Hypernym Infinite Memory / Librarian · CTO research board · Direct IP only
Memory control plane for small-model long-memory inference.
True north: prove that a 4,096-token-class ~3B local model can use controller-curated memory to handle long-memory pressure, isolated recall, and provenance-controlled retrieval at materially lower serving cost than brute-force long context.
Current decision: continue research and build V3 around selected-current payloads, controller-owned metadata, and tail-anchored output contracts. v0.61 shows the tail-contract and tail-schema variants passed both 1024 and 2048 pressure; the original-control 2048 row hit one transport disconnect.
Why this page exists: it is not a pretty dashboard. It is the current decision record: what we can say, what we cannot say, and what the next eval must prove before a CTO should change direction.
QuestionCan a 4,096-context-class small model behave like a useful long-memory system when a controller curates memory?
Evidence80 live rows, 1,199,692 prompt tokens, 84.6% true-north score.
ConstraintFailed-slice rerun coherence stayed at 25.0%; the weak cluster is real.
Current testGranite Pilots 8192 passed with 210,631 reported prompt tokens; the next 16384 probe was rejected at the gateway with HTTP 413 in 0.4259s and undocked True.
Largest observed pressure
198,888
48.6x a 4,096-token native context class. Measured successful pressure, not a base-context increase.
Transport liveness
3 / 4
Known compact transport probes that completed under the request-path wall.
Pocket semantic score
83.3%
Latest pocket-generalization semantic mean across completed rows.
Personal-memory score
84.6%
v0.51 live true-north score across the complete 80-row matrix.
v0.53 semantic score
46.3%
Format/schema/pressure diagnostic. This is a constraint finding, not a demo score.
v0.54 semantic score
85.2%
Runtime-selected minimal-payload diagnostic. This is the strongest V3 direction signal.
v0.55 pressure score
11.1%
Semantic pass under 1024 pressure in deterministic envelope live run. This is the next hard constraint.
v0.57 HTTP 200
9 / 9
Timeout-aware shard completed every row, including 2048-pressure rows, without 503 contamination.
v0.58 HTTP 200
0 / 1
First row stopped after repeated chat-slot busy responses; health was OK, but no quality row completed.
v0.59 HTTP 200
14 / 15
Idle-gated retry avoided the busy cascade; stopped on one fast HTTP 500 Invalid input batch.
v0.60 HTTP 200
6 / 6
Focused agent-loop replay completed all rows. Original controller-expands passed at 1024/2048.
v0.61 HTTP 200
5 / 6
Tail-contract and tail-schema variants passed 4/4 scored rows; original-control 2048 disconnected.
Pilots pressure pass
1,024
Highest strict pass band on the session endpoint; max probe prompt tokens 24,584, max probe latency 58.1629s.
Pilots 2048 pass
52,933
Reported prompt tokens in the independent 2048 extension; elapsed 55.1825s, strict/semantic 100%, forbidden hits 0.
Pilots 4096 pass
105,498
Reported prompt tokens in the independent 4096 extension; elapsed 106.0359s, strict/semantic 100%, forbidden hits 0.
Pilots 8192 pass
210,631
Reported prompt tokens in the independent 8192 extension; elapsed 196.4813s, strict/semantic 100%, forbidden hits 0.
16384 ceiling
HTTP 413
The 16384 probe hit nginx Request Entity Too Large in 0.4259s. This is a request-admission ceiling, not a memory-quality failure.
What This System Is
A controller-mediated memory layer around a small local model. The model is not simply given an infinite prompt; the controller selects, scopes, and verifies memory payloads so the model can answer from long-memory state without brute-force long-context serving.
- Serving target: local/self-hosted Granite GGUF through the isolated direct-IP lane.
- Security target: per-tenant memory isolation, active binding precedence, revoked/stale memory rejection, and empty-result honesty.
- Business target: reduce the cost and reliability penalty of long-memory inference across many users.
What Is Actually Strong
- Pressure handling: successful prompts have exceeded native context by tens of times.
- Structured output: strict JSON / digest-handle patterns are working in prepared evals.
- Control-plane safety: empty result, duplicate conflict, and ambiguous candidate abstention tests are now explicit repeatable harnesses.
- Product-shaped eval: v0.51 live covered research, story canon, relationship boundaries, personal preferences, and long-running agent continuity.
- Cost guard: current prompt-injection runs are dry-run or direct self-hosted; frontier generation endpoints are blocked by policy audit.
Current Verdict, In Human Terms
Deck-safe claim
Small local models can be made materially more useful for long-memory workflows when memory selection, rejection, and provenance handling are moved into a controller layer instead of being left to raw prompt text.
Hard proof
v0.51 completed 80/80 live rows with 80 HTTP 200 responses, 100.0% parse success, 100.0% forbidden-fact absence, and 95.4% required-fact recall.
Hard weakness
The system is not yet robust enough for arbitrary exact long-term recall: failed-slice rerun coherence was 25.0%, especially in update precedence and long-running agent entity binding.
Research action
Do fewer visuals. After each major run, publish only this form: question, evidence, limits, next decision. v0.61 moves the agent-loop direction from "do not shorten" to "keep the original prompt body, but add the output contract at the tail."
Current Eval Harness State
The latest work extended the Granite Pilots lane from the earlier 0/256/1024 ladder through independent 2048, 4096, and 8192 pressure bands, then measured the next ceiling: 16384 was rejected by nginx with HTTP 413 before model inference. The full objective still requires the 27-row first-live subset after a lease/quiet-window gate, but the Pilots evidence now proves a private one-handle session can recall current facts, reject controls, and undock cleanly at 210,631 reported prompt tokens.
| Artifact | Status | Hard fact | Why it matters |
|---|---|---|---|
| Objective audit | not_complete_live_gate_blocked | 13 satisfied · 2 partial live · 6 blocked | Goal remains incomplete until live evidence covers story, relationship, psychology, agent, Q2 sequential state, and full-domain threshold pressure. |
| Final validation | pass | failures 0; warnings 0; suite index verified 1; lock guard verified 1; adaptive plan verified 1 | The manifest, packet, comparison, checklist, first-live subset, finalizer, full-suite artifact index, Granite Pilots session lock guard, and adaptive pressure scheduler agree. |
| Granite Pilots router | dry_run_plan_only | handle velvet-starfall-crusader-9; max active pilots 2; endpoint touched False | New session-aware route uses one durable handle and enforces the shared two-active-pilot policy before live calls. |
| Granite Pilots lock guard | verified | router dry_run_plan_only; probe dry_run_pass; pressure pass; endpoint touched false | All live-capable Pilots tools now share a nonblocking local session-handle lock plus the remote berth gate. This prevents two local runs from accidentally consuming multiple berths while another researcher is active. |
| Granite Pilots adaptive pressure | ready_no_live_endpoint | pass ceiling 8,192; gateway ceiling 16,384; target safe single-shot False | Future pressure runs are now mode-gated: replay 8192 safely, recheck 16384 only as an explicit gateway test, and use chunked session continuation for >8192 accumulated pressure. Safe replay command: bash research/tracks/hypernym-infinite-mim/forge_runner.sh run-granite-pilots-pressure-ladder --live --run-id <run>_p8192_replay --bands 8192 --timeout 300 |
| Granite Pilots live probe | pass | turns 9; semantic 100.0%; strict 100.0%; undocked True | Measured live one-handle session memory: five current private facts recalled across research/story/relationship/psychology/agent domains, stale/rejected/foreign controls absent, berth freed after use. |
| Granite Pilots pressure ladder | pass | bands [0, 256, 1024]; highest strict 1,024; highest semantic 1,024; max prompt tokens 24,584 | Measured live session recall under rising pressure on the same durable handle. 1024 band passed strict JSON recall with no stale/rejected/foreign leakage and then undocked. |
| Granite Pilots 2048 extension | pass | band [2048]; highest strict 2,048; highest semantic 2,048; max prompt tokens 52,933 | Independent single-band extension. 2048 passed strict JSON recall in 55.1825s with no stale/rejected/foreign leakage and then undocked. |
| Granite Pilots 4096 extension | pass | band [4096]; highest strict 4,096; highest semantic 4,096; max prompt tokens 105,498 | Independent single-band extension. 4096 passed strict JSON recall in 106.0359s with no stale/rejected/foreign leakage and then undocked. |
| Granite Pilots 8192 extension | pass | band [8192]; highest strict 8,192; highest semantic 8,192; max prompt tokens 210,631 | Independent single-band extension. 8192 passed strict JSON recall in 196.4813s with no stale/rejected/foreign leakage and then undocked. |
| Granite Pilots 16384 ceiling | partial_or_fail | band [16384]; probe HTTP 413; elapsed 0.4259s; undocked True | Gateway/request-size ceiling. Setup turns were admitted, but the 16384 probe was rejected by nginx before model inference, so this is transport-admission evidence. |
| Next-live evidence matrix | ready_but_gate_blocked | 6 gaps · 27 rows/turns · Q1 4, Q2 15, Q4 8 | Each blocked objective gap has a specific label or turn trace and a stop condition for the next admitted run. |
| Indexed dry run | dry_run_pass | steps: validate-eval-suite, q1-cross-domain-tail-contract, q2-multi-turn-personal-memory, q3-tenant-foreign-boundary-regression, q4-sensitive-preference-boundary-abstention | Harness proof only. It exercises Q1/Q2/Q3/Q4 without endpoint traffic. |
| Indexed live refusal | blocked_by_gate | live endpoint touched: False | Under v0.66, full live suite refuses before touching the direct-IP endpoint. |
| Full-suite finalizer | certification_artifacts_ready | Q3 safety-control: True; Q1/Q2/Q4 live-like: {'q1_scores': False, 'q2_scores': False, 'q4_scores': False} | Consumes the suite artifact index and produces threshold plus version-comparison artifacts. Dry-index use remains harness evidence only. |
| Version packet | v3-candidate | fingerprinted files: 33 | Future V3 or V-next can be rerun against the same objective and compared without changing the contract. |
| Comparator | comparison_scaffold_ready_no_candidate_artifacts | candidate Q1/Q2/Q3/Q4 artifacts supplied: none | Correctly refuses to create a capability delta until live candidate artifacts exist. |
| Launch checklist | ready_but_gate_blocked | steps: 10; first-live minimum rows: 27 | Operator path is explicit: status, audit, validation, dry-run, gate, first-live subset, threshold finalizer, live suite, version compare. |
Artifact index command trace: bash research/tracks/hypernym-infinite-mim/forge_runner.sh compare-version-eval-results --run-id 20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_compare --candidate-version <version> --q1-scores research/tracks/hypernym-infinite-mim/results/v0.63-unscored-domain-drain-resume/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_q1/scores.json --q2-scores research/tracks/hypernym-infinite-mim/results/q2-multi-turn-personal-memory-session/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_q2/scores.json --q3-scores research/tracks/hypernym-infinite-mim/results/q3-tenant-foreign-boundary-regression-plan/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_q3_plan/plan.json --q4-scores research/tracks/hypernym-infinite-mim/results/q4-sensitive-preference-boundary-abstention/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_q4/scores.json
Threshold command trace: bash research/tracks/hypernym-infinite-mim/forge_runner.sh analyze-threshold-boundary --run-id 20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_threshold --live-scores research/tracks/hypernym-infinite-mim/results/v0.63-unscored-domain-drain-resume/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_q1/scores.json --live-scores research/tracks/hypernym-infinite-mim/results/q4-sensitive-preference-boundary-abstention/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_q4/scores.json --q2-live-scores research/tracks/hypernym-infinite-mim/results/q2-multi-turn-personal-memory-session/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1_q2/scores.json
Evidence Ladder
| Layer | Status | Evidence | Meaning |
|---|---|---|---|
| Transport under pressure | measured live | 198,888 prompt tokens, 48.6x native class | The endpoint path can accept and complete much larger prompt pressure than the model's native context class. |
| Quality floor | measured live | Semantic 100.0%; provenance 80.0% | Basic structured recall can work under controlled pressure. |
| Pocket generalization | measured mixed | 825,958 prompt tokens total; exact provenance 28.6% | There is a real exactness pocket, but the system drifts outside that pocket. |
| Empty-result contract | prepared dry-run | v0.47 verifier 100.0% | The harness now tests that no memory selected means no invented placeholder memory. |
| Duplicate conflict rejection | prepared dry-run | v0.48 page safe 100.0%; bundle safe 83.3% | The harness now tests stale/shadow duplicate binding conflicts before live launch. |
| Ambiguous candidate abstention | prepared dry-run | v0.49 abstention modes 100%; aggregate 95.8% | The harness now tests that candidate hints do not become recalled memory without controller selection. |
| Personal entity coherence | measured live | v0.51 true-north 84.6%; coherence 81.2%; rows 80/80; prompt tokens 1,199,692 | First complete live eval shaped like single-user long-term memory across research, story, relationship, psychology, and agent workflows. |
| Timeout-aware pressure retry | measured live | v0.57 strict true-north 74.1%; semantic true-north 74.1%; rows 9/9; prompt tokens 370,933 | Serving contamination from v0.56 was fixed; quality still drops at 1024/2048 pressure. |
| Granite Pilots session pressure | measured live | bands [0, 256, 1024]; highest strict pass 1,024; 1024-band prompt tokens 24,584; max probe latency 58.1629s | The session endpoint preserved five current private facts and rejected stale/rejected/foreign controls under rising pressure on a persistent one-handle memory lane. |
| Granite Pilots 2048 extension | measured live | band [2048]; prompt tokens 52,933; latency 55.1825s; strict/semantic pass 2,048/2,048 | The next pressure boundary passed independently with the same durable handle, max two-pilot policy, no forbidden hits, and mandatory undock. |
| Granite Pilots 4096 extension | measured live | band [4096]; prompt tokens 105,498; latency 106.0359s; strict/semantic pass 4,096/4,096 | The next pressure boundary passed independently with the same durable handle, max two-pilot policy, no forbidden hits, and mandatory undock. |
| Granite Pilots 8192 extension | measured live | band [8192]; prompt tokens 210,631; latency 196.4813s; strict/semantic pass 8,192/8,192 | The next pressure boundary passed independently with the same durable handle, max two-pilot policy, no forbidden hits, and mandatory undock. |
| Granite Pilots 16384 ceiling | measured live | band [16384]; probe status HTTP 413; elapsed 0.4259s; prompt tokens n/a | The current shared gateway rejects the next band before inference. The proven memory pass ceiling is therefore 8192 until request-size limits are changed. |
| Current-fact locking retry | serving blocked | v0.58 attempted 1 row, HTTP 200 0/1, stop reason non_200_after_v058-research_update_precedence-baseline_verbatim_current_facts-p1024 | No memory-quality evidence yet. `/health` can be OK while the single chat slot is busy. |
| Idle-gated fact-locking retry | partial live | v0.59 attempted 15 rows, HTTP 200 14/15, strict/semantic true-north 100.0%/100.0% | Chat-slot idle gate worked; completed rows were clean, but one agent-loop locked-ID row hit HTTP 500 Invalid input batch. |
| Agent-loop input-batch hardening | measured live | v0.60 completed 6 rows, HTTP 200 6/6; original replay coherence 100.0% | The v0.59 500 did not reproduce. Original controller-expands passed; shorter hardened/budgeted variants returned plain text and failed parse. |
| Agent-loop tail output contract | measured live | v0.61 completed 6 rows, HTTP 200 5/6; tail-contract coherence 100.0%; tail-schema coherence 100.0% | Tail-anchored output contract fixed the v0.60 short-prompt plain-text collapse on scored rows. Original-control 2048 had one remote disconnect and should be treated as a serving boundary, not a memory-quality pass. |
v0.51 Live Findings
- Transport: 80/80 rows returned HTTP 200; no non-200 rows recorded.
- Format control: parse success 100.0%; entity id exactness 92.5%.
- Memory safety: forbidden-fact absence 100.0%; no stale/forbidden phrase leakage in scored answers.
- Recall: required fact recall 95.4%; coherence pass 81.2%.
- Focused audit: rerunning the original failed slices produced 25.0% coherence while keeping parse and forbidden-fact absence at 100.0%/100.0%.
v0.51 Constraints
- Latency: p50 2.4586s; p90 89.9753s; max 93.018s across completed rows.
- Pressure behavior: 1024-band rows were slow but scored best: coherence 100.0%.
- Weakest domain: long-running agent task at 56.2%; failures were mostly entity-id exactness and temporal update handling, not fact leakage.
- Temporal updates: some rows remembered core facts but missed the newest update condition; V3 needs explicit update precedence.
Scenario Scorecard
This is the part that should drive product and CTO decisions. It shows where the memory system is already useful, and where V3 should focus.
| Scenario | Coherence | Meaning | Decision |
|---|---|---|---|
| Relationship boundary memory | 100.0% | Best current product fit: safety-sensitive stated preferences and boundaries stayed clean. | candidate demo lane |
| Story-world canon | 87.5% | Strong for creative continuity; still needs explicit supersession rules for canon changes. | keep testing |
| Personal preference / psychology | 81.2% | Useful but must be conservative because wrong continuity can feel personally invasive. | guard heavily |
| Research program continuity | 81.2% | Promising for long-running research, but update precedence is the blocker. | V3 target |
| Long-running agent task | 56.2% | The main failure cluster: facts are often present, but entity binding and task identity drift. | do not demo yet |
v0.52 Targeted Follow-Up
v0.52 is live evidence now. It targeted the v0.51 failure cluster: update precedence and stable memory-key/entity binding under stale and foreign near-duplicate pressure.
- Composite live result: 24 rows, 24 HTTP 200 after focused replacement, 74.1% true-north score.
- Safety preserved: stale-fact absence 100.0%; foreign-fact absence 100.0%.
- Unexpected result: prose memory coherence 91.7%; structured control-plane coherence 66.7%.
- Pressure cliff: 1024-pressure coherence 50.0%; 64-pressure coherence 100.0%.
Convergence Standard
Future pages should only publish after the run has a named hypothesis, stored scores, and a convergence/review artifact. Raw visual output without this context is not a research result.
- Every page must state the research question and why it matters to true north.
- Every chart/table must have a decision attached to it.
- Every institutional claim must distinguish measured fact, inference, and open question.
- Dry-runs may validate harness logic; they must not be framed as live model capability.
v0.53 Format / Schema / Pressure Diagnostic
v0.53 is live evidence. It tested whether the v0.52 failure cluster was caused by record format, output schema, or pressure budget. The result is constraining: every row returned HTTP 200, but quality fell sharply under the diagnostic matrix.
| Slice | Measured result | CTO meaning |
|---|---|---|
| All rows | 36 rows, 36 HTTP 200, 811,194 prompt tokens | Transport is not the blocker in this run; quality and binding are. |
| Strict true north | 25.9% | The strict machine-contract path is not yet robust enough for production memory claims. |
| Semantic true north | 46.3% | Even allowing field recovery, many rows lose rejected-key completeness or admit stale/foreign facts. |
| Best memory format | prose coherence 50.0%; compact KV 25.0%; verbose JSONL 8.3% | More explicit structure did not solve the problem. Prose remains the current reliable baseline. |
| Output contract | strict JSON coherence 55.6%; field-tagged coherence 0.0% | Relaxing the answer envelope did not recover coherence. Use strict JSON plus verifier/repair, not looser text. |
| Pressure cliff | 0-pressure semantic pass 72.2%; 1024-pressure semantic pass 33.3% | Pressure still materially degrades reliable memory behavior. |
v0.54 Runtime-Selected Payload Diagnostic
v0.54 is live evidence. It tested whether the system improves when the controller sends selected memory instead of asking the model to sort through stale and foreign fact bodies. This is the strongest post-v0.53 result and should drive V3.
| Slice | Measured result | CTO meaning |
|---|---|---|
| All rows | 36 rows, 36 HTTP 200, 822,190 prompt tokens | Transport stayed clean; this is a quality/control-plane signal. |
| Strict true north | 77.8% | Recovered sharply from v0.53 strict true-north 25.9%. |
| Semantic true north | 85.2% | Recovered sharply from v0.53 semantic true-north 46.3%. |
| Noisy full payload | coherence 58.3%; semantic 58.3% | Letting the model see stale and foreign fact bodies is worse. |
| Selected current only | coherence 100.0%; semantic 100.0% | Best quality result. Runtime selection works when audit handles are not required. |
| Selected + rejected handles | coherence 66.7%; semantic 91.7% | Promising audit envelope, but exact rejected-key set formatting still needs repair. |
| Verifier repair | repair attempted 13.9%; repair improved 0.0% | The current repair prompt does not fix failures. V3 needs deterministic repair or a stronger constrained decoder. |
v0.55 Deterministic Envelope Diagnostic
v0.55 is live evidence. It tested whether the controller should fill rejected handles and identity metadata instead of making the model emit the full memory-control envelope. The answer is mixed: controller-filled envelopes are better than model-filled full JSON, but pressure recall remains the blocker.
| Slice | Measured result | CTO meaning |
|---|---|---|
| All rows | 18 rows, 18 HTTP 200, 363,059 prompt tokens | Transport stayed clean; failures are quality/control-plane under pressure. |
| Strict true north | 29.6% | Deterministic envelope alone did not recover v0.54 quality. |
| Semantic true north | 33.3% | Semantic recovery is also low because current-fact recall collapses under pressure. |
| Model fills full JSON | coherence 0.0%; semantic 16.7%; rejected-key recall 50.0% | Do not ask the model to own the full audit envelope. Strict coherence was zero in this run. |
| Controller fills rejected handles | coherence 50.0%; semantic 50.0% | Better than full model envelope, but still vulnerable when the model must emit IDs and current facts under pressure. |
| Controller fills all IDs | coherence 66.7%; semantic 66.7%; rejected-key recall 100.0% | Best envelope mode. Controller should own identity and rejection metadata. |
| Pressure split | 0-pressure semantic 77.8%; 1024-pressure semantic 11.1% | The next optimization is not more envelope filling; it is preserving current-fact recall under pressure. |
v0.56 Selected-Current Controller Envelope Pressure Recall
v0.56 is live evidence, but not a completed quality matrix. It tested the v0.54+v0.55 synthesis: selected-current payloads plus controller-owned identity/rejection metadata at 0, 1024, and 2048 pressure. The important result is a serving boundary: the 2048-pressure row exceeded the 150s client timeout and the backend continued processing it, so subsequent sequential rows got single-flight 503 Backend busy.
| Slice | Measured result | CTO meaning |
|---|---|---|
| Completed matrix | 27 attempted rows, 2 HTTP 200, 41,539 prompt tokens counted from successful calls | The matrix ran, but quality conclusions are valid only for the first two completed rows. |
| 0-pressure row | coherence 100.0%; semantic 100.0% | The selected-current/controller envelope path works cleanly without pressure in the observed slice. |
| 1024-pressure row | current-fact recall 66.7%; semantic 0.0% | At 1024 pressure, exact IDs and rejected handles stayed correct, but one current fact was replaced by a weaker memory-key statement. |
| 2048-pressure row | first 2048 row timed out after 150s; later rows returned 503 busy | This is a serving-control boundary, not a memory-quality verdict. V3 harness needs cancellation-aware timeout, retry-after sleep, or per-row cooldown. |
| Rejected handles | rejected-key recall 100.0%; exact-clean 100.0% over completed rows | Controller-owned rejected handles remained exact in completed rows; pressure recall, not metadata exactness, is the unresolved issue. |
v0.57 Timeout-Aware Sharded Pressure Retry
v0.57 is live evidence. It reran the v0.56 question with a smaller 9-row matrix, 300s timeout, cooldowns, and stop-after-timeout behavior. This separates serving control from memory quality: the serving path completed cleanly, but recall quality remained pressure-dependent.
| Slice | Measured result | CTO meaning |
|---|---|---|
| Serving control | 9/9 rows completed, 9/9 HTTP 200, stop_reason None, partial False | The v0.56 503 cascade was harness/serving contamination, not proof that 2048 pressure cannot complete. |
| Token pressure | 370,933 prompt tokens and 1,199 completion tokens across 9 rows | The run is a real pressure eval over the direct-IP lane, not a dry-run or local-only verifier result. |
| Overall quality | strict true-north 74.1%; semantic true-north 74.1%; coherence 77.8% | Useful memory behavior exists, but it is not yet a production-grade arbitrary recall guarantee. |
| 0 pressure | coherence 100.0%; semantic 100.0% | The selected-current/controller-owned envelope is clean when not under added pressure. |
| 1024 pressure | coherence 66.7%; semantic 66.7%; current-fact recall 100.0% | At 1024 pressure, current facts remained present, but one row hit length/parse/entity-output failure. |
| 2048 pressure | coherence 66.7%; semantic 66.7%; current-fact recall 77.8% | 2048-pressure rows can complete under a 300s wall, but research-update precedence lost current facts in one row. |
| Safety invariants | stale-fact absence 100.0%; foreign-fact absence 100.0% | The most important safety invariant stayed clean across this retry: stale and foreign facts were not admitted. |
v0.58 Current-Fact Locking + Output Budget
v0.58 is staged, approved, and dry-run validated, but the first live attempt did not reach quality scoring. The first row returned 503 Backend busy on all four attempts. Preflight and final /health were 200 OK, which means health is not sufficient proof that the single inference slot is idle.
| Slice | Measured result | CTO meaning |
|---|---|---|
| Dry-run | 18/18 rows, strict/semantic true-north 100% | Harness/scoring mechanics are valid; this is not live model evidence. |
| Live attempt | 1/18 rows attempted, 0/1 HTTP 200, stop reason non_200_after_v058-research_update_precedence-baseline_verbatim_current_facts-p1024 | The runner stopped safely before contaminating the matrix. |
| First-row response | HTTP 503 Backend busy after configured retries | The endpoint can be healthy while the chat slot is occupied. Retry must add chat-slot idle probing or a longer cool-down. |
| Quality score | not available | Do not claim fact-ID locking improved recall until a clean live matrix exists. |
v0.59 Chat-Slot Idle-Gated Fact Locking
v0.59 is partial live evidence. It reran the v0.58 fact-locking matrix after proving the chat slot was idle with a tiny authenticated chat probe. This avoided the v0.58 503 busy cascade and produced 14 successful scored rows before one fast HTTP 500 Invalid input batch stopped the matrix.
| Slice | Measured result | CTO meaning |
|---|---|---|
| Idle gate | idle probe success True; attempts 1 | The correct readiness probe is a tiny chat request, not only `/health`. |
| Live attempt | 15/18 rows attempted, 14/15 HTTP 200, stop reason non_200_after_v058-agent_loop_directive-locked_fact_ids_controller_expands-p1024 | The run produced quality evidence before stopping; it is partial, not a full 18-row matrix. |
| Completed-row quality | strict true-north 100.0%; semantic true-north 100.0%; coherence 100.0% | Every completed scored row was coherent and semantically correct under the harness. |
| Safety invariants | stale absence 100.0%; foreign absence 100.0%; rejected-key recall 100.0% | The completed rows preserved the core safety invariant: no stale/foreign memory admitted. |
| Best variant | budgeted fact-ID output: current-fact-ID recall 100.0%; n=4 | Budgeted fact-ID output is the strongest completed V3 direction, but it has only four completed rows in this partial run. |
| Boundary | non-200 row: v058-agent_loop_directive-locked_fact_ids_controller_expands-p1024 | This is not the v0.58 busy failure. It is an input-batch validity edge on a specific agent-loop locked-ID prompt shape. |
v0.60 Agent-Loop Input-Batch Hardening
v0.60 is live evidence. It isolated the v0.59 agent-loop HTTP 500 boundary with a six-row matrix: original replay, hardened JSON-ID instruction, and budgeted ID output at 1024 and 2048 pressure. All six rows returned HTTP 200, so the v0.59 Invalid input batch did not reproduce.
| Slice | Measured result | CTO meaning |
|---|---|---|
| Serving | 6/6 rows completed, 6/6 HTTP 200, stop reason None | The v0.59 500 was not a stable reproducible input-batch failure under identical replay. |
| Original replay | coherence 100.0%; current-fact-ID recall 100.0%; n=2 | The original controller-expands instruction remains the best agent-loop ID contract at 1024/2048 pressure. |
| Hardened JSON IDs | coherence 0.0%; parse 0.0% | The shorter literal-array instruction caused plain-text output, not safer JSON. |
| Budgeted output | coherence 0.0%; parse 0.0% | Budgeted ID output was not robust for the agent-loop slice even though it looked strong in v0.59's completed non-agent rows. |
| Safety invariant | stale absence 100.0%; foreign absence 100.0% | Even failed parse rows did not admit stale or foreign memory text; the failure mode was output-contract loss. |
v0.61 Agent-Loop Tail Output Contract
v0.61 is live evidence. It kept the agent-loop prompt body close to the working v0.60 original-control path and moved the JSON/output contract to the prompt tail. The measured result is not a clean 6/6 transport pass: one original-control 2048 row disconnected after 152s. The two new tail-contract variants passed all four rows they owned at 1024 and 2048 pressure.
| Slice | Measured result | CTO meaning |
|---|---|---|
| Serving | 6 rows attempted, 5/6 HTTP 200, non-200 v061-agent_loop_directive-controller_expands_original_control-p2048 | Do not overclaim a full transport pass. One original-control 2048 row hit RemoteDisconnected. |
| Tail contract | coherence 100.0%; parse 100.0%; n=2 | Tail-appended output contract preserved JSON/ID compliance without the plain-text collapse seen in v0.60's shortened variants. |
| Tail schema example | coherence 100.0%; parse 100.0%; n=2 | The schema-example tail also passed 1024 and 2048 pressure; this is now the best next candidate to rerun at wider probe diversity. |
| Safety invariant | scored-row stale absence 100.0%; foreign absence 100.0%; rejected-key recall 100.0% | On the scored rows, the system kept current facts and rejected stale/foreign handles under pressure. |
What Is Not Proven
- Not proven: general random-access exact recall across arbitrary positions.
- Not proven: production multi-tenant safety under live traffic.
- Not proven: transfer of the same multiplier to frontier-scale models.
- Not proven: stable correction on known failed slices; focused rerun confirmed the weak cluster.
- Not a leaderboard-certified MTRAG result in this board.
V3 Optimization Targets
- Make controller-selected payloads first-class: selected, candidate, stale, revoked, and foreign-tenant states must be impossible to confuse.
- Improve arbitrary-position exactness, especially non-tail provenance recovery.
- Promote focused-row/resume execution into the standard harness so hard suite walls do not create partial matrices.
- Promote update precedence to a first-class memory field: original fact, superseding fact, effective timestamp, and scope.
- Add stable entity-id constraints for agent-task memories; current failures often recall facts but return the wrong entity id.
- Use runtime-selected payloads as the V3 default; v0.54 shows selected-current-only hit 100% coherence in that matrix.
- Make controller-filled identity and rejection metadata the default; v0.55 shows `controller_fills_all_ids` is the best envelope mode.
- Do not assume deterministic envelopes solve recall; v0.55 1024-pressure semantic pass was only 11.1%.
- Add cancellation-aware serving control before larger pressure sweeps; v0.56 showed client timeout can leave the backend busy and contaminate later sequential rows.
- Add pressure-aware output budgeting; v0.57 showed one 1024-pressure row failed through length/parse output despite current facts being available.
- Add current-fact locking for update precedence; v0.57 showed one 2048-pressure research-update row dropped to 33.33% current-fact recall.
- Keep chat-slot idle probing; v0.59 showed it avoids the v0.58 busy cascade.
- Do not over-shorten agent-loop ID prompts; v0.60 showed original controller-expands passed while shorter hardened/budgeted variants returned plain text.
- Use tail-anchored output contracts for agent-loop prompts; v0.61 showed tail-contract and schema-example variants passed 1024/2048 scored rows while preserving parse and ID recall.
- Do not rely on the current repair prompt; v0.54 repair attempts produced 0% measured improvement.
- Expose substrate observability per run: what was selected, what was rejected, and why.
Next Decision
- continue Build V3 update-precedence and stable-entity binding tests from the confirmed failure cluster.
- tighten Replace broad dashboards with decision boards after major rounds only.
- measure Add latency/cost normalization to true-north score so slow 1024-band wins do not hide serving constraints.
- do not claim “infinite perfect memory” or “general exact recall” yet.
Deck-Safe One-Liner
Hypernym Infinite Memory is a memory control plane for model fleets: per-tenant memory stores, controller-curated recall, provenance-handle verification, and lower long-memory serving cost for small local models under extreme context pressure.
Deck-safe caveat: current evidence supports pressure handling and prepared control-plane safety tests; general exact arbitrary recall remains the optimization target.
Data Trace
Every row below is a retrieval handle. A future agent, CTO, or local API client should be able to use these paths to pull the raw score object, reconstruct the claim, and follow the run into the ledger or CXDB handoff. This mirrors the RMT/Hermes lineage frame: run id to score artifact to finding to durable handoff.
| Claim | Run / id | Type | Source path | Used for |
|---|---|---|---|---|
| largest observed pressure | 20260608T202409Z | measured fact | research/tracks/hypernym-infinite-mim/results/byte-threshold/20260608T202409Z/scores.json | 48.56x native-context pressure claim |
| v0.51 personal entity coherence | 20260610T_personal_entity_coherence_combined_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.51-personal-entity-coherence-threshold/20260610T_personal_entity_coherence_combined_live_codex_v1/scores.json | single-user memory true-north baseline |
| v0.52 update precedence / entity binding | 20260610T_update_precedence_entity_binding_combined_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.52-update-precedence-entity-binding/20260610T_update_precedence_entity_binding_combined_live_codex_v1/scores.json | prose-vs-structured prompt comparison |
| v0.53 record format / pressure diagnostic | 20260610T_record_format_schema_pressure_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.53-record-format-schema-pressure-diagnostic/20260610T_record_format_schema_pressure_live_codex_v1/scores.json | strict vs semantic contract failure cluster |
| v0.54 selected payload recovery | 20260610T_runtime_selected_minimal_payload_strict_repair_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.54-runtime-selected-minimal-payload-strict-repair/20260610T_runtime_selected_minimal_payload_strict_repair_live_codex_v1/scores.json | selected-current payload V3 lever |
| v0.55 deterministic envelope constraint | 20260610T_deterministic_rejected_handle_envelope_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.55-deterministic-rejected-handle-envelope/20260610T_deterministic_rejected_handle_envelope_live_codex_v1/scores.json | controller metadata helps; pressure recall still fails |
| v0.56 pressure serving boundary | 20260610T_selected_current_controller_envelope_pressure_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.56-selected-current-controller-envelope-pressure-recall/20260610T_selected_current_controller_envelope_pressure_live_codex_v1/scores.json | 2048-pressure timeout and single-flight busy contamination boundary |
| v0.57 timeout-aware pressure retry | 20260610T_timeout_aware_sharded_pressure_retry_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.57-timeout-aware-sharded-pressure-retry/20260610T_timeout_aware_sharded_pressure_retry_live_codex_v1/scores.json | serving-control fix and pressure-dependent memory quality score |
| v0.58 serving-slot busy stop | 20260610T_current_fact_locking_output_budget_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.58-current-fact-locking-output-budget/20260610T_current_fact_locking_output_budget_live_codex_v1/scores.json | health-ok but chat-slot-busy boundary before quality evidence |
| v0.59 idle-gated fact-locking partial live result | 20260610T_chat_slot_idle_gated_fact_locking_retry_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.59-chat-slot-idle-gated-fact-locking-retry/20260610T_chat_slot_idle_gated_fact_locking_retry_live_codex_v1/scores.json | chat-idle gate success, fact-ID variant signal, and HTTP 500 input-batch boundary |
| v0.60 agent-loop input-batch hardening result | 20260610T_agent_loop_input_batch_hardening_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.60-agent-loop-input-batch-hardening-fact-locking-completion/20260610T_agent_loop_input_batch_hardening_live_codex_v1/scores.json | proves v0.59 HTTP 500 did not reproduce and original controller-expands passed agent-loop 1024/2048 |
| v0.61 agent-loop tail output-contract result | 20260610T_agent_loop_tail_output_contract_live_codex_v1 | measured fact | research/tracks/hypernym-infinite-mim/results/v0.61-agent-loop-tail-output-contract/20260610T_agent_loop_tail_output_contract_live_codex_v1/scores.json | tail contract and schema-example variants passed 1024/2048; original-control 2048 had one transport disconnect |
| current objective readiness audit | 20260610T_objective_completion_matrix_16384_ceiling_codex_v1 | gate/audit artifact | research/tracks/hypernym-infinite-mim/results/objective-readiness-audit/20260610T_objective_completion_matrix_16384_ceiling_codex_v1/audit.json | current completion matrix: 13 satisfied harness requirements, 2 partial live evidence, 6 blocked live-certification requirements; includes 8192 pass and 16384 gateway ceiling |
| current eval-suite validation | 20260610T_eval_adaptive_pressure_plan_codex_v3 | validation artifact | research/tracks/hypernym-infinite-mim/results/eval-suite-manifest-validation/20260610T_eval_adaptive_pressure_plan_codex_v3/report.json | proves the current suite artifact index, audited 8192/16384 boundary, Granite Pilots session router, local session lock guard, adaptive pressure scheduler, packet, comparison, checklist, first-live subset, threshold finalizer, and full-suite certification finalizer are internally consistent |
| Granite Pilots session router | 20260610T_pilot_session_router_dryrun_codex_v1 | session routing guard | research/tracks/hypernym-infinite-mim/results/granite-pilots-session-router/20260610T_pilot_session_router_dryrun_codex_v1/route.json | enforces one durable handle for this track, session field on chat bodies, max two active pilots, and undock-when-idle shared-capacity behavior |
| Granite Pilots live session-memory probe | 20260610T_granite_pilots_session_memory_probe_live_codex_v2 | measured live fact | research/tracks/hypernym-infinite-mim/results/granite-pilots-session-memory-probe/20260610T_granite_pilots_session_memory_probe_live_codex_v2/scores.json | proves the one-handle Granite Pilots lane can recall five current private memory facts across domains, reject stale/rejected/foreign controls, and undock after use under the two-active-pilot cap |
| Granite Pilots live pressure ladder | 20260610T_granite_pilots_pressure_ladder_live_codex_v1 | measured live fact | research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_live_codex_v1/scores.json | measures private session recall under sequential pressure bands 0, 256, and 1024; highest strict pass 1024 with 24,584 reported prompt tokens and no stale/rejected/foreign leakage |
| Granite Pilots 2048 pressure extension | 20260610T_granite_pilots_pressure_ladder_2048_live_codex_v1 | measured live fact | research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_2048_live_codex_v1/scores.json | independently measures the next pressure boundary: 2048 band passed strict and semantic recall with 52,933 reported prompt tokens, 55.18s latency, no forbidden hits, and mandatory undock |
| Granite Pilots 4096 pressure extension | 20260610T_granite_pilots_pressure_ladder_4096_live_codex_v1 | measured live fact | research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_4096_live_codex_v1/scores.json | independently measures the next pressure boundary: 4096 band passed strict and semantic recall with 105,498 reported prompt tokens, 106.04s latency, no forbidden hits, and mandatory undock |
| Granite Pilots 8192 pressure extension | 20260610T_granite_pilots_pressure_ladder_8192_live_claude_v1 | measured live fact | research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_8192_live_claude_v1/scores.json | independently measures the next pressure boundary: 8192 band passed strict and semantic recall with 210,631 reported prompt tokens, 196.48s latency, no forbidden hits, and mandatory undock |
| Granite Pilots 16384 pressure ceiling | 20260610T_granite_pilots_pressure_ladder_16384_live_codex_v1 | measured gateway ceiling | research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_16384_live_codex_v1/scores.json | identifies the current request-size ceiling: setup turns admitted, 16384 probe rejected by nginx HTTP 413 Request Entity Too Large in 0.43s, no memory-quality verdict, mandatory undock |
| Granite Pilots local session lock guard | 20260610T_granite_pilots_pressure_ladder_lock_dryrun_codex_v1 | tooling guard artifact | research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_lock_dryrun_codex_v1/scores.json | proves dry-run lock metadata and validates live-capable Pilots tools now use a shared nonblocking local session-handle lock plus remote berth gate |
| Granite Pilots adaptive pressure plan | 20260610T_granite_pilots_adaptive_pressure_plan_codex_v2 | pressure scheduler artifact | research/tracks/hypernym-infinite-mim/results/granite-pilots-adaptive-pressure-plan/20260610T_granite_pilots_adaptive_pressure_plan_codex_v2/plan.json | encodes 8192 as measured pass ceiling, 16384 as gateway ceiling, and directs >8192 exploration toward chunked session continuation rather than oversized single-shot probes |
| next live evidence matrix | 20260610T_next_live_evidence_matrix_codex_v1 | objective-to-row ledger | research/tracks/hypernym-infinite-mim/results/next-live-evidence-matrix/20260610T_next_live_evidence_matrix_codex_v1/matrix.json | maps each of the six unproved objective gaps to exact labels, turns, runners, stop conditions, and the 27-row first-live certification subset |
| hardened admission lease validation | 20260610T_admission_gate_valid_lease_validation_codex_v1 | gate validation artifact | research/tracks/hypernym-infinite-mim/results/v0.66-admission-control-gate/20260610T_admission_gate_valid_lease_validation_codex_v1/decision.json | proves the admission gate accepts only a fresh track-scoped, endpoint-scoped, sequential-only lease; wrong-track, stale, wrong-endpoint, missing, and parallel-unsafe fixtures are refused |
| runtime runner refusal guard | 20260610T_admission_runtime_runner_refusals_codex_v1 | runtime guard artifact | research/tracks/hypernym-infinite-mim/results/admission-runtime-guard-validation/20260610T_admission_runtime_runner_refusals_codex_v1/report.json | proves first-live, full-suite, Q2, and Q4 runners refuse expired allow decisions at runtime before endpoint traffic |
| indexed full-suite dry-run | 20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1 | harness artifact, not capability evidence | research/tracks/hypernym-infinite-mim/results/personal-memory-eval-suite/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1/suite-run.json | machine-readable Q1/Q2/Q3/Q4 artifact index and ready-to-copy threshold/version-comparison commands |
| full-suite artifact index | artifact-index | machine-readable trace | research/tracks/hypernym-infinite-mim/results/personal-memory-eval-suite/20260610T_personal_memory_eval_suite_indexed_dryrun_codex_v1/artifact-index.json | future live run path reconstruction for Q1/Q2/Q3/Q4 comparisons without manual lookup |
| indexed full-suite live gate refusal | 20260610T_personal_memory_eval_suite_indexed_live_gate_refusal_codex_v1 | gate artifact | research/tracks/hypernym-infinite-mim/results/personal-memory-eval-suite/20260610T_personal_memory_eval_suite_indexed_live_gate_refusal_codex_v1/suite-run.json | proves the full live suite refuses before endpoint traffic while v0.66 blocks memory-quality rows |
| full-suite certification finalizer | 20260610T_full_suite_certification_dry_index_codex_v2 | certification guard | research/tracks/hypernym-infinite-mim/results/full-suite-certification-finalizer/20260610T_full_suite_certification_dry_index_codex_v2/finalizer.json | consumes full-suite artifact-index.json; refuses non-live Q1/Q2/Q4 unless explicitly allowed; treats Q3 as safety-control evidence when page safety passes |
| versioned eval packet v36 | 20260610T_versioned_eval_packet_v3_candidate_codex_v36 | fingerprinted rerun packet | research/tracks/hypernym-infinite-mim/results/versioned-eval-packet/20260610T_versioned_eval_packet_v3_candidate_codex_v36/packet.json | future V3/V-next reruns with identical objective, gates, catalog, scoring contract, and manifest fingerprint |
| version-comparison scaffold v36 | 20260610T_version_comparison_scaffold_codex_v36 | comparison guard | research/tracks/hypernym-infinite-mim/results/version-comparison/20260610T_version_comparison_scaffold_codex_v36/comparison.json | refuses dry-run, gate-refusal, health-only, or missing Q1/Q2/Q3/Q4 artifacts as capability evidence |
| live launch checklist v41 | 20260610T_live_launch_checklist_codex_v41 | operator checklist | research/tracks/hypernym-infinite-mim/results/live-launch-checklist/20260610T_live_launch_checklist_codex_v41/checklist.json | explicit launch order: status, audit, validation, dry-run, gate, first-live subset, threshold finalizer, live suite, version compare |
| current compound ledger | ledger | compiled artifact | .forge/artifacts/hypernym-infinite-memory-research-ledger.json | machine-readable index of source runs |
| current CXDB handoff | cxdb-hypernym-infinite-mim-handoff-20260610 | durable handoff | .forge/artifacts/cxdb-hypernym-infinite-mim-handoff-20260610.json | resume/API import packet |
Compound Research Chain
These are the prior pages and local artifacts this board compounds. Public links are directly openable. Local artifact paths are intended for Forge/CXDB/API retrieval from the workstation or repository.
| Document | Kind | Reference |
|---|---|---|
| Current public board | public page | https://hypernym-infinite-memory-v09.pages.dev/ |
| Previous immutable board with v0.55 | public page | https://2f1e75b6.hypernym-infinite-memory-v09.pages.dev |
| Previous immutable objective-audit board | public page | https://1b03ee20.hypernym-infinite-memory-v09.pages.dev/ |
| Previous immutable board with v0.54 | public page | https://1db5d7c7.hypernym-infinite-memory-v09.pages.dev |
| Institutional 2026-06-09 page | local artifact | .forge/artifacts/hypernym-infinite-memory-institutional-20260609.html |
| CTO findings 2026-06-09 page | local artifact | .forge/artifacts/hypernym-infinite-memory-cto-findings-20260609.html |
| Research ledger JSON | local artifact | .forge/artifacts/hypernym-infinite-memory-research-ledger.json |
| Compound visualization standard | local standard | research/tracks/hypernym-infinite-mim/compound-research-visualization-standard.md |
API Pull Targets
Canonical local query shape for future automation:
jq '.summary' research/tracks/hypernym-infinite-mim/results/<suite>/<run_id>/scores.jsonjq '.latest_results' .forge/artifacts/cxdb-hypernym-infinite-mim-handoff-20260610.jsonjq '.runs.v057_timeout_aware_sharded_pressure_retry_live' .forge/artifacts/hypernym-infinite-memory-research-ledger.jsonjq '.runs.v058_current_fact_locking_output_budget_live' .forge/artifacts/hypernym-infinite-memory-research-ledger.jsonjq '.runs.v059_chat_slot_idle_gated_fact_locking_retry_live' .forge/artifacts/hypernym-infinite-memory-research-ledger.jsonjq '.runs.v060_agent_loop_input_batch_hardening_live' .forge/artifacts/hypernym-infinite-memory-research-ledger.jsonjq '.runs.v061_agent_loop_tail_output_contract_live' .forge/artifacts/hypernym-infinite-memory-research-ledger.jsonjq '.summary' research/tracks/hypernym-infinite-mim/results/granite-pilots-session-memory-probe/20260610T_granite_pilots_session_memory_probe_live_codex_v2/scores.jsonjq '.summary' research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_live_codex_v1/scores.jsonjq '.summary' research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_2048_live_codex_v1/scores.jsonjq '.summary' research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_4096_live_codex_v1/scores.jsonjq '.summary' research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_8192_live_claude_v1/scores.jsonjq '.summary' research/tracks/hypernym-infinite-mim/results/granite-pilots-pressure-ladder/20260610T_granite_pilots_pressure_ladder_16384_live_codex_v1/scores.json- Suggested CXDB lineage key:
research:hypernym-infinite-mim:<run_id>:<finding_id>