Proof bank
Receipts before claims.
Each entry is shaped for audit: ID, status, claim, setup, result, limit, reproduce, related repo/artifact, date, and confidence level.
Receipt anatomy
The proof bank is intentionally repetitive. Strong claims should be boring to audit: same fields, same limits, same reproduction pressure.
Local relaxation can generate global coherence
Toy constraint graph receipt showing a large tension reduction through local relaxation.
Claim: Local relaxation supports the narrow claim that local constraint updates can reduce global tension in a bounded graph.
Result: Tension fell from 156.216618 to 11.862658, a 92.4063% reduction.
Limit: Quadratic constraint graph only; this is not a full-scale reasoning benchmark.
Attractors emerge from local constraint relaxation
Hopfield-style binary graph receipt for noisy recovery and context-biased recovery.
Claim: Local relaxation can recover stable attractor states in a toy binary graph.
Result: 30% noise recovery was around 87%; context bias improved recovery to about 98%.
Limit: Toy attractor graph; requires replication beyond synthetic binary states.
Contradiction localizes as residual/provenance tension
Small graph receipt where a planted contradiction ranks first by residual energy and relief-if-removed.
Claim: Contradictions can localize as residual graph tension in a provenance-aware toy graph.
Result: The planted contradiction ranked #1 by residual energy and relief-if-removed; removing the bad edge reduced tension near zero.
Limit: Small planted contradiction test; requires larger messy-source replication.
Break/Evolve through context splitting
Mechanism receipt showing incompatible regimes become stable after context splitting.
Claim: Break/Evolve-style context splitting supports the narrow claim that mixed regimes can be separated to reduce tension.
Result: Single global relation tension was 786.8852; context split tension was 0.5270, a 99.933% reduction.
Limit: Synthetic two-regime setup; not evidence of general unsupervised concept discovery.
Coherence vs complexity
Synthetic MDL-style receipt where three contexts beat underfit and overfit alternatives.
Claim: A tension-plus-complexity penalty can prefer a compact context split over both underfit and overfit variants.
Result: K=3 was best under the combined penalty.
Limit: Synthetic data; requires replication on public datasets.
Provenance-weighted coherence
Provenance weighting improves error and catches bad sources in a controlled setup.
Claim: Source provenance can support more coherent aggregation than flat averaging in a bounded setup.
Result: Flat averaging error was 2.3185; provenance-weighted error was 0.7714; median error was 0.1996; suspicion caught 12/12 bad sources.
Limit: Controlled source model; real source correlation and incentives are harder.
Strategic Stabilization Graph
Toy planning graph receipt for reducing tension across selected public research assets.
Claim: A strategic graph can identify high-pressure documentation and proof-bank work without pretending the toy model is the real system.
Result: Weighted tension energy fell from 452.05 to 10.7258.
Limit: Strategic toy model; not a real productivity benchmark.
Bayesian update rigidity under regime change
Bayesian-style test showing decay/forgetting is needed when regimes change.
Claim: Overconfident priors can create rigidity under regime changes unless decay or revision pressure is present.
Result: Decay/forgetting was needed to adapt after the regime shift.
Limit: Synthetic Bayesian-style test; requires public replay artifact.
CIG source correlation matters
CIG-style note showing correlated sources should not be treated as independent votes.
Claim: Dependency-aware consensus is needed because source correlation can inflate apparent agreement.
Result: Source correlation changed confidence behavior compared with flat independent-source assumptions.
Limit: Draft receipt until backed by a public dataset and replay command.
Free-energy-style mixed-regime split
Free-energy-style framing for why context splits can solve mixed regimes.
Claim: Context splitting can reduce residual pressure when one model is forced across incompatible regimes.
Result: Context split solved the mixed-regime pressure in the bounded setup; overconfident priors caused rigidity.
Limit: Framing receipt, not a general theory proof.
Seeded Word-Function parser
Seeded parser line for semantic-frame accuracy under a toy grammar.
Claim: Seeded word-function structure can improve inspectability of parser decisions in a toy/synthetic grammar.
Result: Records semantic-frame accuracy and trace structure for the seeded parser direction.
Limit: Toy/synthetic grammar; not a natural-language capability benchmark.
Learned tension weights
Seeded operator line testing learned tension weights under controlled negatives.
Claim: Learned tension weights can become a useful diagnostic signal when paired with hard negatives.
Result: The receipt captures the lesson that hard-negative mining matters for meaningful tension weights.
Limit: Synthetic grammar and small models; requires benchmark expansion.
Hard-negative mining lesson
Negative examples are required before local tension/failure signals become meaningful.
Claim: Failure localization receipts are weak without hard negatives that pressure the model into mistakes.
Result: Hard negatives exposed local tension behavior that easy negatives did not.
Limit: Lesson receipt; not a published full benchmark.
Seeded Operator LM direction
Planned receipt path for seeded operator language-model experiments.
Claim: Seeded operator structure is a research direction, not yet a benchmark claim.
Result: Next result should report parser/frame accuracy, tension diagnostics, and failure modes.
Limit: No broad claim until public benchmark receipts exist.
TS-pi independent route audit
Audit-line receipt for independent pi routes and block-level tension.
Claim: Independent derivation routes can expose block tension and corruption pressure in a bounded audit.
Result: Block tension highlighted disagreement locations for inspection.
Limit: Audit toy line; route independence must be proven, not assumed.
TS-pi corruption localization and selective repair
Audit-line receipt for identifying corrupted blocks and repairing selectively.
Claim: Block tension can support selective repair in a bounded corruption-localization setup.
Result: Corruption localized to high-tension blocks and selective repair targeted those blocks.
Limit: Requires stricter dependency-aware consensus tests.
Provenance consensus and dependency limits
Consensus note where provenance helps, but dependency-aware consensus remains the hard part.
Claim: Agreement should be weighted by provenance and dependency structure, not treated as simple vote count.
Result: Provenance consensus improved over flat voting, while dependency-aware consensus remained the limitation.
Limit: Needs public CIG dataset and replayable contradiction tests before strong benchmark claims.
Verifier-first language vertical slice
Complete bounded path from human text through MeaningGraph and TS-Reasoner to deterministic response and unified receipt.
Claim: The language substrate and verifier operate together end to end without an external language model or an unverified rendering fallback.
Result: 30/30 decisions passed; unsupported accepts, rejected-claim contamination, renderer support violations, and deterministic replay failures were all zero; receipt generation was 100%.
Limit: Bounded deterministic grammar and reasoning families only. This is not a general-purpose chatbot, broad NLP result, or learned language-generation system.
Habitat v2 signed world and verified planning
Bounded persistent symbolic world with signed evidence, semantic merge, cluster activation, deterministic transitions, causal closure, and verifier-gated plans.
Claim: Within the frozen Habitat grammar and action schemas, TSLC and TS-Reasoner can remember, update, query, derive, and plan without unsupported affirmative output or rejected-memory contamination.
Result: Functional decisions were 100 ACCEPT and 20 REJECT. Adversarial decisions were 30 ACCEPT, 12 REPAIR, and 60 REJECT. Every required hard-gate metric was zero and receipt generation was 100%.
Limit: Deterministic symbolic toy habitat only: bounded grammar, four-valued state, turn-order events, positive causal rules, fixed action schemas, open-topology movement, and process-local memory. Not a general chatbot, biological model, or real-world knowledge system.
Habitat v3 verifier-first symbolic agent loop
Bounded deterministic agent cycle with persistent goals, tension-directed effort, explicit topology, verified execution, effect checking, replanning, reflection, lesson approval, multi-agent scheduling, and whole-run replay.
Claim: Within the frozen Habitat v3 grammar, environment, action schemas, and execution budgets, the system can pursue explicit goals through verified symbolic actions without treating desire, tension, plans, expected effects, or reflection as proof authority.
Result: The final functional and adversarial reports have zero unsupported accepts, goal satisfactions, action authorizations/executions, plan steps, unverified commits, contamination, silent overwrites, topology duplicates, tension-as-evidence violations, stale executions, missed replans, unauthorized lesson activations/transfers, scheduling nondeterminism, renderer violations, replay failures, or execution errors. Receipt generation is 100%.
Limit: In-process simulated environment only; bounded regex language, topology, goals, planning, tension, reflection, agents, events, actions, receipts, and execution budgets. No external model, network knowledge, real-world actuation, self-modifying code, biological equivalence, or AGI claim.
TS-003
Local relaxation can generate global coherence
Claim
Local relaxation supports the narrow claim that local constraint updates can reduce global tension in a bounded graph.
Setup
Fixed-seed quadratic constraint graph with local update rules and tension telemetry.
Result
Tension fell from 156.216618 to 11.862658, a 92.4063% reduction.
Limit
Quadratic constraint graph only; this is not a full-scale reasoning benchmark.
Reproduce
Run the fixed-seed local relaxation script and compare initial/final tension.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
TS-CoreRelated artifact/model
artifact pending
TS-004
Attractors emerge from local constraint relaxation
Claim
Local relaxation can recover stable attractor states in a toy binary graph.
Setup
Hopfield-style binary graph with 30% noise and a context-bias condition.
Result
30% noise recovery was around 87%; context bias improved recovery to about 98%.
Limit
Toy attractor graph; requires replication beyond synthetic binary states.
Reproduce
Replay the seeded attractor graph with and without context bias.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
TS-CoreRelated artifact/model
artifact pending
TS-005
Contradiction localizes as residual/provenance tension
Claim
Contradictions can localize as residual graph tension in a provenance-aware toy graph.
Setup
9 nodes and 16 constraints with one planted contradiction.
Result
The planted contradiction ranked #1 by residual energy and relief-if-removed; removing the bad edge reduced tension near zero.
Limit
Small planted contradiction test; requires larger messy-source replication.
Reproduce
Replay the 9-node graph, compute residual energy by edge, remove the top edge, and compare tension.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
CIGRelated artifact/model
artifact pending
TS-006
Break/Evolve through context splitting
Claim
Break/Evolve-style context splitting supports the narrow claim that mixed regimes can be separated to reduce tension.
Setup
Two incompatible regimes: y-x approximately +2 and y-x approximately -2.
Result
Single global relation tension was 786.8852; context split tension was 0.5270, a 99.933% reduction.
Limit
Synthetic two-regime setup; not evidence of general unsupervised concept discovery.
Reproduce
Fit one global relation, then split by context and compare tension plus complexity penalty.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
TS-CoreRelated artifact/model
artifact pending
TS-008
Coherence vs complexity
Claim
A tension-plus-complexity penalty can prefer a compact context split over both underfit and overfit variants.
Setup
K=1..6 context candidates evaluated under MDL/tension plus complexity penalty.
Result
K=3 was best under the combined penalty.
Limit
Synthetic data; requires replication on public datasets.
Reproduce
Replay K=1..6 context scoring and verify the minimum at K=3.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
TS-CoreRelated artifact/model
artifact pending
TS-009
Provenance-weighted coherence
Claim
Source provenance can support more coherent aggregation than flat averaging in a bounded setup.
Setup
Mixed-source aggregation with trusted, noisy, and adversarial sources.
Result
Flat averaging error was 2.3185; provenance-weighted error was 0.7714; median error was 0.1996; suspicion caught 12/12 bad sources.
Limit
Controlled source model; real source correlation and incentives are harder.
Reproduce
Replay the seeded aggregation and suspicion-scoring script.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
CIGRelated artifact/model
artifact pending
TS-011
Strategic Stabilization Graph
Claim
A strategic graph can identify high-pressure documentation and proof-bank work without pretending the toy model is the real system.
Setup
Selected proof-bank docs, Start Here page, Obsidian vault, research statement, and demo video as nodes in a stabilization graph.
Result
Weighted tension energy fell from 452.05 to 10.7258.
Limit
Strategic toy model; not a real productivity benchmark.
Reproduce
Replay the weighted stabilization graph with the selected asset nodes.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
boggersthefish-siteRelated artifact/model
artifact pending
TS-013
Bayesian update rigidity under regime change
Claim
Overconfident priors can create rigidity under regime changes unless decay or revision pressure is present.
Setup
Synthetic regime-change sequence with fixed and decayed priors.
Result
Decay/forgetting was needed to adapt after the regime shift.
Limit
Synthetic Bayesian-style test; requires public replay artifact.
Reproduce
Replay fixed-prior and decayed-prior update curves under a regime switch.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
CIGRelated artifact/model
artifact pending
TS-014
CIG source correlation matters
Claim
Dependency-aware consensus is needed because source correlation can inflate apparent agreement.
Setup
Claim/evidence graph with correlated source branches and confidence updates.
Result
Source correlation changed confidence behavior compared with flat independent-source assumptions.
Limit
Draft receipt until backed by a public dataset and replay command.
Reproduce
Build correlated and independent source graphs, then compare confidence propagation.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
CIGRelated artifact/model
artifact pending
TS-015
Free-energy-style mixed-regime split
Claim
Context splitting can reduce residual pressure when one model is forced across incompatible regimes.
Setup
Mixed regimes scored with pressure, fit, and complexity terms.
Result
Context split solved the mixed-regime pressure in the bounded setup; overconfident priors caused rigidity.
Limit
Framing receipt, not a general theory proof.
Reproduce
Replay mixed-regime scoring with and without context split and prior decay.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
TS-CoreRelated artifact/model
artifact pending
TS-016
Seeded Word-Function parser
Claim
Seeded word-function structure can improve inspectability of parser decisions in a toy/synthetic grammar.
Setup
Synthetic grammar with seeded semantic-frame parser and fixed evaluation prompts.
Result
Records semantic-frame accuracy and trace structure for the seeded parser direction.
Limit
Toy/synthetic grammar; not a natural-language capability benchmark.
Reproduce
Replay fixed prompts against seeded parser exports and compare semantic frames.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
TensionLMRelated artifact/model
TS-018
Learned tension weights
Claim
Learned tension weights can become a useful diagnostic signal when paired with hard negatives.
Setup
Seeded operator/word-function examples with learned local tension weights.
Result
The receipt captures the lesson that hard-negative mining matters for meaningful tension weights.
Limit
Synthetic grammar and small models; requires benchmark expansion.
Reproduce
Train with and without hard negatives and compare failed-step/tension localization.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
TensionLMRelated artifact/model
TS-019
Hard-negative mining lesson
Claim
Failure localization receipts are weak without hard negatives that pressure the model into mistakes.
Setup
Proof/language traces compared under easy negatives and hard negatives.
Result
Hard negatives exposed local tension behavior that easy negatives did not.
Limit
Lesson receipt; not a published full benchmark.
Reproduce
Run the same detector with easy negatives and hard negatives, then compare localization quality.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
Proof RankerRelated artifact/model
TS-020
Seeded Operator LM direction
Claim
Seeded operator structure is a research direction, not yet a benchmark claim.
Setup
Operator-token traces, local tension telemetry, and fixed prompts.
Result
Next result should report parser/frame accuracy, tension diagnostics, and failure modes.
Limit
No broad claim until public benchmark receipts exist.
Reproduce
Pending: publish fixed prompts, checkpoint, seed, and export script.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
TensionLMRelated artifact/model
TS-021
TS-pi independent route audit
Claim
Independent derivation routes can expose block tension and corruption pressure in a bounded audit.
Setup
Multiple independent pi routes with block tension and provenance notes.
Result
Block tension highlighted disagreement locations for inspection.
Limit
Audit toy line; route independence must be proven, not assumed.
Reproduce
Replay independent route generation and compare block tension by segment.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
CIGRelated artifact/model
artifact pending
TS-022
TS-pi corruption localization and selective repair
Claim
Block tension can support selective repair in a bounded corruption-localization setup.
Setup
Injected corruption in a pi audit route with block-level consensus checks.
Result
Corruption localized to high-tension blocks and selective repair targeted those blocks.
Limit
Requires stricter dependency-aware consensus tests.
Reproduce
Inject known block corruption, run audit tension, repair top blocks, and compare consensus.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
CIGRelated artifact/model
artifact pending
TS-023
Provenance consensus and dependency limits
Claim
Agreement should be weighted by provenance and dependency structure, not treated as simple vote count.
Setup
Claim/evidence graph with independent and dependent sources, contradictions, and provenance edges.
Result
Provenance consensus improved over flat voting, while dependency-aware consensus remained the limitation.
Limit
Needs public CIG dataset and replayable contradiction tests before strong benchmark claims.
Reproduce
Create a small claim/evidence graph, introduce conflicting evidence, and verify dependency-aware confidence updates.
Date
2026-05-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
CIGRelated artifact/model
artifact pending
TS-024
Verifier-first language vertical slice
Claim
The language substrate and verifier operate together end to end without an external language model or an unverified rendering fallback.
Setup
Thirty fixed relational, Boolean, ambiguity, unsupported-inference, and planning cases, each replayed three times through the deterministic bridge and structured verifier.
Result
30/30 decisions passed; unsupported accepts, rejected-claim contamination, renderer support violations, and deterministic replay failures were all zero; receipt generation was 100%.
Limit
Bounded deterministic grammar and reasoning families only. This is not a general-purpose chatbot, broad NLP result, or learned language-generation system.
Reproduce
Clone TS-Reasoner-v0 and ts-chat-language as siblings, run ./scripts/run_vertical_slice.sh, then python -m ts_vertical_slice.evaluation.
Date
2026-06-20 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
ts-chat-languageRelated artifact/model
TS-025
Habitat v2 signed world and verified planning
Claim
Within the frozen Habitat grammar and action schemas, TSLC and TS-Reasoner can remember, update, query, derive, and plan without unsupported affirmative output or rejected-memory contamination.
Setup
120 functional cases across six required families plus 102 adversarial cases across 17 attack categories, each replayed twice through TurnReceipt v2.
Result
Functional decisions were 100 ACCEPT and 20 REJECT. Adversarial decisions were 30 ACCEPT, 12 REPAIR, and 60 REJECT. Every required hard-gate metric was zero and receipt generation was 100%.
Limit
Deterministic symbolic toy habitat only: bounded grammar, four-valued state, turn-order events, positive causal rules, fixed action schemas, open-topology movement, and process-local memory. Not a general chatbot, biological model, or real-world knowledge system.
Reproduce
Clone the habitat-v2 branches as siblings, run ./scripts/run_vertical_slice.sh, then PYTHONPATH=../TS-Reasoner-v0:. .venv/bin/python -m ts_vertical_slice.habitat_evaluation.
Date
2026-06-21 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
ts-chat-languageTS-026
Habitat v3 verifier-first symbolic agent loop
Claim
Within the frozen Habitat v3 grammar, environment, action schemas, and execution budgets, the system can pursue explicit goals through verified symbolic actions without treating desire, tension, plans, expected effects, or reflection as proof authority.
Setup
240 frozen functional cases, 240 frozen adversarial cases, ten continuous demonstrations, 633 TS-Reasoner tests, 109 TSLC tests, and all 30/165/120/102 legacy evaluation cases.
Result
The final functional and adversarial reports have zero unsupported accepts, goal satisfactions, action authorizations/executions, plan steps, unverified commits, contamination, silent overwrites, topology duplicates, tension-as-evidence violations, stale executions, missed replans, unauthorized lesson activations/transfers, scheduling nondeterminism, renderer violations, replay failures, or execution errors. Receipt generation is 100%.
Limit
In-process simulated environment only; bounded regex language, topology, goals, planning, tension, reflection, agents, events, actions, receipts, and execution budgets. No external model, network knowledge, real-world actuation, self-modifying code, biological equivalence, or AGI claim.
Reproduce
Clone the habitat-v3-agent-loop branches as siblings, run scripts/run_habitat_v3.sh, then PYTHONPATH=../TS-Reasoner-v0:$PYTHONPATH python3 -m ts_vertical_slice.habitat_v3_evaluation.
Date
2026-06-21 Source: Manual proof-bank receipt record Displayed as the receipt record update date; external artifact links are shown separately where available.
Related repo
ts-chat-language