We kept the lineage. We changed the system.
You and Natasha built the original science and got TraceScore running in GCP. We kept that provenance and V1 as the control. We also changed the estimator path, added new decision math, rebuilt how the work runs in GCP, and made it all inspectable through one harness.
- Repository
- github.com/preciseai-inc/precise
- Architecture
docs/atlas/pipeline.html- Start here
cd harness && npm run precise -- docs
This was not just a wrapper job.
peer noteHey Matt,
This is not an introduction to your own system. You were the method architect, Natasha wrote the production TraceScore engine, and Abhay's research supplied the reference Shapley implementations. The starting point was already a deliberate BigQuery to Shapley to BigQuery production job with windowing, multi-segmentation, routing, and append-only run records. We preserved that lineage, and we kept the original engine intact as a benchmark.
But saying we merely productionized around it is too soft. We made reviewable changes inside the active engine: explicit seed plumbing, an exact path for additive games, an optional vectorized complementary-contribution path, explicit ratio-denominator semantics, and corrected variance and degrees-of-freedom accounting. We also built CI and interaction work that is real and tested but not yet on the regular job route.
Then we added math the first system did not have: a leak-free walk-forward calibration court, conformal and regime work, attainable par, a measured next-dollar frontier, exact Owen hierarchy attribution, and experimental learned-value and gradient paths. Those pieces are at different stages, and the repository says which are live, connected, optional, or still POCs.
We also changed the operating shape. TraceScore can now fan out through immutable manifests, deterministic shard seeds, idempotent outputs, canonical merges, and run and cost receipts. The gold replay matched 672 of 672 keys to roughly machine precision. On the frozen 150-segment sampled benchmark, the vectorized path used about 26 times fewer compute-seconds and the best fleet path finished about 12 times faster end to end than preserved V1. A newer GCP spine now adds dispatch, registry, admin, services, jobs, and dated estate observations without disturbing the protected original estate.
The harness is the common way into all of that, not the whole accomplishment. It lets a human, product, job, or agent call the real core, inspect its claims and floors, run comparable variants, and see the difference between source code, a connected capability, and something actually observed in GCP.
I am not looking for a verdict on somebody else's finished system. Use Codex to pull on the threads, understand why each change exists, reproduce whatever matters, and then help us deepen the science and expand the system. I had it make the map and the prompt below so you can get to the consequential parts quickly.
Call me after you have poked at it.
Adam
Here is the evolution in one pass.
Matt's method · Natasha's engine · Abhay's references · V1 control
exact + CC · seeded RNG · fast path · explicit objective semantics
calibration · par · frontier · hierarchy · floors · seal
products · jobs · agents · receipted GCP · estate evidence
Inside the math
- Additive games now settle exactly at any size
- Sampled paths can be seeded and made dramatically faster
- Ratio singularities are explicit instead of accidental
- Variance, effective degrees of freedom, and missing cells are handled honestly
- Calibration, par, frontier, hierarchy, and learned-math seams are new
Inside the operating system
- Cloud Run fan-out is content-addressed, replayable, and receipted
- Gold replay is 672 of 672; sampled fleet runs are materially faster
- Dispatch, registry, admin, governed gateway, and ledgers are new, at different deployment stages
- The estate now separates declared code from dated GCP observation
- The harness gives every client one truthful route into the system
The provenance and source lineage remain canonical. V1 is preserved as the control. The active fork carries explicit, reviewable math and runtime changes instead of hiding them in a clean-room rewrite.
The original exact and complementary-contribution routing remains, with explicit seed plumbing, additive exactness, an opt-in fast path, and explicit denominator policy. TypeScript and Python exact settlement share hand-computed golden games. The harness can call the real Python exact and serial CC core, but it does not yet bridge every router.
Sampled CIs, pairwise interactions, exact Owen values, conformal calibration, and learned-value or gradient work exist at different levels of maturity. They should be evaluated as options, not described as product-live by association.
Immutable workloads, HMAC shard seeds, deterministic task claims, canonical outputs, replay verification, and cost receipts ran in precise-pipeline-prod. The 672-key replay preserved the known answer; sampled paths have statistical parity, not a false bit-for-bit claim.
trace-staging-486619 is the protected original Trace estate. precise-pipeline-prod is the newer operating spine. The July 14 probe observed the newer spine only, so we should not imply the fork replaced the original live jobs.
Tests prove behavior. Evals enforce invariants and floors. Ablations compare estimators, policies, or models. Evidence-bound experiments decide adoption. Products, jobs, UIs, and agents can all use that same contract.
Here is where I want us to push it next.
This is a working agenda, not a request for a ceremonial review.
I am coming at this as a product and systems builder. I built Madhive, I know the shape a large media system needs, and I have a clear sense of where these tools are going. You have the deeper scientific read. I want you to use the system, talk with Codex about what it found and why, and help us understand, extend, and improve it.
Graduate the right estimator changes
Work through which parts of the fork deserve to become normal TraceScore behavior and which should remain explicit options: additive exactness, fast CC, denominator policy, surfaced intervals, hierarchy, and interaction work.
Make the decision layer scientifically real
The measured frontier is useful now. The curved-gradient core and tests also exist, so this is not a from-scratch research problem, but it remains unmounted and untrained. We need the right estimand, saturation model, counterfactual, uncertainty, and outcome join before it can drive a larger media move.
Let Konstant propose what Precise should test
RSL-native VFAST compresses pre-aggregated support, opposition, activation, pair energy, uncertainty, and orphan pressure. Syncopation nominates combinations whose joint value may differ from their parts. Orphan detection surfaces coherent recurring activity the current representation did not name.
The clean seam is not to drop those scores into TraceScore. Konstant can propose fields, blind spots, candidate pairs, and diverse opportunity sets. Precise tests whether each representation adds held-out media outcome value, owns the economic number, and uses exact, CC, interaction, or Owen math to explain the complete measured game.
This is a descendant of the same architectural insight you and Natasha proved: keep the value function separable and cheap enough to evaluate many coalitions. We already use this pattern in Polymarket, with borrowed VFAST weights treated as an unprivileged challenger beside label-permutation, time-shift, and signed-edge-rewire controls. There is no media adapter or evidence yet.
Close the operating circuit
Connect the full TraceScore router, one real product caller, the live outcome join, remote worker control, and common telemetry. The pieces exist, but they are not yet one self-grading production loop.
Five courts worth running
- Player basis. Compare current segment keys with Context-selected fields and orphan-augmented fields on the same campaign, advertiser, and time holdouts.
- Pair nomination. Let VFAST or syncopation nominate a small set of interactions, then require held-out full-game value and the Precise interaction court to clear them.
- Online decision. Compare measured frontier, additive value model, curved gradient, and VFAST-gated mixture on exact gap, band calibration, outcome value, and action utility.
- Orphan as regime. Test orphan pressure as direct signal versus a change-point or mixture gate. The latter may be the more defensible use.
- Diverse opportunity book. Test whether a diversity-aware Konstant selector can remove redundant recommendations while preserving total held-out value and executable capacity.
The repo will show you both the science and the system.
Start with the map, then inspect the math fork, the new courts, the deployment receipts, and the gap between repository intent and observed GCP state.
gh repo clone preciseai-inc/precise
cd precise/harness
npm ci
npm run precise -- docs
npm run precise -- describe
npm run precise -- estate coverage
npm run precise -- estate deployments --observed trueMATH.md
The diligence-grade map of the actual estimators, calibration, and decision math.
trace-score-job/FORK-NOTES.md
The reviewable changes inside the original engine lineage.
burst/docs/BENCH.md
The real GCP parity, speed, determinism, failure, and cost receipts.
calibration-breadth/README.md
The leak-free model court and the challengers it refused.
CONNECTIONS.md
What calls real code now, what is partial, and what is still a POC.
PRECISE-ESTATE-BACKPLANE.md
The observed production estate, telemetry boundary, and honest gaps.
Let your coding agent do the repo walk with you.
This prompt treats you as Precise's Chief Scientist. It finds what changed inside the estimator, what math was added, what ran in GCP, and where the harness or estate still overstates reality.
Preview the coding-agent prompt
You are working with Matt Barlin, Precise's Chief Scientist and method architect. Matt helped build the original Precise science and GCP services. Do not onboard him to his own system or explain his own math back to him.
Your job is to help Matt inspect, understand, and extend what changed inside and around that work: the TraceScore fork, entirely new math and decision courts, distributed execution, production paths, harness, estate observations, and current connection gaps.
Repository:
https://github.com/preciseai-inc/precise
Target handoff commit:
856dd76c5fb2ee9a0fe225092b41c47ccfebe4dd
Clone or update the repository safely. Preserve and report any local changes. Verify that the target commit is available before treating this prompt as repository truth. If it is not on the remote, say so plainly and continue only with Matt's direction.
Start read-only. Do not deploy, mutate GCP, or rewrite an existing service during the repository walk.
Read these completely:
1. AGENTS.md
2. .claude/skills/precise-harness/SKILL.md
3. .claude/skills/boundary/SKILL.md
Then read these completely:
1. packages/trace-score-job/FORK-NOTES.md
2. docs/atlas/MATH.md
3. packages/calibration-breadth/README.md
4. packages/burst/README.md
5. packages/burst/docs/BENCH.md
6. docs/atlas/CONNECTIONS.md
7. docs/atlas/PRECISE-ESTATE-BACKPLANE.md
8. packages/polymarket-observatory/README.md
Then run:
cd harness
npm ci
npm run precise -- docs
npm run precise -- skill
npm run precise -- describe
npm run precise -- estate coverage
npm run precise -- estate drift
npm run precise -- estate deployments --observed true
npm run precise -- optional-surfaces
Read the files returned in data.readInOrder. Treat generated estate status and docs/atlas/CONNECTIONS.md as the declared status, then check them against code, tests, receipts, and observed deployment evidence. Report any mismatch rather than inheriting it.
Give Matt a compact working map:
1. Preserve the original authorship and V1 control, then show the exact mathematical and runtime changes made in the active TraceScore lineage.
2. Separate changes inside the estimator from entirely new calibration, decision, hierarchy, and learned-math work.
3. Show which paths are live in the regular job, optional, harness-connected, prototype, POC, or unbuilt.
4. Show what the burst, dispatch, registry, gateway, and newer GCP spine actually added, including the captured parity and benchmark evidence.
5. Explain the difference among the harness tree, estate catalog, deployment observations, and CONNECTIONS.md.
6. Report any disagreement among code, docs, receipts, estate metadata, and observed GCP state.
7. Bring back questions where Matt's scientific or operational history is more authoritative than the repository.
8. Explain how the isolated Polymarket lab reuses Precise math, harness courts, ablation, and GCP machinery, what its first null result actually established, and what reusable improvements came back into Precise.
9. If the sibling konstantin.ai repository is available, inspect lib/resonance/v-fast-rsl.ts and the orphan/resonance path. Explain how Konstant can propose representations while Precise tests held-out media value and owns the economic decision. Do not treat the older syncopation/v-fast.ts path or its default weights as validated Precise science. Return concrete player-basis, pair-nomination, online-decision, orphan-regime, and diverse-slate courts.
Keep it technical, candid, and compact. Do not describe this as a wrapper around frozen code. Do not imply the fork replaced the protected original GCP jobs. Do not promote CIs, interactions, Owen, conformal, VFAST, orphan, syncopation, or learned paths merely because code exists. Treat products, jobs, UIs, and agents as clients of the same contract. Distinguish tests, evals, ablations, and evidence-bound experiments. Do not make Polymarket the center of Precise.
After the first map, become Matt's working research partner. Answer follow-up questions by tracing claims to code and receipts. Reproduce a result when useful. Help him turn a scientific idea into a swappable variant, invariant, falsification control, and evidence-bound court. Implement only when Matt asks, preserve the V1 control, and keep maturity labels honest. Follow his direction.Then use Codex to take it further.
This is not a request for a verdict. Ask Codex why a boundary exists, make it reproduce a result, have it add a challenger behind a harness seam, or build a new court with you. I will keep pushing the product and system shape. I would love you to push the science, and the harness gives us a clean way to make the two meet.