7]. +Figure 1 shows the two correction paths used in the model-reuse experiment. + +![Two correction paths: policy correction permits checked reuse; report replacement requires fresh judgment.](assets/correction-paths.png) + +Figure 1. Policy-only correction permits checked reuse; replacing the report requires fresh judgment [7]. Reuse is opt-in and same-branch, with matching versioned contract, exact request bytes, ordered candidates/inputs and admission history. The reused judgment is one accepted realization, not an independent sample or proof of truth. + ## 4. Results -In the atomic revision experiment, 128 rules with 32 dense edits took a median 1,972.5 ms for per-branch edits versus 146.7 ms for atomic edits. The median paired speedup was 13.46 times, range 13.24–13.56 across six trials. For 1,024 rules and 128 sparse edits, it was 1.40 times, range 1.34–1.68. Atomic edits improved edit time in all 30 matched comparisons. The lighter full scan had the lowest total time in every matched trial; for the dense cell its median total was 27.8 ms versus 369.1 ms for atomic indexed revision. The baseline's reduced historical contract prevents attributing the difference to indexing alone [6]. +Table 1 summarizes the three main experimental findings. + +Table 1. Observed outcomes and their limits [6-8]. Timings, token counts and success rates describe separate experiments; they are not a combined performance measure. + +| Experiment | Observed result | Scope and limitation | +| --- | --- | --- | +| Atomic revision | Dense cell: 13.46x median paired edit-time speedup (range 13.24-13.56). | Six paired trials; fixed synthetic trace. Lighter full scan wins total time. | +| Model-result reuse | Repair: 1 call / 193 tokens. Fresh: 2 calls / 404 tokens. Same checked plan. | One observation per route after prompt revision; no general savings estimate. | +| Recovery policies | Each of three policies: 0/6 held-out successes. | Shared trajectories; strict JSON failures; every retry late. | -In a later bulk setup experiment, the 1,024-rule fixture's median setup time was 3,690 ms using individual production calls versus 214 ms using bulk calls. The median paired setup ratio was 16.62, range 14.89–34.55, across six trials. The scan retained the lowest median total in all five cells and finished first in 29 of 30 matched trials [6]. These are fixed synthetic, single-writer traces on hosted machines, not a general break-even result. +**Atomic revision reduced edit time.** In the atomic revision experiment, 128 rules with 32 dense edits took a median 1,972.5 ms for per-branch edits versus 146.7 ms for atomic edits. The median paired speedup was 13.46 times, range 13.24–13.56 across six trials. For 1,024 rules and 128 sparse edits, it was 1.40 times, range 1.34–1.68. Atomic edits improved edit time in all 30 matched comparisons. The lighter full scan had the lowest total time in every matched trial; for the dense cell its median total was 27.8 ms versus 369.1 ms for atomic indexed revision. The baseline's reduced historical contract prevents attributing the difference to indexing alone [6]. -In the revised-prompt model example, repair reused one report judgment, made one new call and used 145 input plus 48 output tokens. A clean rebuild made two calls and used 326 input plus 78 output tokens. Both produced the same checked canary plan; the observed savings were 211 of 404 tokens. Replacing the report forced a fresh judgment and a hold decision. The earlier prompt had produced an incorrect passing status for an 11/12 report, which the validator rejected; the prompt was revised before the reported successful run. Each route is one observation, not evidence of typical savings or semantic reliability [7]. +**Bulk setup reduced initialization cost.** In a later bulk setup experiment, the 1,024-rule fixture's median setup time was 3,690 ms using individual production calls versus 214 ms using bulk calls. The median paired setup ratio was 16.62, range 14.89–34.55, across six trials. The scan retained the lowest median total in all five cells and finished first in 29 of 30 matched trials [6]. These are fixed synthetic, single-writer traces on hosted machines, not a general break-even result. -A separate frozen recovery-policy evaluation illustrates the remaining model risk. Ten synthetic cases generated 20 Qwen calls. All responses included Markdown fences and failed the unchanged strict JSON contract. On the six held-out cases, the stop, affordable-retry and evidence-gated-retry policies each had zero successes. These policy outcomes are paired replays of shared trajectories, not independent policy arms. All retry receipts arrived after the 60-second operation deadline; the deadline was soft and did not interrupt dispatched calls [8]. +**Scoped reuse saved work in one correction.** In the revised-prompt model example, repair reused one report judgment, made one new call and used 145 input plus 48 output tokens. A clean rebuild made two calls and used 326 input plus 78 output tokens. Both produced the same checked canary plan; the observed savings were 211 of 404 tokens. Replacing the report forced a fresh judgment and a hold decision. The earlier prompt had produced an incorrect passing status for an 11/12 report, which the validator rejected; the prompt was revised before the reported successful run. Each route is one observation, not evidence of typical savings or semantic reliability [7]. + +**Frozen recovery failed the contract and deadline.** A separate frozen recovery-policy evaluation illustrates the remaining model risk. Ten synthetic cases generated 20 Qwen calls. All responses included Markdown fences and failed the unchanged strict JSON contract. On the six held-out cases, the stop, affordable-retry and evidence-gated-retry policies each had zero successes. These policy outcomes are paired replays of shared trajectories, not independent policy arms. All retry receipts arrived after the 60-second operation deadline; the deadline was soft and did not interrupt dispatched calls [8]. ## 5. Discussion and limitations The implementation makes evidence dependence explicit at an application boundary and preserves the cost of work that becomes unusable. The measurements show that atomic batching can improve a costly per-branch path without establishing that indexed revision beats a lighter scan. The model example shows that a carefully scoped cached judgment can be reused in one controlled correction, not that agent decisions improve in the wild. The strict recovery failures show how validation, model formatting and time budgets can dominate the usefulness of such a mechanism. -Future work should evaluate real correction streams and user decisions, test concurrency and adversarial callers, compare equally durable baselines, and run preregistered, independently measured model evaluations. The frozen incident-routing workload and bounded controller on main are offline preparations; they do not establish model performance. A reviewer package contains the pinned source snapshot, evidence index and offline verification commands. Repository and workflow links require access while the repository remains private. +Future work should evaluate real correction streams and user decisions, test concurrency and adversarial callers, compare equally durable baselines, and run preregistered, independently measured model evaluations. The frozen incident-routing workload and bounded controller on main are offline preparations; they do not establish model performance. A reviewer package contains the pinned source snapshot, evidence index and offline verification commands. The repository and measured source snapshot are publicly accessible. ## 6. Conclusion @@ -60,6 +82,8 @@ Versioned evidence and conservative checks can prevent this prototype from servi Measured main-branch snapshot: 20fab84a6baa525452191ecf80eeee339f0b5ac7, https://github.com/k2jac9/kairotensor. Raw benchmark CSVs, prompts, model traces and evaluation protocols are in docs/benchmarks and docs/evaluations. Historical CI executions use their recorded source commits, not the current main commit. The accompanying reviewer package preserves the source archive and its checksum; exact experimental sources and commands are documented in [6-8]. +Public source and evidence: [measured snapshot 20fab84](https://github.com/k2jac9/kairotensor/tree/20fab84a6baa525452191ecf80eeee339f0b5ac7). The public snapshot includes the measured source and raw evidence; no GitHub sign-in is required. + ## References [1] Moreau, L. and Missier, P. (eds.), 2013. PROV-DM: The PROV Data Model. W3C Recommendation. https://www.w3.org/TR/prov-dm/ @@ -82,6 +106,20 @@ Measured main-branch snapshot: 20fab84a6baa525452191ecf80eeee339f0b5ac7, https:/ An operator could use admission controls to suppress inconvenient evidence, present a selective record as complete, or convey unwarranted authority through a provenance-like interface. The store does not establish source authenticity, prevent tampering by a privileged operator, detect omitted dependencies, guarantee semantic correctness, or erase copies already delivered. Deployment would need access control, independent auditing, transparent inclusion and exclusion policies, and protection for stored sensitive information. These are recommendations, not implemented protections. The reported examples use synthetic reports and deployment data; this statement does not establish the privacy properties of future applications. +### Provenance of the work + +Alejandro Castellanos confirms that no KairoTensor-specific ideas, code or experiments predate the sprint. The repository history begins on 13 November 2026, and the measured revision and model experiments were committed on 13-14 November. This provenance statement is the author's declaration; commit timestamps alone cannot establish when work began. + +### Reproducibility boundary + +Table 2. What the retained evidence can establish [6-8]. Recomputing a recorded result is distinct from re-executing a model or independently authenticating its execution. + +| Evidence class | Available check and limit | +| --- | --- | +| Retained artifacts | Pinned source, benchmark CSVs, archived model requests/responses and the recovery ledger. Hashes identify bytes; stored metadata is not independent authentication. | +| Recomputed offline | Benchmark summaries and archived recovery consistency checks can be repeated with zero model calls. They reproduce calculations on recorded data. | +| Requires fresh evidence | Re-executing model requests, estimating typical accuracy or savings, representative latency, provider billing and authentic execution require further evidence. Archived traces alone do not establish these claims. | + ## LLM usage statement AI assistance was used to help develop code and prepare this report. The author is responsible for the claims and final submission. Automated checks recomputed benchmark summaries and reconciled the archived recovery records with their SQLite ledger; these checks establish internal consistency, not independent replication of model execution. diff --git a/docs/submission/verification/editorial-20261116/README.md b/docs/submission/verification/editorial-20261116/README.md new file mode 100644 index 0000000..8a8a0cb --- /dev/null +++ b/docs/submission/verification/editorial-20261116/README.md @@ -0,0 +1,41 @@ +# Editorial successor verification + +Base `eebfdef26be08f4edb9fc6d975fcea20f0ffd04f`. The measured research source remains +`20fab84a6baa525452191ecf80eeee339f0b5ac7`. + +Changes implement the approved reference-paper review: concrete correction +example, Figure 1, finding labels, consistent typography/spacing, page numbers, +and Table 2 on reproducibility. The author provenance declaration moves intact +to the appendix. The 192-word abstract and all four original Results paragraphs +are preserved; prior evidence, archives and verification receipts are unchanged. + +[report-check.json](report-check.json) binds the native revision, original export +and final numbered PDF. Seven pages: five main including title, references on +page 6 and appendix on page 7. The diagram and caption are together, both tables +are legible, and section headings remain with their following text. All pages +were rendered and inspected. + +Native Google Docs edits used revision guards. The PDF-only numbering helper +uses hash-pinned pypdf 7.1.0. Verification compares every page's body pixels, text +and URI links before/after numbering; only footer digits may differ. A native +export duplicate title-destination warning is retained and documented in the +asset guide. The original export remains available in the retained packet. + +One formatting batch was rejected atomically because page-break-before is not +allowed in table cells; readback confirmed no revision change before a corrected +request. The first numbering candidate inherited the export's flipped coordinate +system; wrapping original content in a saved graphics state corrected the footer. +The first provenance-check probe accidentally removed a space while constructing +its expected string; the checker now slices the exact baseline declaration. +These candidates and failures are retained; no artifact was changed to satisfy +an incorrect expected value. + +No new application tests, model calls, training or dataset expansion occurred. +The source-bound historical tests remain distinct from this document verification. +Authenticated cloud readback and anonymous access are separate checks; an uploaded +link alone does not establish reviewer access. No event submission is claimed. + +Full packet: `kairotensor-artifacts/editorial-pass-20261116`, including native +snapshots, render candidates, raw export, numbered PDF, scripts, hashes, storage +receipts and independent review. Post-upload delivery receipts are kept outside +the already-hashed reviewer bundle to avoid a self-referential manifest. diff --git a/docs/submission/verification/editorial-20261116/delivery-check.json b/docs/submission/verification/editorial-20261116/delivery-check.json new file mode 100644 index 0000000..f73b245 --- /dev/null +++ b/docs/submission/verification/editorial-20261116/delivery-check.json @@ -0,0 +1,64 @@ +{ + "cloud_readback": { + "checked_at_utc": "2026-11-16T08:30:25.103732+00:00", + "authenticated_downloads": [ + { + "file_id": "1dTGN3-QSCzFGtf2vwILFT7V1CRl_CmIQ", + "bytes": 256985, + "sha256": "86bf058fcaf0cf5ed886af7b198a7a6aff9978bac5e2994a5a217e9f8d4c692c", + "matches_local": true + }, + { + "file_id": "1Nc2X8NwHdDpKTSU2kknw0V_V5cqOEBOP", + "bytes": 3580195, + "sha256": "40c645a87766ad7e054a5c7a6e918b2a5349cd9396f7daf7b32a8f1fd35eb700", + "matches_local": true + } + ], + "public_access_verified": false + }, + "package": { + "all_payloads_read_back_and_hashed": true, + "broken_markdown_links": [], + "measured_source_sha256": "e5928a4daf5b9d68cfbec0f58afdfd53d65f4202e3f1c5d54a4284baa1c0b821", + "old_packages_unchanged": true, + "original_report_payloads": 45, + "outer_payloads": 20, + "pilot_manifest_payloads": 149, + "private_submission_fields_included": false, + "public_access_verified": false, + "raw_evidence_and_historical_receipts_unchanged": true, + "report_manifest_payloads": 54, + "report_zip_sha256": "96e7d160131f57615ab341103fa8438f92e324cb1fac6129c673acad41daba27", + "zip": "KairoTensor-submission-evidence-20261116.zip", + "zip_bytes": 3580195, + "zip_sha256": "40c645a87766ad7e054a5c7a6e918b2a5349cd9396f7daf7b32a8f1fd35eb700" + }, + "anonymous_read": { + "checked_at_utc": "2026-11-16T08:29:58.751850+00:00", + "authentication_sent": false, + "inherited_proxy_used": false, + "files": [ + { + "file_id": "1dTGN3-QSCzFGtf2vwILFT7V1CRl_CmIQ", + "status": 200, + "final_host": "accounts.google.com", + "content_type": "text/html; charset=utf-8", + "bytes": 946902, + "exact_payload_matches": false + }, + { + "file_id": "1Nc2X8NwHdDpKTSU2kknw0V_V5cqOEBOP", + "status": 200, + "final_host": "accounts.google.com", + "content_type": "text/html; charset=utf-8", + "bytes": 946622, + "exact_payload_matches": false + } + ], + "all_exact_payloads_accessible": false + }, + "event_submission_receipt": null, + "repository_visibility_changed": false, + "scope": "Post-upload receipt retained outside the already-hashed evidence bundle." +} diff --git a/docs/submission/verification/editorial-20261116/report-check.json b/docs/submission/verification/editorial-20261116/report-check.json new file mode 100644 index 0000000..9ac38cb --- /dev/null +++ b/docs/submission/verification/editorial-20261116/report-check.json @@ -0,0 +1,33 @@ +{ + "base": "eebfdef26be08f4edb9fc6d975fcea20f0ffd04f", + "native_revision": "AHj4eMS4SUmUDkRDSWnGiE15rtWgxLibLMBLnn8yvBaZle3sk6DE5gWF8lYv5Vrljf_w-Lq53M9vcTtojtDHTyAFVYhPJ3REKsamLvh9pNc", + "native_export_sha256": "8f211cacfe61610edbb240f0b85871ae5f0abf05d9d22f11d18ce73e0c0eac51", + "pdf_sha256": "86bf058fcaf0cf5ed886af7b198a7a6aff9978bac5e2994a5a217e9f8d4c692c", + "pdf_bytes": 256985, + "pdf_pages": 7, + "main_pages_including_title": 5, + "references_page": 6, + "appendix_page": 7, + "abstract_words": 192, + "abstract_unchanged": true, + "original_result_paragraphs_preserved": 4, + "provenance_declaration_preserved": true, + "reproducibility_table_cells_verified": 8, + "page_numbers_verified": [ + 1, + 2, + 3, + 4, + 5, + 6, + 7 + ], + "pdf_body_pixels_unchanged_by_numbering": true, + "links_unchanged_by_numbering": true, + "visual_review": "All seven pages inspected; diagram/caption together, both tables legible, section openings retained with their text.", + "native_export_warning": "Duplicate title destination h.bzdusysp7krw in Google export, with two positions on the same title page. pypdf logs normalization; original retained. Text, URI links and all body pixels verified unchanged.", + "new_model_calls": 0, + "application_tests_rerun": false, + "event_submitted": false, + "reviewer_access": "Requires fresh permission and anonymous-read verification; earlier check was owner-only." +} diff --git a/docs/submission/verification/improvements-20261116/README.md b/docs/submission/verification/improvements-20261116/README.md new file mode 100644 index 0000000..1e37d94 --- /dev/null +++ b/docs/submission/verification/improvements-20261116/README.md @@ -0,0 +1,36 @@ +# Template-review verification, 16 November 2026 + +Documentation successor to `03bcfb0adc5879aa1c937a15340a85f0a648ba90`. +The measured source remains `20fab84a6baa525452191ecf80eeee339f0b5ac7`. + +Changes: explicit contributions, a numbered/captioned native results table, +status-independent measured-scope wording and an evidence-package link. Native +Google Docs edits used required-revision guards and were read back before export. + +[report-check.json](report-check.json) binds the final PDF and native revision. +It checks the unchanged 192-word abstract and four original result paragraphs, +all 12 table cells against the editable manuscript, page/size limits and the PDF +evidence-link annotation. All eight pages were rendered and visually reviewed. +The report is six pages including the title, plus references and appendix; this +meets the eight-page maximum but exceeds the four-page recommendation. + +No source-code, historical experiment, measured-source archive or prior receipt +was changed. No new model, training, or application-test run occurred. Existing +source-bound tests remain historical evidence, not tests of this document edit. +The source and frozen recovery archive are retained byte for byte in the package. + +The owner authorized read-only reviewer access to the final PDF and combined +bundle. The current Drive connector cannot create an anonymous-reader permission. +The links must be set to "Anyone with the link / Viewer" and verified without +authentication before relying on reviewer access. Repository visibility is +unchanged; no event-submission receipt is established. + +The full retained packet includes before/after native snapshots, revision-guarded +edit plans, rendered pages, verification scripts, storage admission, package +manifests and review findings. It is under +`kairotensor-artifacts/report-improvements-20261116-apdv5hsp`. + +The post-upload delivery receipt is `delivery-check.json` in this repository. +It is outside the already-hashed reviewer bundle and records exact cloud/local +byte matches. Anonymous GETs at 08:07 UTC reached Google's sign-in page instead +of either payload. This is a concrete delivery limitation, not an evidence failure. diff --git a/docs/submission/verification/improvements-20261116/delivery-check.json b/docs/submission/verification/improvements-20261116/delivery-check.json new file mode 100644 index 0000000..781fc4e --- /dev/null +++ b/docs/submission/verification/improvements-20261116/delivery-check.json @@ -0,0 +1,64 @@ +{ + "cloud_readback": { + "checked_at_utc": "2026-11-16T08:06:57.696680+00:00", + "authenticated_downloads": [ + { + "file_id": "1dTGN3-QSCzFGtf2vwILFT7V1CRl_CmIQ", + "bytes": 165283, + "sha256": "da591098a638a0cddee86bcc1e42a5e7ec8bfd912d7768ae4bf5e38ce26705ca", + "matches_local": true + }, + { + "file_id": "1Nc2X8NwHdDpKTSU2kknw0V_V5cqOEBOP", + "bytes": 3018218, + "sha256": "ada372f553f26a038826e17b57fc176bd7cb05b6b3ad6ff594b8dab69ba27069", + "matches_local": true + } + ], + "anonymous_access_established": false + }, + "combined_package": { + "all_payloads_read_back_and_hashed": true, + "broken_markdown_links": [], + "measured_source_sha256": "e5928a4daf5b9d68cfbec0f58afdfd53d65f4202e3f1c5d54a4284baa1c0b821", + "old_packages_unchanged": true, + "original_report_payloads": 42, + "outer_payloads": 14, + "pilot_manifest_payloads": 149, + "private_submission_fields_included": false, + "public_access_verified": false, + "raw_evidence_and_historical_receipts_unchanged": true, + "report_manifest_payloads": 45, + "report_zip_sha256": "58c8a03d2876db4647f8c816ae85197f1bd67f27b588eecb05ca2171e55c9a17", + "zip": "KairoTensor-submission-evidence-20261116.zip", + "zip_bytes": 3018218, + "zip_sha256": "ada372f553f26a038826e17b57fc176bd7cb05b6b3ad6ff594b8dab69ba27069" + }, + "anonymous_read": { + "checked_at_utc": "2026-11-16T08:07:19.242550+00:00", + "authentication_sent": false, + "inherited_proxy_used": false, + "files": [ + { + "file_id": "1dTGN3-QSCzFGtf2vwILFT7V1CRl_CmIQ", + "status": 200, + "final_host": "accounts.google.com", + "content_type": "text/html; charset=utf-8", + "bytes": 947128, + "exact_payload_matches": false + }, + { + "file_id": "1Nc2X8NwHdDpKTSU2kknw0V_V5cqOEBOP", + "status": 200, + "final_host": "accounts.google.com", + "content_type": "text/html; charset=utf-8", + "bytes": 946600, + "exact_payload_matches": false + } + ], + "all_exact_payloads_accessible": false + }, + "scope": "This post-upload receipt is retained in the repository and local packet, outside the already-hashed reviewer bundle.", + "repository_visibility_changed": false, + "event_submission_receipt": null +} diff --git a/docs/submission/verification/improvements-20261116/report-check.json b/docs/submission/verification/improvements-20261116/report-check.json new file mode 100644 index 0000000..1e56788 --- /dev/null +++ b/docs/submission/verification/improvements-20261116/report-check.json @@ -0,0 +1,24 @@ +{ + "document_revision": "AHj4eMQw_hyQIRahlYtzjyvb79_98CW8JWz3jon9x0ArTa_W82jp354MHPBFgBMShqQtFt1fwK5JFsPqb9KAs_i7SEuxDAgIShs8H4AoVp8", + "pdf_sha256": "da591098a638a0cddee86bcc1e42a5e7ec8bfd912d7768ae4bf5e38ce26705ca", + "pdf_bytes": 165283, + "pdf_pages": 8, + "main_pages_including_title": 6, + "reference_pages": [ + 7 + ], + "appendix_pages": [ + 8 + ], + "abstract_words": 192, + "abstract_unchanged": true, + "original_result_paragraphs_unchanged": 4, + "contribution_bullets": 3, + "native_table_cells_matched_to_manuscript": 12, + "evidence_link_is_pdf_annotation": true, + "new_model_calls": 0, + "event_submitted": false, + "visual_review": "All eight pages reviewed; contributions, caption/table, references and appendix legible. Pages 1-3 text/layout unchanged by final spacing-only edit.", + "reviewer_access": "Pending sharing change and anonymous verification; link presence does not prove access.", + "previous_pdf_sha256": "a755007e453b633cae4f19d38fd1c1f557c365a04d07ddfffb52d5753bf10c5a" +} diff --git a/docs/submission/verification/presentation-20261116/README.md b/docs/submission/verification/presentation-20261116/README.md new file mode 100644 index 0000000..74bbeb7 --- /dev/null +++ b/docs/submission/verification/presentation-20261116/README.md @@ -0,0 +1,18 @@ +# Public presentation successor — 16 November 2026 + +This successor adds original repository artwork, a research-first overview and a +documentation map. Detailed API prose is retained in `docs/store-guide.md`. The +paper keeps seven pages and its academic template, with teal headings, pale table +headers and public pinned-source links. Its 192-word abstract, Results section, +author provenance and empirical boundaries are unchanged. + +[Export checks](report-check.json) bind the PDF, native export and current text. +The existing page-number helper preserves native content and URI annotations. +The duplicate native title-destination warning remains documented; no warning-free +export is claimed. Visual inspection covers all seven pages and both SVGs. + +This is documentation and presentation work. No model calls, training, new +measurements, production-code changes or event submission are part of it. Prior +receipts describe earlier exact files and remain unchanged. Current cloud upload, +independent review and public publication receipts are retained separately from +already-hashed packages so those receipts cannot recursively change the package. diff --git a/docs/submission/verification/presentation-20261116/delivery-check.json b/docs/submission/verification/presentation-20261116/delivery-check.json new file mode 100644 index 0000000..c18b32e --- /dev/null +++ b/docs/submission/verification/presentation-20261116/delivery-check.json @@ -0,0 +1,43 @@ +{ + "scope": "Post-upload receipt, kept outside the already-hashed reviewer package", + "package": { + "corrected_pilot_unchanged": true, + "historical_receipts_preserved": 39, + "measured_source_sha256": "e5928a4daf5b9d68cfbec0f58afdfd53d65f4202e3f1c5d54a4284baa1c0b821", + "outer_links": 12, + "outer_payloads": 23, + "pilot_payloads": 149, + "private_submission_fields_included": false, + "report_links": 27, + "report_payloads": 58, + "report_zip_sha256": "a692df5ae8207491ba212c459554c95fff41538e9527c2dae780515ce5128733", + "zip_bytes": 3804246, + "zip_sha256": "e9723df1e39c310ac899fad1bf42e808c4aed7f15b1a63be3e6cd939bb334d9b" + }, + "cloud": { + "checked_at_utc": "2026-11-16T08:47:21.938353+00:00", + "authenticated_readback": [ + { + "file": "kairotensor-report.pdf", + "sha256": "5350a885ee9be61300b2743ef35f78bed674e283578a5c4aa487edd4f0b9b476", + "bytes": 256329, + "matches_cloud": true + }, + { + "file": "KairoTensor-submission-evidence-20261116.zip", + "sha256": "e9723df1e39c310ac899fad1bf42e808c4aed7f15b1a63be3e6cd939bb334d9b", + "bytes": 3804246, + "matches_cloud": true + } + ], + "public_drive_access_verified": false, + "repository_public_verified_separately": true + }, + "visual_review": [ + "seven PDF pages", + "SVG hero and lifecycle", + "GitHub-rendered Markdown in approximate light and dark browser styles" + ], + "new_application_tests": false, + "new_model_calls": 0 +} diff --git a/docs/submission/verification/presentation-20261116/report-check.json b/docs/submission/verification/presentation-20261116/report-check.json new file mode 100644 index 0000000..19ddddb --- /dev/null +++ b/docs/submission/verification/presentation-20261116/report-check.json @@ -0,0 +1,26 @@ +{ + "pdf_pages": 7, + "abstract_words": 192, + "results_section_unchanged": true, + "author_provenance_unchanged": true, + "native_pdf_and_markdown_text_verified": true, + "native_revision": "AHj4eMSI1LxACd25i8cAOm-KHh0tFKHNhtS2W2pF1Fb7GOUpCq-FY7X-FRSVaig4P375ARk0cB7XbWGlAnqBEiszGqyRGpitc6JigWs_MKQ", + "pdf_sha256": "5350a885ee9be61300b2743ef35f78bed674e283578a5c4aa487edd4f0b9b476", + "pdf_bytes": 256329, + "native_export_sha256": "88a59eb79347a339f3937135baeb08f8aa61d0df529fd37eef4f6220cb4b6e48", + "old_readme_contracts_retained": true, + "relative_links_checked": 115, + "svg_contrast_ratios": { + "#F4F6FC": 14.01, + "#C1CBE0": 9.29, + "#66E4EA": 9.99, + "#C4AFFB": 7.84, + "#F5CE7A": 10.09, + "#AEBED9": 8.05, + "print_teal_on_white": 8.19 + }, + "svg_has_accessible_titles_descriptions": true, + "new_model_calls": 0, + "historical_source": "20fab84a6baa525452191ecf80eeee339f0b5ac7", + "numbering_warning": "Duplicate title destination in native export; existing helper normalizes it; text and URI annotations checked." +} diff --git a/docs/submission/verification/public-access-20261116/anonymous-check.json b/docs/submission/verification/public-access-20261116/anonymous-check.json new file mode 100644 index 0000000..96dc2c6 --- /dev/null +++ b/docs/submission/verification/public-access-20261116/anonymous-check.json @@ -0,0 +1,50 @@ +{ + "checked_at_utc": "2026-11-16T08:36:20.491715+00:00", + "authentication": "none; no Authorization header or cookies; environment proxy disabled", + "checks": [ + { + "name": "repository", + "url": "https://api.github.com/repos/k2jac9/kairotensor", + "status": 200, + "final_url": "https://api.github.com/repos/k2jac9/kairotensor", + "bytes": 5061, + "sha256": "aa5e2a1eb1aed22d0599dcf85b2c5a45e8657fb0307ff587d8533aa06ad108e3", + "content_type": "application/json; charset=utf-8", + "private": false, + "visibility": "public", + "default_branch": "main", + "html_url": "https://github.com/k2jac9/kairotensor", + "pushed_at": "2026-11-16T07:32:55Z" + }, + { + "name": "readme", + "url": "https://raw.githubusercontent.com/k2jac9/kairotensor/main/README.md", + "status": 200, + "final_url": "https://raw.githubusercontent.com/k2jac9/kairotensor/main/README.md", + "bytes": 15687, + "sha256": "aab4d29de9d90d22d03865cca51fadbd9bd7205ef71948ce11268e3f4d2d4a3d", + "content_type": "text/plain; charset=utf-8" + }, + { + "name": "measured_source", + "url": "https://raw.githubusercontent.com/k2jac9/kairotensor/20fab84a6baa525452191ecf80eeee339f0b5ac7/README.md", + "status": 200, + "final_url": "https://raw.githubusercontent.com/k2jac9/kairotensor/20fab84a6baa525452191ecf80eeee339f0b5ac7/README.md", + "bytes": 14057, + "sha256": "2a0a280b78091f542e9b0256577c856f5ebf35f70f08351415e359cd8ef6b8fe", + "content_type": "text/plain; charset=utf-8", + "matches_pinned_source": true + }, + { + "name": "remote_report", + "url": "https://raw.githubusercontent.com/k2jac9/kairotensor/main/docs/submission/kairotensor-report.pdf", + "status": 200, + "final_url": "https://raw.githubusercontent.com/k2jac9/kairotensor/main/docs/submission/kairotensor-report.pdf", + "bytes": 139531, + "sha256": "a755007e453b633cae4f19d38fd1c1f557c365a04d07ddfffb52d5753bf10c5a", + "content_type": "application/octet-stream", + "is_pdf": true, + "matches_final_editorial_pdf": false + } + ] +}